AI Department, Not AI Chatbot

A contractor in the Mid-South hired an out-of-state AI agency last summer to install a chatbot on his website and a booking assistant tied to his phone line. The promise was simple: never miss another after-hours lead. The pitch deck looked great. The demo worked. The contract was twelve months.

Three weeks in, the bot started telling callers it had booked them an appointment when no event was actually being created on the calendar. Customers showed up at his office on days nobody was expecting them. Other customers got booked, then got conflicting confirmation texts an hour later because a separate workflow had assigned them to a different slot. By month two, his receptionist was manually reviewing every booking that came through, which defeated the entire point of the install.

When he reached out to the agency, the response was that the bot was “learning” and the issues would smooth out. They were not smoothing out. He cancelled the contract early, ate the cancellation fee, and went back to taking calls himself.

That story is not unusual right now. Most of the AI agency pitches landing in the inbox of a small business owner today are selling some version of that contractor’s experience. A single bot, with a system prompt, hooked into one part of the business, with no team behind it and no way to catch its own mistakes.

This post is about why that pattern keeps happening, and what to ask for instead.

A Single Agent Is a Single Point of Failure

The reason most AI agent deployments break is structural. They are designed as one model trying to do everything: read the context, decide what to do, execute the action, write the response, and check its own work. When any one of those steps goes wrong, the whole thing goes wrong, and there is no second set of eyes inside the system to catch it.

The failures are usually quiet. The bot does not throw an error message. It confidently tells the customer their order has shipped when it has not. It books an appointment that never lands on the calendar. It quotes a price that does not exist in the catalog. It tells a lead the business does something it does not actually offer. From the outside, everything looks fine. The bot is responsive, the dashboard is green, the conversation looks coherent. The work just is not getting done.

These failures are hard to detect from the inside of a single-agent system because the agent that is supposed to do the work is also the agent that reports on the work. There is no Critic in the loop. The bot grades its own homework.

Real teams of humans do not work that way. A good operations team has roles. Someone gathers the information. Someone plans the work. Someone executes. Someone reviews. Someone communicates the result. That separation of concerns is what makes the team reliable, because failure in one role is caught by another.

AI agent teams should be built the same way.

The Seven Roles of a Real AI Team

There is a useful taxonomy emerging in the AI agent space that breaks the work into seven roles. Most production-quality AI deployments are some combination of these.

Learner. Reads the world. Pulls in research, source documents, customer history, market context. In SMB work, this is the agent that looks up the lead before drafting a response, or pulls the contractor’s last three invoices before quoting a new job. Without a Learner, the rest of the team is operating on whatever happened to be in the prompt.

Planner. Decomposes a request into steps. Decides which tools to use and in what order. For a service business, this is the agent that decides whether a new lead gets a same-day callback, a quote request, or a routing to the owner. Without a Planner, every task either gets a one-shot answer or spirals into trial and error.

Doer. Executes the actual work. Writes the email, creates the calendar event, posts the message, updates the CRM record. Most of what people call an AI agent is actually just a Doer with no team behind it. That is why those deployments break.

Tool Operator. Handles external integrations. Knows how to call APIs, format inputs, parse outputs. The unsexy plumbing that determines whether the work actually lands in the systems you care about. When a chatbot says it booked an appointment but no event hits the calendar, this is the role that failed.

Critic. Reviews what the Doer produced. Checks for hallucinations, off-brand language, missed steps. Without a Critic, every output goes to production unreviewed. This is the role most demos skip and most production agents desperately need. It is also the single biggest difference between a system that works for the first thirty leads and a system that still works at lead three thousand.

Supervisor. Coordinates the team. Catches stalls, manages handoffs, decides when to escalate to a human. The role that keeps the team from quietly drifting into failure modes. When something unusual happens, the Supervisor is what flags it instead of letting it slip through.

Presenter. Packages the result for the human who asked. Translates work-product into the report, message, summary, or update that lands in the inbox or chat thread. The interface between the team and the human running it.

A real production AI deployment for a small business probably needs four or five of these. A single chatbot has, at most, a Doer and a Presenter. That is why it breaks.

How Vectis Runs Its Own AI Team

Vectis Studio runs its own internal AI team built on this principle. Two agents are live now, with a third in design.

Alfred. Chief of staff. Supervisor and Presenter. Coordinates the team, surfaces what matters in a daily morning briefing, schedules and triggers the work that the other agents do. Every weekday, Alfred delivers a structured rundown of what changed overnight in the news that affects the business, what opportunities are time-sensitive, what action items are queued, and what to watch. If you have ever worked with a great executive assistant who actually ran your day, that is Alfred.

Q. Researcher. Learner and Tool Operator. Sweeps news sources, GHL changelogs, local civic news, competitive intel, and Anthropic product updates twice a day. Pulls signal from noise. Writes structured intel reports that Alfred uses for the briefings. When the city council releases its zoning agenda, Q sees it in the morning sweep. When a competitor launches a new pricing tier, Q catches it on the Monday competitive scan. When a tool we use changes its API, Q flags it the day it happens.

