{
  "name": "aiNOW anonymized case studies",
  "description": "Machine-readable feed of aiNOW's anonymized (NDA) AI project case studies: industry, TL;DR summary, before/after results, tech stack, full case study and FAQ. English.",
  "provider": {
    "name": "aiNOW",
    "url": "https://ainow.ge"
  },
  "citationPolicy": "Cite as aiNOW (https://ainow.ge).",
  "count": 12,
  "cases": [
    {
      "slug": "kids-ecommerce-size-assistant",
      "url": "https://ainow.ge/projects/kids-ecommerce-size-assistant",
      "title": "Kids clothing store: AI size assistant",
      "industry": "E-commerce",
      "slogan": "Size matching by height and age, 24/7 in WhatsApp",
      "summary": "An online kids clothing store with heavy evening traffic could not keep up with a stream of identical size questions: five operators worked in shifts, and at peak a customer waited up to 40 minutes for a reply and often left for a competitor. We built an AI assistant in WhatsApp that understands natural questions, matches models by height and age and files the order into the CRM itself. Routine requests are almost fully closed by the bot, replies arrive in seconds, and chat to cart conversion went up.",
      "techStack": [
        "WhatsApp Business API",
        "Bitrix24",
        "RAG",
        "Google Sheets"
      ],
      "metrics": [
        {
          "label": "Reply time",
          "before": "up to 40 minutes",
          "after": "seconds"
        },
        {
          "label": "Routine size questions",
          "before": "on the operators",
          "after": "~60% on the bot"
        },
        {
          "label": "Cart conversion",
          "before": "lower from waiting",
          "after": "+~15%"
        }
      ],
      "caseStudy": {
        "context": "An online kids clothing store with about 60 thousand visits a month and a wide range from several dozen brands. Sales run mostly through messengers, and in the evening, when parents are free after work, the flow of requests grew several times over. Five operators worked in shifts but could not keep up with the peak: in the busiest hours a customer waited up to 40 minutes for a first reply, and some buyers simply closed the chat.",
        "diagnostic": "We reviewed a month and a half of correspondence and saw that more than half of the messages were the same question: will the size fit the child by height and age. Parents got confused because size grids differ between brands, and they asked again and again. A noticeable share of working time also went to manually copying contacts and orders into the CRM, and because of the long wait some warm clients drifted to competitors.",
        "problem": "Simple button bots only irritated people here: a parent needs advice from a real seller, not a menu of options. The task was to teach the assistant to understand free-form wording like my son is five, height one meter ten, what should I take for autumn and reply with a specific model rather than generic words. At the same time it has to keep brand size grids in mind, see stock levels and never invent availability that does not exist.",
        "solution": "We built an AI assistant in WhatsApp that understands questions in plain language and recommends specific models with a little room to grow, checking against the size database and current stock. It advises the size by brand, offers alternatives when an item is missing, and files the assembled order into the CRM itself with no manual copying. The tone is set by examples from the store real chats, so answers sound human. In hard cases, such as a refund, the dialog moves to an operator gently.",
        "steps": "The launch took about five weeks. First we exported the dialog history and gathered the frequent questions and parent phrasings from it. Then we built a size matching database for each brand and linked it to warehouse stock. Separately we worked out the boundaries: where the assistant answers on its own and where it must call a human. After that we ran the wording on part of the traffic, tuned the tone from real dialogs and only then rolled the assistant out to the whole flow of requests.",
        "impact": "Now the bot closes about 60 percent of routine size requests on its own, and the first reply arrives in seconds instead of the old forty minutes. Manual copying of orders into the CRM is gone: the client card is assembled automatically. Operators stopped answering identical questions and switched to complex cases and upsells. Chat to cart conversion grew by roughly 15 percent, because the buyer gets advice right away, while the urge to buy is still warm.",
        "limitations": "The assistant does not handle non-standard returns without a receipt, disputes over item quality or delivery complaints. It does not promise timelines that are not in stock and does not invent availability. As soon as a question goes beyond matching and booking, the bot hands the dialog to the senior administrator together with the chat history.",
        "methodology": "We compared CRM data for 30 days before and 30 days after the full launch of the assistant: first reply time, the share of requests closed without an operator and chat to cart conversion. All figures are rounded and given in ranges so a specific store cannot be identified from them.",
        "anonymityNote": "We do not disclose the brand, domain or employee names under NDA. The messenger and spreadsheet names are given only to show the real integration stack, not as advertising. All figures are rounded and given in ranges so the company cannot be recognized from them.",
        "faq": [
          {
            "q": "How does the bot account for different brands?",
            "a": "The size charts of each manufacturer are loaded into the database, so the size advice is tied to the specific brand rather than averaged."
          },
          {
            "q": "What if an item is out of stock?",
            "a": "The bot immediately offers close alternatives in the same size and price range instead of sending the client to search on their own."
          },
          {
            "q": "When does a live operator step in?",
            "a": "For returns, complaints and any non-standard situations the bot hands the dialog to an operator itself, together with the chat history."
          }
        ]
      }
    },
    {
      "slug": "coffee-chain-catering-requests",
      "url": "https://ainow.ge/projects/coffee-chain-catering-requests",
      "title": "Coffee chain: B2B catering intake",
      "industry": "Coffee",
      "slogan": "Qualifying corporate requests and a draft estimate in minutes",
      "summary": "A city coffee chain with its own bakery was losing large B2B catering leads: they were handled by hand over email, calls and messengers, and the first reply could take up to three hours. We built an AI assistant on Telegram and Instagram that qualifies the request itself, checks the kitchen load and sends a ready estimate. Request handling dropped to a couple of minutes, losses on peak days fell by about three times, and managers now work only warm deals.",
      "techStack": [
        "Telegram",
        "Instagram",
        "Google Sheets",
        "Make"
      ],
      "metrics": [
        {
          "label": "Time to estimate",
          "before": "up to 3 hours",
          "after": "a couple of minutes"
        },
        {
          "label": "Lost requests at peak",
          "before": "lost on holidays",
          "after": "about 3x fewer"
        },
        {
          "label": "Manager focus",
          "before": "manual pre-sales",
          "after": "only warm deals"
        }
      ],
      "caseStudy": {
        "context": "A city chain of around a dozen and a half coffee shops with its own bakery. Beyond retail, the company actively sells office catering, corporate breakfasts and custom cakes, and it is exactly these B2B orders that bring the largest tickets. Requests for them arrived mixed across every channel at once: email, Instagram Direct, WhatsApp and calls to a shared number. One manager effectively sorted them by hand, and on a busy day the first reply to a client could take up to three hours.",
        "diagnostic": "We pulled a month and a half of chats and the call log and broke down where the time went. More than a third of the manager working day was eaten by manually sending the same menu presentations and delivery terms. On holidays, when there are three times as many requests, the line overloaded: some calls dropped and requests from different channels got lost, because there simply was no single list. Every missed catering request is not a trifle but a large corporate ticket that walked to a competitor.",
        "problem": "The task was not to hang up a chatbot for show, but to close real qualification. The assistant should clarify the date, event format, number of guests and budget itself, check it against the real kitchen load and immediately assemble a draft estimate, while sounding human and on brand. Ideally the manager is left only to check the details and confirm the booking, rather than running the entire conversation from scratch.",
