The AI chatbot that turned roadside questions into sales

StartAI reports that its AI sales agent deployed for Serbian roadside-assistance company Atlas24h handled 7,559 conversations in July 2026, produced 2,776 qualified leads and 1,705 contracts confirmed against Atlas24h's production system, attributing €113,722 in revenue to those sales. The lead-to-contract rate rose from 42.9% in June to 61.4% in July after in-chat purchases were enabled, though StartAI's case study notes there was no randomised control and conversation volume also grew sharply. The vendor-published figures replace a prior web chatbot that captured zero leads and zero measurable revenue, and raise questions about attribution and what counts as revenue.

10 min read Atlas24h connected its chatbot to purchases. StartAI reports €113,722 from 1,705 contracts, raising questions about attribution and what counts as revenue. TL;DR - Serbian roadside-assistance company Atlas24h replaced an informational chatbot with an AI sales agent that identifies intent, checks eligibility, collects order data and either completes the purchase or passes a qualified lead to sales. - Vendor StartAI reports 7,559 conversations, 2,776 qualified leads and 1,705 contracts confirmed against Atlas24h’s production system in July 2026, producing €113,722 in attributed revenue. - The lead-to-contract rate rose from 42.9% in June to 61.4% in July after in-chat purchases were enabled, but there was no randomised control and conversation volume also grew sharply. - The copyable lever is to connect a conversational surface to one bounded transaction and verify the result against the system of record instead of counting chats as conversions. A driver with a broken car does not need a very articulate FAQ. They need the right plan, a completed contract and a tow truck when things go wrong. Atlas24h https://atlas-servis.com/en/ had already put a chatbot on its website. According to its software vendor, the bot answered questions about coverage, prices and renewals, then watched the customer leave. It collected no leads, could not finish a purchase and generated no revenue that the company could measure back to web chat. The company did not need a chatbot with nicer manners. It needed the conversation to do the next piece of work. Atlas24h and Novi Sad software company StartAI rebuilt the channel around intent, qualification, data collection, action and verification. The result is a useful test of where conversational AI conversion comes from when the chat box stops being a destination and becomes part of the sales system. StartAI’s case study https://www.startai.tech/en/case-studies/atlas24h-chatbot/ reports that the system handled 7,559 conversations in July 2026, created 2,776 qualified leads and produced 1,705 contracts that Atlas24h’s production system confirmed. The vendor attributes €113,722 in revenue to those sales. The figures are detailed and commercially useful. They are also vendor-published and come without a randomised control. The old bot stopped at the useful bit Atlas24h has provided roadside assistance in Serbia and across Europe since 2008. Its company profile https://atlas-servis.com/en/about-us/ says it has more than 150 employees, over 200 tow trucks positioned around Serbia and European coverage through a partner network. Customers buy annual packages that vary by vehicle, geography and the help included. That makes the website conversation more than a support exchange. A visitor may need to distinguish a national plan from one covering the rest of Europe, confirm eligibility, supply a registration plate and address, and activate a contract. The old bot could explain the products but could not move the customer through those steps. StartAI says the baseline for web chat was zero captured leads and zero measurable revenue. That is a channel baseline, not the company’s overall sales baseline. People could still buy through the website, app, telephone or human sales team. The experiment asked a narrower question. Could a conversation that already attracted purchase questions carry the customer far enough to produce a contract the operating system recognised? The agent received one job and two exits The Atlas24h AI sales agent first identifies whether someone wants a new package, a renewal, an order update or general information. For purchase intent, it checks eligibility and collects the contact, vehicle and package details required for a contract. From there, the customer can leave by one of two useful doors. An eligible buyer can complete the in-chat purchase flow without waiting for a salesperson. Someone who is interested but does not finish becomes a structured lead for human follow-up. A follow-up email covers people who supply contact details and leave before the transaction is complete. Those branches require different outcome measures. Many conversational systems measure containment, the share of chats that end without a person, as though every human handoff were a failure. Atlas24h kept the handoff as a commercial path. In July, StartAI reports that 771 people entered the direct purchase route and 559 bought, while 2,005 qualified leads entered follow-up and 1,146 later became confirmed contracts. The technical boundary came after the conversation. Each claimed sale was checked against Atlas24h’s contract database through a production API attribution step. That prevented a form submission, an agreeable chat or a model’s success label from becoming the revenue number. The system of record decided whether the sale existed. July changed the action as well as the volume The published sequence begins in May 2026. StartAI reports 1,847 conversations, 458 qualified leads and 196 confirmed sales, equal to a 42.8% lead-to-contract rate. June brought 4,598 conversations, 1,301 leads and 558 sales, while the conversion rate stayed almost flat at 42.9%. In July, conversation volume reached 7,559 and qualified leads reached 2,776. The reported contract count rose to 1,705 and lead-to-contract conversion to 61.4%. StartAI says July was when the chatbot gained the ability to complete purchases inside the conversation. That timing makes the in-chat action the most interesting intervention. It also stops short of a clean causal result. July had more than four times May’s conversation volume, and the public case does not explain traffic sources, marketing spend, seasonality, offer changes or how users were assigned. Roadside-assistance demand can move with summer travel. A month-to-month chart cannot tell us what July would have looked