What a vertical AI agent actually does

Almost every company that says it has AI in customer service today actually has a chatbot with a well written prompt. It reads naturally, it sounds helpful, and it stops at the exact moment the customer asks for something that requires looking at a system: where is my order, move my appointment, cancel this subscription.

The vertical agent exists to get through that wall. The difference is not the language model — it is the same model. It is three things the chatbot does not have: tools, memory and ownership of a process.

Tools: what separates answering from resolving

A tool is a function the agent is allowed to call. Look up an order by order number. Create a deal in the CRM. Check open slots in a calendar. Issue an invoice.

When a customer asks where their order is, the chatbot produces a plausible paragraph about shipping times. The agent with a tool opens the order, reads the tracking status and answers with the real position of the package.

In the conversation the difference looks subtle. In the operation it is enormous: one resolves the problem, the other postpones it. And postponing has a price — the customer comes back, now annoyed, and now to a human.

The part that gets underestimated: a tool needs preconditions. An agent allowed to book without checking whether the calendar exists will schedule into thin air. An agent allowed to discount without checking margin will discount where there is none. The quality of a vertical agent lives far more in the rules around the tools than in the text it produces.

That is also where most of the build time goes. Writing the prompt is an afternoon. Deciding what happens when the inventory API returns stale data, when the customer gives a partial order number, or when two systems disagree about the same record — that is the work.

Memory: without it, every conversation starts from zero

A language model remembers nothing between one conversation and the next. If the customer said yesterday that they moved, they will have to say it again today.

Memory here does not mean keeping the chat history — that is context, and it fits in the model window. Memory is what the company knows about that person: what they bought, what they complained about, which plan they pay for, what the last interaction was and how it ended.

That is what lets the agent open the conversation already knowing who it is talking to, instead of asking the same three screening questions the customer answered last week.

It also changes what the agent is allowed to assume. Recognising a returning customer is useful; acting on that recognition without confirming identity is a data protection problem waiting to happen. Phone numbers change hands, shared inboxes exist, and a household often has one device. Recognition informs the conversation, it does not authorise disclosure.

Ownership: the agent answers for a process, not for a reply

This is the part that changes how you measure.

A chatbot is judged by how good the answer is. A vertical agent is judged by what happened to the process: how many conversations were resolved without a human, how long they took, how many turned into revenue, what each resolution cost.

That implies the agent needs to know when to stop. An agent that insists on solving what it cannot solve is worse than one that escalates quickly. The gates that matter most in production are not the ones that make it act — they are the ones that make it stop:

• Irreversible action limits. Cancelling a subscription, issuing a credit note, refunding money: either it goes through human confirmation, or it has a very explicit ceiling.
• A volume circuit breaker. If the agent has sent too many messages to the same contact in a short window, something is looping. It stops on its own.
• Escalation on uncertainty. When tools fail in sequence, the right move is to hand the case to a person with the context already assembled — the boundary rules are written out in the post on the sales pipeline run by agents — not to try a fourth time with different wording.
• An audit trail. Every tool call logged, with a trace identifier. Without it, when something goes wrong, nobody can reconstruct what happened.

Vertical means it knows your business

A generic agent knows how to talk. A vertical agent knows how to talk about a domain, with the vocabulary, the rules and the systems of that domain.

The retail agent understands carts, delivery windows, return policy and the real catalogue behind the storefront — the ground covered in agentic ecommerce. The clinic agent understands insurance coverage, specialty, procedure duration and the difference between a first visit and a follow up. Each one with the tools of its own world. That is why the same platform produces very different results depending on how much of the domain was actually modelled.

It is also why vertical agents deliver faster than AI for everything projects: the scope is narrow enough for the rules to be written down and tested.

Where to start without getting stuck

The most common mistake is starting with the most visible process instead of the most repetitive one. Opening with exception handling — the hard case, the angry customer — picks precisely the ground where AI has the least advantage.

Start with volume and repetition: order status, scheduling, first line screening, qualifying the lead that came from an ad. These are processes where the correct answer can be derived from a system, and where the difference between answering in ten seconds and answering in two hours moves conversion.

With that running and measured, scope grows on its own — and it grows on evidence rather than opinion. The same logic applies on the buying side, where a machine is increasingly the one reading your data: that is the subject of agent commerce.

The second most common mistake is treating the agent as a technology project instead of a process change. Whoever runs the operation has to be in the room from day one, because that person knows the real exceptions — the customer who always asks for a discount, the order that disappears at the carrier, the rule nobody ever wrote down. An agent designed only by people reading documentation gets the happy path right and breaks on every edge case, which is exactly where human support spends its time.

It also pays to agree in advance on how the agent will fail. Not whether — how. The practical question is: when the integration goes down, does the agent tell the customer, stay silent, or call a human? All three answers are defensible in different contexts, and choosing beforehand keeps you from discovering the answer on the worst possible day.

What to take from this

A vertical agent is not a better chatbot. It is a piece of the operation that is now executed by software that talks. That changes who answers for the result, changes what you measure, and changes the design: most of the work is in the tools, the preconditions and the gates, not in the prompt.

When someone shows you a demo of AI holding a great conversation, there is only one useful question: what can it make happen in the systems? If the answer is nothing, it is a chatbot — and chatbots already existed in 2016.