        "solution": "We built an assistant that works on Telegram and Instagram Direct as a single entry point for B2B. In the dialog it carefully collects the event parameters, queries the bakery load sheet and checks whether the team can take the order on the requested date. It then builds a personal offer with a preliminary estimate and, if the client is ready, records the request in a shared base and hands the manager a card with all the details. The tone is set by examples from the brand real correspondence, so the assistant sounds like a living employee rather than a faceless form, and it does not invent prices that are not in the price list.",
        "steps": "The project took about six weeks. First we digitized the menu, catering rates and kitchen load rules so the bot had something to rely on. Then we connected it to the sheets and set up the preliminary estimate calculation. We spent a separate week limiting the assistant imagination: strict wording boundaries and no promises beyond the price list. After that we ran the scenario on part of the incoming requests, tuned the tone and wording from real dialogs and only then rolled the assistant out to the full flow.",
        "impact": "After stabilization the assistant handles the first pass of almost all incoming B2B requests itself. The time from a request to a ready estimate dropped from several hours to a couple of minutes, and the share of lost requests on peak days fell by about three times, because everything now lands in a single list instead of dissolving between channels. The manager stopped being an inbox sorter and does what actually brings money: running warm deals, agreeing large orders and growing the partner network. The average catering ticket, meanwhile, did not drop.",
        "limitations": "The assistant is deliberately limited in its powers. It does not approve discounts above a set threshold, does not resolve disputes about food quality and does not confirm non-standard delivery terms. As soon as the conversation goes beyond the scenario, the bot does not improvise but hands the dialog to the department head together with the full context of the correspondence, so a human makes the call.",
        "methodology": "We measured from the request log and CRM data for six weeks before launch and six weeks after stabilization. We compared time to first estimate, the share of handled requests and conversion to a paid invoice. All figures are rounded and given in ranges so the company cannot be identified from them.",
        "anonymityNote": "The chain name, location addresses and employee names are hidden under NDA. The messenger and spreadsheet trademarks are named only to show the real integration stack, not as advertising. The figures are given in ranges and rounded so a specific company cannot be identified from them.",
        "faq": [
          {
            "q": "How is catering delivery cost calculated?",
            "a": "The bot takes the distance from the kitchen to the address on the map and applies the company rate grid, so the client sees the preliminary cost right away, without waiting for a manager."
          },
          {
            "q": "What if the kitchen is overloaded on the requested date?",
            "a": "The assistant sees the load from the sheet and does not promise the impossible: it offers the nearest free date or an alternative format, and passes disputed cases to a manager."
          },
          {
            "q": "How accurate is the preliminary estimate?",
            "a": "It is a calculation from the price list and the event parameters, the final price is confirmed by a manager after clarifying details. The bot stays within the price list and does not invent numbers."
          }
        ]
      }
    },
    {
      "slug": "dental-clinic-booking",
      "url": "https://ainow.ge/projects/dental-clinic-booking",
      "title": "Dental clinic: online booking without calls",
      "industry": "Dental",
      "slogan": "24/7 appointment booking in WhatsApp with live slot checks",
      "summary": "A private dental clinic was losing patients outside the working day: a noticeable share of calls came in the evening and on weekends and went unanswered, while the administrator was torn between the phone and the front desk during appointment hours. We connected an AI assistant in WhatsApp that offers free slots of the needed doctor itself, puts the booking into the calendar and reminds about the visit. Simple booking and rescheduling now go without a call, and first-visit bookings grew by about a quarter thanks to off-hours.",
      "techStack": [
        "WhatsApp Business API",
        "Google Calendar",
        "Custom CRM"
      ],
      "metrics": [
        {
          "label": "Off-hours requests",
          "before": "unanswered",
          "after": "booked 24/7"
        },
        {
          "label": "First-visit bookings",
          "before": "work hours only",
          "after": "+~25%"
        },
        {
          "label": "Administrator",
          "before": "phone and desk",
          "after": "less routine"
        }
      ],
      "caseStudy": {
        "context": "A private dental clinic with several doctors and a tight appointment schedule. The main flow of patients came by phone, and during working hours the line was busy almost constantly. At the same time a noticeable share of calls fell on evenings and weekends, when the clinic is already closed, and those requests simply went unanswered. Patients dislike waiting long on the line or calling back, so some of them left for another clinic that answered faster.",
        "diagnostic": "We reviewed a month of the call and booking log and saw a pattern: about a third of all requests are a simple booking or a visit reschedule that do not need a doctor at all. Such calls ate the administrator time exactly during the busiest appointment hours, when patients are already standing at the desk. A separate problem was the lack of reminders: without them patients forgot about the visit, and empty windows stayed in the schedule that no one managed to fill.",
        "problem": "We needed booking available at any time of day, not only during reception hours. The assistant has to understand natural wording like book me with a therapist on the weekend closer to noon and see the free slots precisely, so it does not put two patients at the same time. Separately we had to split ordinary booking from emergencies: with acute pain a person should reach the on-duty doctor at once, rather than going through a long booking scenario.",
        "solution": "We connected an AI assistant in WhatsApp that works like a round the clock front desk. It understands the request in plain language, checks the schedule and offers the nearest free windows of the needed doctor, and then puts the booking into the calendar itself. The day before the visit the patient gets a reminder, and rescheduling is possible right in the same chat. Signs of acute pain are recognized as priority, and the assistant routes the person to the on-duty doctor at once, without extra questions.",
        "steps": "The project took about a month. First we exported the doctors schedule and reception rules so the assistant understood which slots and of what duration it could offer. Then we set up the logic of booking, rescheduling and reminders, plus a separate priority route for emergencies. We ran the scenario on some patients, listened to how the answers sounded, tuned the wording to the clinic tone and only after that opened booking through the assistant to everyone without exception.",
        "impact": "Simple booking and rescheduling now go without a single call, at any time of day. First-visit bookings grew by about a quarter, and almost all of that growth came from evenings and weekends that used to be simply lost. Day-before reminders noticeably cut forgotten visits, and there are fewer empty windows in the schedule. The administrator stopped being torn between the phone and the desk and works with patients in the office rather than sorting identical calls.",
        "limitations": "The assistant only handles the organization of booking. It does not advise on treatment, does not assess the complexity of a case and does not name the price of specific procedures before an examination. Any medical question it passes to a doctor or the administrator, so the diagnosis and treatment plan always come from a human, not from the bot.",
        "methodology": "We compared booking data for 30 days before and 30 days after launch: the number of first-visit bookings, the share of requests in off-hours and the number of no-shows. The figures are rounded and given in ranges so a specific clinic cannot be identified from them.",
        "anonymityNote": "The clinic asked us not to name the brand, address or doctor names, so all details are depersonalized. The messenger and calendar names are given only to show the real integration stack, not as advertising. The figures are rounded and given in ranges so a specific clinic cannot be identified from them.",
        "faq": [
          {
            "q": "What about emergencies?",
            "a": "The bot recognizes signs of acute pain as priority and routes the patient to the on-duty doctor at once, bypassing the usual booking scenario."
          },
          {
            "q": "Can a visit be rescheduled?",
            "a": "Yes, the patient reschedules right in the same chat, and the calendar slot updates automatically, without a call to the administrator."