Draper. Content. Doer, Critic, and Presenter for written work. In design now, coming online soon. Reads the intel that Q produces, drafts content in the Vectis voice, checks it against brand and quality rules, and packages it for distribution. Named for the most quintessential ad man in fiction. The agent is the craft, not the character.

That is three agents covering the seven roles. Each one specialized, each one accountable for a defined set of work. When something goes wrong, you can trace which role failed and fix that specifically. When something goes well, you know which role earned it.

Two agents are not enough to handle every kind of work. Three is closer. Five is a real department. The point is not the count. The point is that each role has someone accountable for it, and no role is asked to do work that belongs to a different role.

This is the difference between an AI tool and an AI department.

What This Looks Like for a Small Business

Most small businesses do not need three named agents with names from spy fiction. They need one or two coordinated systems that handle a real category of work. Here is what that can look like for a home services contractor in plumbing, HVAC, or roofing, the kind of business that Vectis works with most often.

A new lead comes in through the website, Google Business Profile, or a referral text.

The Learner reads the lead. Looks up the address, checks if the contact exists in the CRM, pulls history if so. If it is an existing customer, Learner pulls the last three jobs and any open invoices. If it is a new contact, Learner runs a quick search for the address to confirm it is in the service area.

The Planner decides the next move. New customer with a leak gets a same-day callback. Existing customer with a quote request gets a different sequence. After-hours emergency gets escalated to the owner’s phone with priority alert. Routine inquiry from outside the service area gets a polite decline with a referral.

The Tool Operator creates the contact in the CRM, drops them in the right pipeline stage, and triggers the right workflow. If the Planner decided this is an emergency, the Tool Operator also fires the alert to the owner. If it is a quote request, the Tool Operator pulls the latest pricing sheet and prepares the variables the Doer will need.

The Doer drafts the response, whether SMS, email, or a scheduled callback prompt. Uses the customer history that the Learner pulled, the path that the Planner picked, and the current pricing that the Tool Operator prepared.

The Critic checks it. Right tone, right service, right address, no hallucinated pricing, no scheduling conflicts, no promises the business does not keep. If anything looks off, the Critic flags it for human review instead of letting it go out.

The Supervisor decides whether the response goes out automatically, gets queued for human review, or pings the owner directly because something is unusual. New customer with a routine question is automatic. Existing customer with a high-value quote request is queued for owner approval. Any kind of complaint, billing dispute, or unusual phrasing is escalated immediately.

The Presenter logs what happened in a format the owner can review at end of day. Not a wall of conversation logs, but a structured summary: this many leads, this many automatic responses, this many flagged for review, this many escalated, this many pending.

That is a system that can handle a hundred leads a day reliably. A single chatbot trying to do all of that without a team behind it will work for the first thirty leads and then start dropping them in ways that are hard to trace until a customer calls to complain about an appointment that never made it onto anyone’s calendar.

The cost difference between the two approaches is smaller than you would think. The reliability difference is enormous.

What to Ask the Next AI Agency That Pitches You

If someone is selling you an AI agent, the answer to a few specific questions tells you what you are actually buying.

Ask which of the seven roles their system covers. If they cannot name them, the system is one bot with a clever prompt. If they can name two or three, you are getting closer to a real deployment.

Ask what happens when the bot makes a mistake. If the answer is “the model will improve over time,” the system has no Critic. Mistakes will compound. Ask specifically: who or what reviews the bot’s output before it reaches the customer?

Ask how you will know if it stops working correctly. If the answer is “you’ll see it in the dashboard,” find out what the dashboard actually measures. A green dashboard with a hallucinating bot is the worst of both worlds. The Critic and Supervisor roles exist specifically to catch the failures that look fine on the surface.

Ask what happens when something unusual comes in. A complaint, an angry customer, a request that does not fit the script. If the bot just handles it, that is a problem. The Supervisor role exists to catch unusual cases and route them to a human before damage is done.

Ask who you call when something breaks at 9 PM on a Saturday. The agency that sold the contractor in the opening of this post had a support email and a 48-hour response window. That is not enough when a customer is showing up at the wrong time on a Sunday morning.

If they describe a single bot with a clever prompt, you are buying a Doer with no team. It will work in the demo. It will start drifting in production. When it breaks, you will be the one explaining it to your customer.

If they describe a coordinated set of roles with handoffs, review steps, and accountability for each piece of work, you are buying a system that can run reliably. You will spend a little more upfront. You will spend much less on cleanup.

This is the choice. Not whether to use AI. The choice is whether what you put in front of your customers is a single bot with a system prompt, or a real department that can do the work and catch its own mistakes.

Vectis builds the second one.