Putting An AI Agent Development Company Into Live Channels Our AI Agent Development Company delivered an agent that worked well in a test interface. Its first week on live chat produced a complaint, a regulatory question, and a customer who typed the same thing eleven times. Channels are where agents meet reality. Here is what we learned about deploying through them, and what I now specify to any AI Agent Development Company A channel is a set of constraints, not a destination The framing we lacked at first. Chat allows a turn by turn conversation with short responses and visible typing. Email allows length and arrives asynchronously. Voice has no scrollback and no ability to present options visually. A messaging platform has its own formatting and delivery rules. We built once and deployed everywhere, which produced an experience designed for none of them. Decide how the handover to a person works Before anything goes live. Ours transfers with the full conversation, a summary of what was attempted, and a flag saying why the agent stopped. The customer is told clearly that they are now with a person. Our first version transferred with no context at all, so the customer repeated everything. That is worse than no agent, because it adds a step to an already frustrating path. Make the escalation obvious and always available Not hidden behind failed attempts. Every interaction offers a route to a person from the first message. About eleven per cent of users take it immediately and that is a perfectly reasonable choice. Our earlier design required two failed attempts before offering escalation. The complaint in week one was about exactly that. Tell people what they are talking to Plainly, at the start. Ours says so in the first message. There is no ambiguity and nobody has objected to the disclosure itself. Several jurisdictions are moving towards requiring this and it is the right thing regardless. Any AI Agent Development Company proposing an undisclosed deployment is proposing a problem. Handle repetition as a signal The eleven times customer taught us this. A person repeating themselves is telling you the agent is not understanding. That is detectable and should trigger an escalation rather than another attempt. Our rule now escalates after two substantially similar inputs. That one change removed most of our worst interactions. Scope what the agent may discuss Narrowly and explicitly. Ours handles order status, delivery changes, returns, and account details. Anything else goes to a person immediately rather than being attempted. The regulatory question in week one was a complaint about a financial product. The agent attempted it, politely and wrongly, and that category is now a hard transfer. Assess the channel before deploying to it Part of a wider readiness question. An AI agent readiness assessment should cover each channel separately. What the volume is, what the current handling time is, who monitors it, and what the regulatory position is for that medium. Our voice channel failed that assessment and we have not deployed there. Recording, consent, and the absence of scrollback made it a different problem and we did not pretend otherwise. Keep one source of truth across channels Otherwise you run several products. The same question asked in chat, email, and a messaging platform should produce the same answer, from the same knowledge, with the same policy applied. Our first deployment had separate configurations per channel, which drifted apart within two months and produced two contradictory answers to the same question. Measure containment honestly The number most often reported misleadingly. Containment means the customer's issue was resolved. It does not mean the conversation ended without a transfer, which is what most reporting counts. We measure resolution by whether the customer contacted us again within seven days about the same matter. That number is substantially lower than our transfer avoidance rate and it is the one we act on. Watch the first week properly With people present. Our second deployment had two team members reading live conversations for the first three days, with the authority to disable the agent instantly. They used that authority once, for about forty minutes, while a routing problem was fixed. An AI Agent Development Company that does not propose this is underestimating the first week. What our channel deployment requires ● A design per channel, not one design deployed everywhere. ● Transfer with full context and a stated reason. ● An escalation route offered from the first message. ● Clear disclosure that the customer is talking to software. ● Escalation triggered by repeated similar input. ● An explicit topic scope with hard transfers outside it. ● A readiness assessment run separately for each channel. ● One shared knowledge and policy source across all channels. ● Containment measured by repeat contact, not by transfer avoidance. ● Live human monitoring with instant disable authority in week one. Common questions Should an agent handle every channel? No. Our voice channel failed assessment and remains human handled, which was the right call. How should escalation be offered? From the first message, always available. Hiding it behind failed attempts produces complaints rather than containment. Must you disclose? We think so regardless of jurisdiction, and the regulatory direction supports it. No customer has objected to the disclosure. What does containment actually mean? Resolution, measured by repeat contact. An AI Agent Development Company reporting transfer avoidance is reporting a different and more flattering number. Does this change the supplier conversation? It should. Generative AI Development Services proposals often cover the capability and not the channel, and the channel is where the first week's problems come from. Who designs the escalation experience? Your service team, with the supplier implementing. Generative AI Development Services suppliers do not know what your contact centre can absorb. What is the most common early failure? Transfers without context. The customer repeats everything and the agent has made the experience worse than it was before. Where we are now Two channels live with a design each, disclosure in the first message, escalation always available, and containment measured by whether people came back. The agent itself is the same one that worked in the test interface. Everything that went wrong in week one was about the channel it had been placed into.