          },
          {
            "q": "Are visit reminders sent?",
            "a": "Yes, the day before the appointment the patient gets a reminder in the same messenger, which noticeably reduces the number of forgotten visits."
          }
        ]
      }
    },
    {
      "slug": "beauty-salon-booking",
      "url": "https://ainow.ge/projects/beauty-salon-booking",
      "title": "Beauty salon: 24/7 stylist booking",
      "industry": "Beauty",
      "slogan": "Booking and reminders on Instagram and WhatsApp, fewer no-shows",
      "summary": "A beauty salon with several stylists took bookings only by phone during working hours, while evening messages on Instagram were handled with delay and some clients simply never reached a booking. We connected an AI assistant on Instagram and WhatsApp that picks a stylist and time itself, puts the booking into the calendar and reminds about the visit a day ahead. Now about 40 percent of bookings come outside working hours, and no-shows are noticeably fewer thanks to reminders.",
      "techStack": [
        "Instagram",
        "WhatsApp Business API",
        "Google Calendar",
        "Meta Ads"
      ],
      "metrics": [
        {
          "label": "Off-hours bookings",
          "before": "almost none",
          "after": "~40% of bookings"
        },
        {
          "label": "No-shows",
          "before": "no reminders",
          "after": "noticeably fewer"
        },
        {
          "label": "Instagram reply",
          "before": "delayed",
          "after": "instant, 24/7"
        }
      ],
      "caseStudy": {
        "context": "A beauty salon with several stylists and services that differ in length, from a quick manicure to complex coloring. Bookings were taken only by phone and only during working hours, while the administrator combined calls with serving guests at the desk. The main flow of requests, meanwhile, came in the evening through Instagram Direct, when clients browsed the feed after work. Replying at once was not possible, and messages often hung unanswered until the next day.",
        "diagnostic": "We looked at where clients were lost and saw two holes. The first is speed: if a reply came several hours later, the person managed to book at a nearby salon that answered faster. The second is no-shows: without reminders clients forgot about the visit or mixed up the time, and empty windows appeared in the tight stylist schedule that there was no longer anyone to fill. Both problems hit revenue directly, since a free window with a stylist is lost money for the shift.",
        "problem": "We needed booking available when it suits the client to write: in the evening and on weekends, straight from Instagram or WhatsApp. The assistant has to understand which service is needed and how long it lasts, pick a suitable stylist and not put two people at the same time. Separately we needed automatic reminders and easy rescheduling: so a client can move a visit in a couple of messages, and the freed window immediately becomes available to others.",
        "solution": "We connected an AI assistant that replies both in Instagram Direct and in WhatsApp as a single booking point. In the dialog it clarifies the service, picks a free stylist and a convenient time, puts the booking into the calendar and sends a reminder a day before the visit. The client can reschedule or cancel right in the chat, and the window frees up for others at once. The tone is tuned to the salon style, so the assistant sounds friendly rather than a dry booking form.",
        "steps": "The project took about a month. First we gathered the list of services with their durations and price ranges so the assistant would calculate time in the schedule correctly. Then we set up the stylist schedules, the logic of matching and reminders, and the rescheduling and cancellation scenarios. We launched in stages: first on Instagram, where the main evening flow was, then connected WhatsApp. During the run-in we listened to real dialogs and adjusted the wording to the salon tone before opening booking to everyone.",
        "impact": "After launch about 40 percent of all bookings began arriving outside working hours, in the evening and on weekends, exactly the time that used to be almost entirely lost. Day-before reminders noticeably cut no-shows, and the tight stylist schedule stood idle with empty windows less often. Clients book in a couple of messages without waiting for a call, and the administrator works with guests in the salon instead of sorting the evening correspondence. Stylist load became more even across the week.",
        "limitations": "The assistant does not agree complex requests about the look, the coloring tone or an individual price: such details it passes to the stylist before the visit. It does not promise a result that depends on the initial state of the hair, and does not replace a live consultation. Disputed and non-standard requests always go to a human.",
        "methodology": "We compared bookings and no-shows for a month before and a month after launch: the share of bookings in off-hours, the number of no-shows and the overall fill of the stylist schedule. All figures are rounded and given in ranges so a specific salon cannot be identified from them.",
        "anonymityNote": "We do not disclose the salon name, address or stylist names under NDA. The platform and calendar names are given only to show the real integration stack, not as advertising. The figures are rounded and given in ranges so the company cannot be recognized from them.",
        "faq": [
          {
            "q": "What if the desired stylist is busy?",
            "a": "The bot offers the nearest free time with that stylist or picks another with similar services, so the client does not leave without a booking."
          },
          {
            "q": "Are reminders sent?",
            "a": "Yes, a reminder comes a day before the visit in the same messenger, and this is exactly what noticeably cuts no-shows."
          },
          {
            "q": "Can a booking be canceled?",
            "a": "Yes, cancel and reschedule are available right in the chat, and the freed window immediately becomes available to other clients."
          }
        ]
      }
    },
    {
      "slug": "realestate-lead-qualification",
      "url": "https://ainow.ge/projects/realestate-lead-qualification",
      "title": "Premium real estate: lead qualification",
      "industry": "Real estate",
      "slogan": "Filtering out off-target leads by budget, district and property type",
      "summary": "A premium real estate agency was paying for ads, but most of the incoming leads were off-target, and agents spent days chatting with people who were not ready to buy. The first reply sometimes came hours later, and the client interest cooled down. We added an AI assistant on WhatsApp that clarifies budget, district, property type and timeline itself, filters out random requests and hands the agent a ready card with matching options. Now agents mostly work with warm leads, and the first contact takes minutes.",
      "techStack": [
        "WhatsApp Business API",
        "Bitrix24",
        "NLU",
        "Google Sheets",
        "Meta Ads"
      ],
      "metrics": [
        {
          "label": "First contact",
          "before": "hours of waiting",
          "after": "minutes"
        },
        {
          "label": "Leads on the agent desk",
          "before": "everyone at once",
          "after": "only warm ones"
        },
        {
          "label": "Filtering off-target",
          "before": "by hand in chat",
          "after": "at the entry"
        }
      ],
      "caseStudy": {
        "context": "A premium real estate agency with a stream of requests from paid ads and social media. The properties are expensive, the deal cycle is long, and the ad traffic came in mixed: alongside real buyers wrote the curious, renters and people for whom the property was out of budget. Several agents sorted every request by hand in chat. As a result, the most expensive time of the specialists went to conversations that almost never ended in a viewing.",
        "diagnostic": "We broke down the funnel from request to viewing and saw the bottleneck right at the start. Only a small share of requests reached a viewing, and the first reply sometimes took a client several hours, because the agent was busy at a meeting or with another client. During that time interest cooled and the person went to whoever answered faster. Meanwhile, agents could not tell a hot lead from a random one until they had already spent time on it.",
        "problem": "We needed not just an autoresponder but fast qualification at the entry. The assistant should itself, and in a living tone, clarify budget, the district of interest, property type and purchase timeline, and then understand whether the request is on target. Hot requests must go to the agent at once and flagged as priority, while off-target ones are closed gently, without burning the specialists time and without scaring the person off with a dry form.",
        "solution": "We built an assistant on WhatsApp that meets every ad lead within seconds. It asks a short, polite chain of questions, recognizes the client free-form wording and, based on the answers, matches the request against the property base. If the parameters fit and the person is ready to deal within a foreseeable term, the request is flagged as hot and goes to the agent along with a card: budget, district, matching options and the whole dialog history. Off-target requests are closed gently, and the contact is kept in the base for the future so the agent does not return to them.",
        "steps": "The launch took about four weeks. First, together with the head of sales, we described the criteria of a target request: budget thresholds, priority districts, property types and signs of readiness to deal. Then we linked the assistant to the property base and the CRM so the card would reach the right agent immediately. Separately, we tuned the tone of the questions so they sounded like a talk with a consultant, not an interrogation. After a trial on part of the ad traffic, we rolled the assistant out to the whole request flow.",
        "impact": "After launch, the share of viewings that really lead to a deal grew noticeably, because mostly prepared clients now reach the agent. Time to first contact dropped from several hours to a couple of minutes, and requests stopped cooling in the queue. Agents were relieved of sorting random requests and focused on what brings commission: viewings, negotiations and deal support. The ad budget started working more efficiently, since not a single paid lead is left without a fast reply.",
        "limitations": "The assistant deals only with qualification and the first contact. It does not discuss the final price, does not bargain over terms and does not handle the legal part of the deal: document checks, the deposit and the contract. All of that stays with the agent, and the assistant deliberately does not replace the property viewing and live negotiations.",
        "methodology": "We compared conversion from request to a scheduled viewing and time to first reply for 30 days before launch and 30 days after. We counted from CRM data on the same ad traffic. All figures are rounded and given as ranges so the agency cannot be identified from them.",
        "anonymityNote": "The agency name, property addresses and agent names are hidden under NDA. The messenger and ad tool names are given only to show the real integration stack, not as advertising for the services. The figures are given in aggregate and rounded, so a specific company cannot be worked out from them.",
        "faq": [
          {
            "q": "How does the assistant tell that a lead is on target?",
            "a": "It matches budget, district, property type and purchase timeline against the agency criteria and flags matches as priority. Disputed cases still go to the agent so a live person makes the call."