like without the new purchase route. The vendor gives a second comparison. Atlas24h recorded 2,177 sales from its human call centre in July 2025. Total sales reached 3,905 in July 2026, of which StartAI attributes 1,811 to its chat and voice systems. Human call-centre output was 2,094, down 3.8%. That supports the claim that the AI channels added measurable volume without a similar increase in human output, though it still does not isolate the chatbot or control for changes between the two years. The result is attribution before it is causation The strongest number is not the 61.4% conversion rate. It is the match between a conversational lead and a contract in the production system. That turns AI chatbot revenue from a model-reported outcome into a checkable business event. The weakest number is the zero-to-€113,722 story. The old informational bot produced no measurable web-chat revenue because it had no sales path or attribution chain. Once the new system captured contacts, handed leads to sales and logged completed purchases, revenue became visible. Some of that revenue may be incremental. Some customers may have bought by phone or another route anyway. The case publishes no holdout that separates the two. The funnel also mixes direct and assisted outcomes. Only 559 of the 1,705 July contracts closed entirely inside chat. Another 1,146 followed a human sales path after the agent captured the lead. Both are valid commercial outcomes, and they answer different operating questions. Direct sales test whether the agent can complete a bounded transaction. Assisted sales test whether it can supply salespeople with leads salespeople can follow up. StartAI built the software and published the figures. Atlas24h chief operating officer Marios Skutelis is quoted in the case, but the public material does not include raw logs, contract values, confidence intervals, cancellation rates, gross margin or an outside audit. It also leaves the model, retrieval setup, guardrails and human review policy undisclosed. The copyable lever is smaller than the agent A smaller company does not need to reproduce StartAI’s model stack. It can copy the operating sequence. 1. Choose one conversation with clear purchase intent and a transaction narrow enough to audit. 2. List the eligibility checks and fields that a person currently gathers before the order enters the system. 3. Give the agent two explicit exits, a completed transaction and a structured human follow-up. 4. Match every claimed outcome to the CRM, contract database or payment system before counting revenue. 5. Keep a holdout or phased rollout so the team can measure sales that would not have happened anyway. 6. Track cancellations, complaints, bad eligibility decisions and human rework beside conversion. The experiment should split users before engagement. Comparing people who chose to chat with people who did not will flatter the treatment because their intent differs. Randomly offer the sales agent to one group and the informational experience to another, then measure confirmed contracts per eligible visitor, not contracts per qualified lead alone. Keep direct and assisted sales separate and deduplicate anyone who crosses channels. That is where this case differs from Sterling Insurance’s controlled quote experiment https://industrycontents.com/ai-insurance-sales-agent-quote-conversion/ . Sterling kept a comparison while its agent intervened before abandonment. Atlas24h publishes a richer transaction funnel and stronger outcome verification, but a weaker counterfactual. Put the two together and the next test is obvious. Give an action-taking agent a production endpoint and keep a control that never receives it. A Balkan use case makes language part of the system The European relevance is direct. Atlas24h and StartAI are based in Novi Sad, Serbia, and the sale concerns roadside coverage across Serbia and Europe. StartAI says its platform handles Serbian, Croatian, Bosnian, Slovenian and English, including regional speech patterns and code switching. In a fragmented market, the workflow cannot assume that the customer will ask, qualify and buy in the English used to design the first prototype. That does not prove the language layer produced the lift. It explains why the architecture is more than a generic chat widget. The transaction has to preserve the customer’s language, selected package, eligibility facts and order data all the way into the operating system. The model can hold the conversation. Growth comes when the rest of the stack remembers what the conversation was for. Sources and disclosure - StartAI case study on the Atlas24h sales chatbot https://www.startai.tech/en/case-studies/atlas24h-chatbot/ , reporting July 2026 production results - Atlas24h website https://atlas-servis.com/en/ and company profile https://atlas-servis.com/en/about-us/ - Atlas24h on LinkedIn https://rs.linkedin.com/company/atlas24h - StartAI company profile https://www.startai.tech/en/about and StartAI on LinkedIn https://www.linkedin.com/company/startaitech StartAI supplied the system and published the case. Atlas24h is identifiable and its chief operating officer is quoted, while the reported funnel and revenue have not been independently audited in the sources reviewed here. FAQ What did Atlas24h change in its chatbot? It replaced an informational bot with an agent that identifies purchase intent, checks eligibility, collects contract data and either completes an eligible purchase or sends a structured lead to the sales team. How much revenue did StartAI report? StartAI reports €113,722 from 1,705 contracts confirmed against Atlas24h’s production system in July 2026. Of those contracts, 559 closed through the direct in-chat route and 1,146 followed human sales follow-up. Did the agent cause the reported revenue? The case supports attribution to leads and purchases that passed through the system. It does not establish incrementality because there was no randomised control, traffic volume changed sharply and the vendor does not disclose seasonality or other commercial changes. What can another growth team copy? Connect one bounded conversational task to a real transaction, preserve a human follow-up route and verify every claimed sale in the operating system. Then keep a control so attributed revenue can be separated from sales that would have happened anyway. Compare more measured AI experiments in the Growth Signal Index https://industrycontents.com/ic-lab/growth-signal-index/ and the Growth Signal archive https://industrycontents.com/category/growth-signal/ .