          },
          {
            "q": "What happens to off-target leads?",
            "a": "The assistant replies politely and keeps the contact in the base for the future without distracting the agent. Some of these clients ripen later, and then they are worked with directly."
          },
          {
            "q": "Does the assistant replace the agent at a viewing?",
            "a": "No, it brings the client to the meeting and passes the agent the full context. The viewing, negotiations and closing of the deal are handled only by a human."
          }
        ]
      }
    },
    {
      "slug": "freight-logistics-status",
      "url": "https://ainow.ge/projects/freight-logistics-status",
      "title": "Freight logistics: shipment status and intake",
      "industry": "Logistics",
      "slogan": "Status by reference number and new requests in one chat",
      "summary": "A logistics company was drowning in identical where is my cargo calls. Dispatchers were pulled onto them all day, and at peak times customers simply could not get through and grew frustrated, even though the status was already in the accounting system. We added an AI assistant on Telegram: it returns status by reference number in seconds and takes new orders in the same chat, carefully collecting the route, cargo type and contacts. The load of reference calls on dispatchers fell by about a third, and new requests stopped getting lost between channels.",
      "techStack": [
        "Telegram",
        "Custom CRM",
        "REST API",
        "Google Sheets"
      ],
      "metrics": [
        {
          "label": "Status calls",
          "before": "manual stream",
          "after": "about a third fewer"
        },
        {
          "label": "Reply to client",
          "before": "waited for a dispatcher",
          "after": "instant by number"
        },
        {
          "label": "New requests",
          "before": "got lost",
          "after": "all in the CRM"
        }
      ],
      "caseStudy": {
        "context": "A transport and logistics company with a constant flow of cargo and a large client base. Every customer wants to know exactly where their cargo is and when it will arrive, and is used to calling the dispatcher directly for the answer. The status information was already in the accounting system, but only an employee could pull it by hand. As a result, dispatchers, whose real job is planning routes and solving problems on the road, spent half the day answering the same kind of questions by phone.",
        "diagnostic": "We broke down the structure of requests over several weeks and saw a clear picture. The bulk of calls was a simple question about status, whose answer was already in the system and did not require the dispatcher to decide anything. At peak hours the line was busy, and clients with a new request or an urgent problem simply could not get through. And the new requests themselves came now to email, now to a messenger, now by phone, and some got lost without a single intake point.",
        "problem": "The task split in two. First, to give the client the cargo status automatically, by reference number and at any time of day, so as not to occupy a dispatcher with it. Second, to take new orders in the same window through a clear form and enter them into the system at once, so no request is lost. At the same time the assistant must honestly say when there is no data, rather than inventing the cargo location.",
        "solution": "We built an assistant on Telegram as a single window for freight clients. When a person sends a reference number, the assistant queries the accounting system through an integration and returns the current status in real time: where the cargo is, at what stage and the rough timeline. If the client comes with a new order, the same bot collects the route, cargo type and volume and contacts step by step, and then enters a structured request into the CRM. When there is no data for the number or the situation is non-standard, the assistant does not guess but switches the conversation to a live dispatcher.",
        "steps": "The project took about a month. First we connected the assistant to the status accounting system via an API, so answers were always taken from the source rather than from copies. Then we described the new request form: which fields are required and how they fit into the CRM. Separately, we walked through the scenarios where the bot is required to hand the dialog to a human. We launched carefully, first on regular clients who ask about status most often, and after a trial we opened the assistant to everyone.",
        "impact": "After launch, reference calls about status dropped by about a third, because clients began getting an answer in the chat within seconds and at any time, including nights and weekends. Dispatchers won back time for direct work: planning routes, monitoring trips and dealing with real problems. New requests stopped dissolving between channels, since they now have one entry point and a single format in the CRM. Clients feel calmer: the cargo can always be checked by themselves, without waiting for an answer by phone.",
        "limitations": "The assistant is responsible for information and taking requests, but it does not handle disputes. Claims about damaged or delayed cargo, cost recalculation and negotiations over a complaint it does not run: such requests go straight to the responsible manager with full context. The bot also does not change the route and does not confirm non-standard shipping terms.",
        "methodology": "We counted the share of reference status calls in the overall flow of requests to dispatchers for a month before launch and a month after. Additionally, we checked how many new requests reached the CRM without loss. The figures are averaged and given in ranges so a company cannot be identified from them.",
        "anonymityNote": "The company name, routes and employee names are hidden under NDA. The messenger and technical integrations are named only to show the real stack of the solution, not for advertising. All figures are averaged and rounded, so a specific carrier cannot be recognized from them.",
        "faq": [
          {
            "q": "Where does the assistant get the cargo status?",
            "a": "From the company accounting system by reference number, in real time through an integration. The client sees the same data as the dispatcher, but without waiting on the line."
          },
          {
            "q": "What if there is no data for the number?",
            "a": "The assistant honestly says so and switches the client to a dispatcher, rather than inventing the cargo location."
          },
          {
            "q": "Can a new shipment be arranged in the same chat?",
            "a": "Yes, the bot collects the route, cargo type and contacts step by step and enters the request into the CRM. A manager does the pricing and confirms the details."
          }
        ]
      }
    },
    {
      "slug": "fitness-club-retention",
      "url": "https://ainow.ge/projects/fitness-club-retention",
      "title": "Fitness club: membership renewals and reminders",
      "industry": "Fitness",
      "slogan": "Membership expiry reminders and renewal in one chat",
      "summary": "A premium fitness club was losing members for no real reason: the membership ended, no one reminded them to renew in time, and the person simply stopped coming. Administrators could not call the whole base, and without reminders churn grew toward the end of the term. We added an AI assistant on WhatsApp that writes the member itself a few days before expiry, offers convenient renewal options and records the decision in the CRM. Retention rose by several percentage points, and administrators stopped spending hours on manual calls.",
      "techStack": [
        "WhatsApp Business API",
        "Custom CRM",
        "Google Calendar"
      ],
      "metrics": [
        {
          "label": "Member retention",
          "before": "fell at term end",
          "after": "up several points"
        },
        {
          "label": "Reminders",
          "before": "manual calls",
          "after": "on time, to all"
        },
        {
          "label": "Renewal",
          "before": "a front-desk visit",
          "after": "in one chat"
        }
      ],
      "caseStudy": {
        "context": "A premium fitness club with a large base of regular members and memberships of various lengths. The core of the revenue here is not the first sale but the renewal: a member who renews on time stays for months ahead. But expiry reminders rested on the administrators, who already have plenty of work at the front desk. Calling everyone whose term is running out was physically impossible, and some members quietly dropped off.",
        "diagnostic": "We looked at how the base behaves toward the end of the membership term and saw a predictable dip. Churn grew noticeably in exactly the last days of the membership: without a reminder the person put off renewal, then stopped coming and in the end did not return. Manual calls covered only part of the base, depended on the administrator workload and did not scale at all. It turned out the club was losing already acquired and loyal members not because of service quality, but because of silence at the right moment.",
        "problem": "We needed to close the gap between the membership ending and the member decision to renew. The assistant should remind them in advance and without pressure that the term is coming to an end, offer clear options and let them renew right in the chat, without a trip to the front desk. It was important not to turn care into spam: one fitting message on time instead of a stream of mailings, with an easy way to opt out.",
        "solution": "We built an assistant on WhatsApp that watches the terms across the base and reaches out itself a few days before the membership ends. In a calm, friendly tone it reminds the member that the term is coming to an end and offers convenient renewal options based on which membership the person used. The decision is recorded in the CRM at once, and those ready to renew or wanting to discuss terms are passed by the assistant to a manager with the chat history. The message arrives once and on time, and opting out takes one tap.",
        "steps": "The launch took about three weeks. First we exported the terms of active memberships and synced them with the assistant so the reminder went out exactly when needed. Then we scripted the scenario: how many days ahead to write, which options to offer and how the message sounds, so it reads as care rather than a sale. We ran it on a small group of members, tuned the wording from real replies and only then switched reminders on for the whole base.",
        "impact": "After launch, retention rose by several percentage points, because members began renewing on time rather than remembering the membership a month after it ended. Administrators were freed from endless calls and switched to live work in the gym and at the front desk. Renewal turned into a short dialog in the messenger, which is more convenient both for the member and for the club. Importantly, the reminders did not cause irritation: people take one timely message as normal service.",
        "limitations": "The assistant is responsible for a timely reminder and processing the renewal, but it does not resolve individual disputes. Freezing a membership due to illness, refunds or recalculation for missed days it does not approve: such cases go straight to a manager. The bot also does not sell expensive personal packages, where a live talk with a trainer or manager is needed.",
        "methodology": "We compared retention and the share of renewals for two months before connecting the assistant and two months after. We counted from CRM data over seasonally comparable periods, so as not to confuse the effect with ordinary swings in attendance. All figures are rounded and given as ranges.",
        "anonymityNote": "We do not disclose the club name, addresses or member base under NDA. The messenger and systems are named only to show the real integration stack, not as advertising. The figures are rounded and given in ranges, so a specific club cannot be recognized from them.",
        "faq": [
          {
            "q": "Do the reminders feel intrusive?",
            "a": "The assistant writes once and in advance, in a calm tone, and opting out of the mailing takes one tap. It is taken as care, not as spam."
          },
          {
            "q": "Can a membership be renewed right in the chat?",
            "a": "Yes, the member picks an option in the chat, the assistant records the renewal in the CRM and, if needed, passes the data to the administrator. There is no need to go to the front desk."
          },
          {
            "q": "What about freezing and refunds?",
            "a": "Individual cases of freezing or refunds the assistant does not decide itself, but passes straight to a manager with the chat history so a human agrees the terms."
          }
        ]
      }
    },
    {
      "slug": "legal-firm-intake",
      "url": "https://ainow.ge/projects/legal-firm-intake",
      "title": "Law firm: client intake and routing",
      "industry": "Legal",
      "slogan": "Capturing the case essence and routing to the right lawyer",
      "summary": "A multi-practice law firm was losing requests in a shared inbox: they were not structured, some reached the wrong lawyer, and the first reply sometimes took more than a day. We built an AI assistant on WhatsApp and Telegram that carefully clarifies the essence of the case, determines the practice area and hands a structured card to the right lawyer with an urgency flag. The first reply sped up from a day to hours, and off-target questions are filtered out before they occupy a specialist.",
      "techStack": [
        "WhatsApp Business API",
        "Telegram",
        "RAG",
        "Custom CRM"
      ],
      "metrics": [
        {
          "label": "First reply",
          "before": "over a day",
          "after": "hours, not days"
        },
        {
          "label": "Routing",
          "before": "some misrouted",
          "after": "straight to the right lawyer"
        },
        {
          "label": "Lawyer focus",
          "before": "sorting a shared inbox",
          "after": "only relevant cases"
        }
      ],
      "caseStudy": {
        "context": "A law firm ran cases across several practice areas at once and received requests by email, WhatsApp and Telegram. Everything flowed into one shared inbox with no tagging by topic. A duty staffer sorted the stream, manually forwarding letters to colleagues. As a result a client facing a deadline could wait a full day just for the question to reach a lawyer of the right specialization, and often walked to another firm in that time.",
        "diagnostic": "We went through a month of incoming messages and saw that the problem was not volume but the lack of structure. Requests arrived without key details, so the duty staffer kept re-asking the same things: which practice area, are there deadlines, what exactly happened. Some letters were forwarded to the wrong lawyer and came back in circles, and some simply sat unread in the inbox. Every lost request is not only a missed fee but a reputational risk.",
        "problem": "The task was delicate: capture the essence of the case without making the client feel like a cold questionnaire. A legal request is often personal and anxious, and a dry ten-field form scares people off more than it helps. The assistant should clarify the practice area, the nature of the question and any deadlines in a human way, separate target requests from clearly-not-ours, and hand the lawyer a ready card rather than a raw stream of text. And it may give no legal assessment at all.",
        "solution": "We built an assistant that meets the client on WhatsApp and Telegram as a single reception desk. In a calm dialog it clarifies the practice area, the substance of the question and any deadlines, and along the way hints at which documents are worth preparing. It then determines the case profile, flags urgency and hands the lawyer a structured card with all the details. The conversation keeps the firm's restrained tone, the assistant gives no legal advice and promises no outcome, but honestly says a real lawyer will give the assessment.",
        "steps": "The project took about four weeks. First, together with the firm, we laid out the practice by area and described the routing rules: which signals lead to which lawyer. Then we connected the assistant to the CRM so every card lands in a shared log and does not get lost. Separately we agreed the boundaries: the assistant collects facts but does not consult. We ran the scenario on one channel, tuned the wording from real dialogs so the tone sounded caring, and only then opened intake on all channels.",
        "impact": "After launch the assistant handles the first pass of almost all requests itself. The time to a first substantive reply dropped from a day to a few hours, because the case goes straight to the right lawyer with a ready card instead of wandering the shared inbox. Off-target questions are filtered at the entrance and no longer take the specialists' expensive time. Instead of sorting mail, the lawyers work on cases themselves, and from the very first message the client feels heard.",
        "limitations": "The assistant deliberately does not consult. It does not interpret rules, assess a case outlook or name the cost of services: any legal conclusion stays with a live lawyer. As soon as a question calls for judgment rather than fact-gathering, the dialog goes to a specialist together with the full context of the correspondence.",
        "methodology": "We measured from the request log and CRM data for a month before launch and a month after stabilization. We compared the time to a first substantive reply, the share of correctly routed cases and the number of lost letters. All figures are rounded and given in ranges so the firm cannot be identified from them.",
        "anonymityNote": "The firm name, practice areas and lawyer names are hidden under NDA. The messengers and CRM are named only to show the real integration stack, not as advertising. We disclose nothing from the content of specific cases, and the figures are rounded to ranges.",
        "faq": [
          {
            "q": "Does the bot give legal advice?",
            "a": "No, it only clarifies the essence of the case and passes it to the right lawyer. Any legal assessment is given by a live specialist."
          },
          {
            "q": "How is the urgency of a request determined?",
            "a": "By signals in the message itself, such as document filing deadlines or a hearing date. Such cases are flagged as priority and reach the lawyer first."
          },
          {
            "q": "What happens to a request outside the firm's profile?",
            "a": "The bot politely explains that the firm does not handle such matters and does not waste a lawyer's time. The contact is kept in the base."
          }
        ]
      }
    },
    {
      "slug": "auto-body-repair-estimate",
      "url": "https://ainow.ge/projects/auto-body-repair-estimate",
      "title": "Auto body shop: photo estimate and booking",
      "industry": "Auto service",
      "slogan": "Damage estimate from photos and repair booking in chat",
      "summary": "An auto body shop was drowning in the same questions of how much it will cost and when they can come: mechanics broke off work to answer, and some clients left without waiting. We built an AI assistant on Instagram and WhatsApp that takes a photo of the damage, gives a clear preliminary price range and immediately offers a free time for inspection. The mechanics stopped being pulled into chats, and noticeably more clients now make it from a request to a visit.",
      "techStack": [
        "Instagram",
        "WhatsApp Business API",
        "Computer Vision",
        "Google Calendar"
      ],
      "metrics": [
        {
          "label": "Price reply",
          "before": "waited for a mechanic",
          "after": "range right in chat"
        },
        {
          "label": "Mechanics' time",
          "before": "pulled into chats",
          "after": "back on the cars"
        },
        {
          "label": "Chat to visit",
          "before": "some walked away",
          "after": "noticeably more arrive"
        }
      ],
      "caseStudy": {
        "context": "An auto body shop worked mainly through Instagram and WhatsApp and got dozens of requests a day. Almost every one started the same way: the client sent a couple of photos of a dent or a scratch and asked how much it would cost and when they could drop by. The mechanics had to answer straight from the shop floor, breaking off from painting and panel work. In busy hours the messages piled up, replies lagged, and some clients managed to message the shop next door.",
        "diagnostic": "Going through several weeks of chats, we saw that the overwhelming majority of questions come down to a rough estimate from a photo and a booking for an inspection. That needs no mechanic qualification, yet each time it yanks one out of work and breaks the pace in the shop. It also turned out that clients cool off fast: if a clear price range does not arrive in the first minutes, interest fades and the person goes where they got an instant reply. Here the speed of the first answer decides more than the price itself.",
        "problem": "We needed an assistant that recognizes a typical kind of damage from a photo and a short description, honestly names a preliminary price range and immediately offers a time for inspection. It was important not to over-promise: a photo estimate is always a range, not a final number, because hidden damage shows only on the lift. The assistant should sound confident but careful, explain that a mechanic will confirm the exact cost, and under no circumstances invent prices out of thin air.",
        "solution": "We built an assistant that meets the client on Instagram and WhatsApp, asks for an overall shot and a couple of close-ups and reads the nature of the damage from them. Drawing on a base of the shop's typical jobs and prices, it names a preliminary range and explains what the final figure depends on. Then it shows free windows and books the inspection into the calendar. The tone is friendly and without pressure: the assistant openly warns that a mechanic will confirm the exact sum after inspecting on the lift.",
        "steps": "The project took about five weeks. First we gathered examples of damage with real outcomes and sorted them into typical jobs with prices. Then we set up recognition of common cases from photos and tied the price ranges to that base. Separately we wrote honest caveats about hidden damage so the assistant would not over-promise. We connected booking to the calendar, ran the scenario on part of the Instagram inbound, tuned the wording from real dialogs and rolled the assistant out to both channels.",
        "impact": "After launch the assistant handles almost all first requests itself. The mechanics are no longer pulled into chats and keep the pace in the shop, while the client gets a clear price range and an offer to book within the first minutes. Thanks to the speed and clarity, noticeably more people now make it from a request to a visit: before, many dropped off at the waiting stage. The average ticket did not slip, because a mechanic still sets the final cost after inspection.",
        "limitations": "The assistant gives only a preliminary range from a photo and makes no claim to precision. Hidden damage, the state of the body geometry and the real scope of work show only at inspection, so a mechanic always names the final price. The assistant does not resolve disputed or complex cases itself but calls in a human right away.",
        "methodology": "We compared data for a month before launch and a month after stabilization: the speed of the first reply, conversion from request to visit and the share of messages that distracted mechanics. All figures are rounded and given in ranges, and the price ranges in the examples are illustrative so the shop cannot be identified from them.",
        "anonymityNote": "We do not disclose the shop name, address or photos of clients' cars under NDA. The messengers and tools are named only to show the real stack, not as advertising. The specific amounts in the examples are illustrative, and all figures are rounded to ranges.",
        "faq": [
          {
            "q": "How accurate is the photo estimate?",
            "a": "It is a preliminary range for typical jobs, not a final price. A mechanic names the exact cost after inspecting on the lift, when hidden damage is visible."
          },
          {
            "q": "What photos should I send?",
            "a": "An overall view of the damaged spot and a couple of close-ups from different angles. The assistant will suggest which shots to take if something is missing."
          },
          {
            "q": "Can I book an inspection right away?",
            "a": "Yes, after the preliminary estimate the assistant shows free windows and books it into the shop calendar. You can reschedule in the same chat."
          }
        ]
      }
    },
    {
      "slug": "furniture-maker-quote",
      "url": "https://ainow.ge/projects/furniture-maker-quote",
      "title": "Custom furniture: draft quote in chat",
      "industry": "Furniture",
      "slogan": "Collecting order parameters and a draft estimate in minutes",
      "summary": "A custom furniture maker got many requests about how much a wardrobe or a kitchen would cost, but estimates were calculated by hand over several days, and clients left for whoever answered faster. We built an AI assistant on WhatsApp that clarifies dimensions, material and configuration step by step and instantly returns a clear draft estimate from the price list. The price reply sped up from days to minutes, and managers now spend time only on clients ready for a measurement.",
      "techStack": [
        "WhatsApp Business API",
        "Google Sheets",
        "RAG",
        "Custom CRM"
      ],
      "metrics": [
        {
          "label": "Estimate reply",
          "before": "several days",
          "after": "minutes"
        },
        {
          "label": "Collecting parameters",
          "before": "by hand",
          "after": "step by step in chat"
        },
        {
          "label": "Manager focus",
          "before": "every request in a row",
          "after": "only measurement-ready"
        }
      ],
      "caseStudy": {
        "context": "The workshop made furniture to order: wardrobes, walk-in closets and kitchens fitted to a specific room. Requests came mostly on WhatsApp and from ads, and almost every one opened with a price question. The catch is that an honest estimate depends on dimensions, material, hardware and configuration, and collecting those parameters and calculating everything by hand took the manager up to several days. In that time the client managed to get an answer from a nimbler competitor.",
        "diagnostic": "We looked at what a request's path is made of and found two bottlenecks. First: the manager pulled the same intake details out of the client by hand every time, asking ten clarifications in the chat. Second: the price calculation itself was not hard but not fast either, and it was exactly the hot leads that got lost in the queue. In the furniture market clients compare several workshops at once, so whoever gave a clear figure first tended to win the order.",
        "problem": "We needed an assistant to take over the pre-sales routine: collect dimensions, material, facade type and hardware step by step and in a human way, then immediately calculate a clear draft estimate from the workshop's price list. The balance mattered: give the client a fast figure without turning it into a promise of an exact price, since the final sum is still confirmed at measurement. And the assistant must never invent materials or amounts that are not in the price list.",
        "solution": "We built a WhatsApp assistant that walks the client through a short, clear chain: what the piece is, the opening dimensions, the carcass and facade material, the hardware needed. Along the way it relies on the digitized price list and the workshop's calculation rules and instantly returns a draft estimate with a breakdown of what drives the sum. Interested clients it passes to a manager for an exact calculation and a measurement visit. The tone is calm and advisory, and any figures stay strictly within the price list, with no improvisation.",
        "steps": "The project took about a month. First we digitized the price list, materials and calculation rules so the assistant had a firm footing rather than guesses. Then we built the question flow to collect exactly what a draft estimate needs and not tire the client with extras. Separately we boxed the assistant in with hard limits: only items from the price list, no invented prices. We ran it on ad inbound, tuned the wording from real dialogs and rolled it out to the whole request flow.",
        "impact": "After launch the assistant handles the initial parameter gathering and issues a draft estimate on almost all inbound itself. The price reply that used to stretch over days now arrives in minutes, and the client gets a clear figure while their interest is still hot. Managers stopped being form-fillers and work only with those already ready for a measurement and a contract. Thanks to the speed, more requests reach the measurement stage, and the pre-sales load has dropped noticeably.",
        "limitations": "The assistant calculates only a draft range from the price list and does not take on complex custom projects. Curved facades, rare materials, non-standard mechanisms and fitting to uneven walls are assessed by a person at the measurement. The exact price is always confirmed by a manager, not the bot.",
        "methodology": "We compared a month before launch and a month after stabilization: the speed of the estimate reply, conversion from request to measurement and the share of requests that reached a manager warm. All figures are rounded and given in ranges, and the amounts in the example estimates are illustrative so the workshop cannot be identified from them.",
        "anonymityNote": "The workshop name, price list and project portfolio are hidden under NDA. The messenger, spreadsheets and CRM are named only to show the real integration stack, not as advertising. All amounts in the example estimates are illustrative, and the figures are rounded to ranges.",
        "faq": [
          {
            "q": "How accurate is the draft estimate?",
            "a": "It is a preliminary range from the price list and the given parameters, not a final price. A manager confirms the exact cost after measuring the space."
          },
          {
            "q": "What parameters does the assistant ask about?",
            "a": "The type of piece, the opening dimensions, the carcass and facade material and the hardware needed. That is enough to calculate a clear draft range."
          },
          {
            "q": "What about complex custom projects?",
            "a": "For non-standard pieces the bot does not calculate the price itself but offers a visit by a measurer and passes the request to a manager."
          }
        ]
      }
    },
    {
      "slug": "hr-recruitment-screening",
      "url": "https://ainow.ge/projects/hr-recruitment-screening",
      "title": "Recruitment: first-touch candidate screening",
      "industry": "Recruitment",
      "slogan": "Screening applicants by key criteria and booking interviews",
      "summary": "A recruitment agency received hundreds of applications for active roles, and recruiters could not reply to each one quickly: strong candidates cooled off in the queue and left for another employer. We built an AI assistant on Telegram that asks the initial questions itself, checks the answers against the role requirements and immediately offers interview slots to suitable candidates. The first reply now arrives in minutes, recruiters stopped drowning in repetitive correspondence and focus on live interviews.",
      "techStack": [
        "Telegram",
        "Custom CRM",
        "NLU",
        "Google Calendar"
      ],
      "metrics": [
        {
          "label": "First reply to a candidate",
          "before": "hours in a queue",
          "after": "in minutes"
        },
        {
          "label": "First-touch screening",
          "before": "manual, one by one",
          "after": "run by the assistant"
        },
        {
          "label": "Recruiter focus",
          "before": "flood of same questions",
          "after": "live interviews"
        }
      ],
      "caseStudy": {
        "context": "A recruitment agency with several recruiters ran searches for dozens of active roles at once for different employers. Popular positions drew hundreds of applications a day from many sources: job boards, social media ads and referrals. The small team physically could not process each one quickly, so applications piled up in a common queue with no unified tracking. By the time a recruiter reached a strong candidate, that person had often already interviewed at another company and stopped responding.",
        "diagnostic": "We pulled several weeks of the recruiters correspondence and looked at where the time went. It turned out that a large part of the day was spent on the same first questions: work experience, preferred schedule, salary expectations and readiness to start soon. These answers are easy to check against formal criteria, yet they were the ones eating hours. As a result the strongest candidates, who receive several offers at once, most often accepted someone else offer before their turn in the queue arrived.",
        "problem": "The goal was not to filter people out to save effort, but to stop losing strong ones to a slow reaction. The assistant should ask the initial questions itself in a calm, respectful tone, match the answers to the requirements of the specific role and immediately offer an interview time to a suitable candidate. A formal decline should sound polite and human, while any borderline profile must reach a recruiter rather than being rejected automatically. In essence, we needed to take the first-touch routine off the team and leave it the live assessment.",
        "solution": "We built an assistant on Telegram as a single point of first contact with candidates. In the dialog it introduces itself, explains the role and clarifies experience, schedule and expectations in turn, recognizing free-form wording rather than only buttons, and checks the answers against the role requirements from the base. For a suitable candidate the assistant immediately shows free slots and puts a meeting into the recruiter calendar, and passes the card with the answers into the CRM. The tone is set by examples of the agency real correspondence, so the exchange feels human rather than like a form.",
        "steps": "The project took about four weeks. First, together with the recruiters, we described the criteria for each group of roles: what is mandatory, what is desirable and what disqualifies right away. Then we set up the questions themselves, the answer-checking logic, polite decline wording and the handoff of borderline cases to a human. We first launched on several roles, listened to real dialogs, tuned the tone and wording and only then connected the assistant to the whole flow of applications.",
        "impact": "After the run-in the assistant handles the first touch on almost all applications. A candidate gets the first reply in minutes at any time of day rather than after hours of waiting, so strong profiles no longer cool off in the queue, and suitable ones immediately pick a convenient interview slot. Recruiters stopped sorting the same first messages and spend their time on what truly needs a human: live interviews and the final assessment. The share of candidates who reach an interview, meanwhile, grew noticeably.",
        "limitations": "The assistant does not issue a hiring verdict and does not assess soft skills, cultural fit or motivation: that stays with the recruiter at the live interview. It does not reject borderline profiles but flags them and passes them to a human with the full answers. The team, not the bot, always makes the final call.",
        "methodology": "We measured from the application log and CRM data for a month before launch and a month after stabilization. We compared time to first reply, time to a scheduled interview and the share of candidates reaching an interview. All figures are rounded and given as ranges so a specific agency or search cannot be identified from them.",
        "anonymityNote": "The agency name, the employers and the specific roles are hidden under NDA. The messenger and service names are given only to show the real integration stack, not as advertising. All figures are presented in aggregate, in rounded ranges, so the company or its clients cannot be identified from them.",
        "faq": [
          {
            "q": "Could the bot filter out a strong candidate?",
            "a": "The assistant does not reject borderline or disputed profiles but flags them and passes them to a recruiter for a manual check. Only those who clearly fail the mandatory requirements get an automatic decline."
          },
          {
            "q": "How is the interview scheduled?",
            "a": "For a suitable candidate the assistant immediately shows free slots and puts the meeting into the recruiter calendar. The invite and a reminder come in the same chat."
          },
          {
            "q": "Does chatting with a bot feel impersonal?",
            "a": "The tone is set by examples of the agency real correspondence, so the assistant communicates respectfully and in a human way. As soon as a question goes beyond the scenario, a recruiter takes over the dialog."
          }
        ]
      }
    },
    {
      "slug": "cleaning-company-scheduling",
      "url": "https://ainow.ge/projects/cleaning-company-scheduling",
      "title": "Cleaning company: intake and area-based quote",
      "industry": "Cleaning",
      "slogan": "Area-based price and scheduling in one chat",
      "summary": "A cleaning company took requests in three places at once: by phone, on WhatsApp and on Instagram, and some simply got lost when passed between staff. Clients, meanwhile, wanted to know the price and a free time straight away rather than after a long back-and-forth. We built an AI assistant on WhatsApp that clarifies the area, cleaning type and address, calculates the price by the rates and puts the visit into the schedule at once. Requests now gather in one base and no longer get lost, and agreeing the details takes a couple of messages.",
      "techStack": [
        "WhatsApp Business API",
        "Google Calendar",
        "Google Maps",
        "Custom CRM"
      ],
      "metrics": [
        {
          "label": "Price and time quote",
          "before": "long back-and-forth",
          "after": "a couple of messages"
        },
        {
          "label": "Request tracking",
          "before": "lost between channels",
          "after": "one base"
        },
        {
          "label": "Requests handled",
          "before": "some slipped away",
          "after": "noticeably higher"
        }
      ],
      "caseStudy": {
        "context": "The cleaning company worked with private flats and small offices and took requests across three channels at once: calls, WhatsApp and Instagram Direct. Requests were written down however staff could: in notes, inside chats, on paper at the dispatcher desk. When they were passed between employees and shifts, some simply got lost. Clients, meanwhile, wanted the simplest thing: to hear the price and a free time straight away, but instead they were told the team would check and call back, and waited a long time for a reply.",
        "diagnostic": "We walked through every channel and saw the overall picture: there was no single list of requests, so the same request could be worked by two people while another was picked up by no one. Requests were lost especially painfully in the evening and on weekends, when the dispatcher was already off the line. A lot of time went into calculating the price by hand: an employee re-asked the area and cleaning type, dug through the rates and only then named a price. Because of this delay some clients left for whoever answered faster.",
        "problem": "The task was to bring every request into one point and remove manual recalculation. The assistant should collect a request in a clear scheme, calculate the preliminary price itself by area and cleaning type, estimate the trip by address and immediately offer a free window in the schedule. It was important not to let the bot fantasize about prices: it calculates strictly by the company rate grid and does not promise what is not in the price list. And every request from every channel should land in a shared log so nothing gets lost.",
        "solution": "We built an assistant on WhatsApp as a single entry for requests. In the dialog it clarifies the area, the cleaning type (maintenance, deep, after renovation) and the address, estimates the travel zone on the map, calculates a preliminary price by the company rates, shows free windows and puts the visit into the schedule. It gathers every request into a shared base, so a request from Instagram or by phone also lands in the same log. The tone is human, and the bot takes the price only from the price list and does not invent it.",
        "steps": "The project took about three weeks. First we digitized the rates and rules: how much a square meter costs for each cleaning type, how the trip is counted by district, what surcharges exist. Then we set up the calculation, the crew schedule, the collection of requests into a single base and linked the channels so that requests from the phone, WhatsApp and Instagram landed in one log. We ran the scenario on part of the requests, tuned the wording and calculation logic from real dialogs and rolled the assistant out to the full flow.",
        "impact": "After launch, requests from all channels gather in one log, so they stopped getting lost when passed between shifts, and the share of handled requests grew noticeably. The client sees the preliminary price and a free time straight away, at any hour, and the agreement shrank from a long back-and-forth to a couple of messages. The dispatcher no longer recalculates the cost by hand or merges requests from three places: they confirm visits and deal with complex cases. The average ticket, meanwhile, did not drop, because the calculation still runs strictly by the rates.",
        "limitations": "The assistant prices only typical sites accurately. Non-standard areas, heavy soiling, industrial cleaning and disputed square footage it does not estimate blindly, but offers a free estimator visit and passes such a request to a human. It also does not approve individual discounts above the rates: a manager decides that.",
        "methodology": "We measured from the request log and CRM data for a month before launch and a month after stabilization. We compared the share of handled requests, the time to agree the price and booking, and the number of lost requests. All figures are rounded and given as ranges so a specific company cannot be identified from them.",
        "anonymityNote": "The company name, addresses and rates are hidden under NDA, and the prices in the examples are illustrative. The messenger, maps and calendar names are given only to show the real integration stack, not as advertising. The figures are presented in aggregate and rounded so a specific company cannot be identified from them.",
        "faq": [
          {
            "q": "How is the cleaning price calculated?",
            "a": "The assistant takes the area and cleaning type, applies the company rate grid and, accounting for the trip to the address, shows a preliminary price right away, without waiting for a dispatcher. A manager confirms the final cost if the site is non-standard."
          },
          {
            "q": "What about non-standard sites?",
            "a": "The bot does not price complex or atypical cases blindly: it offers a free estimator visit and passes the request to a human. That way the client gets a fair price rather than a randomly inflated or lowered one."
          },
          {
            "q": "Will a request from Instagram or by phone also not get lost?",
            "a": "Yes, all channels are merged into one base, so a request from any source lands in the shared log and does not get lost when passed between shifts."
          }
        ]
      }
    }
  ]
}