How we build
Why AI setups break in week two

The demo was perfect. The first week went fine. Then, somewhere in week two, an agent filed a document in the wrong place, or missed a new kind of request, or stopped working because a form changed overnight. Nothing dramatic. Just the first sign that a real business is messier than a demo, and the start of the question most setups skip: who maintains AI for a small business once the builder has moved on?
This is normal, and it is not a reason to give up on AI. It is a reason to plan for the keeping, not only the building.
Why does a demo work when a real week does not?
A demo runs on the cases someone thought of. A real week brings the ones nobody did. The caller who speaks Spanish. The client who replies to a three-week-old email with a new question. The document scanned upside down. The request that is half one thing and half another.
A demo is also run by the person who built it, who knows which questions to ask and which to avoid. Your clients do not know that, and they should not have to.
An agent, in plain words, is a piece of AI software with one job and a written rulebook. It is good at the cases its rulebook covers. When a case falls outside the rulebook, a well-built agent hands it to a person. A badly kept one guesses. Over time, the gap between the rulebook and the real week grows, unless someone closes it.
What actually breaks?
Three kinds of change, and none of them is about the AI getting worse.
Your tools change. A web form gets a new field. A carrier updates its portal and moves a button. A vendor renames a menu. A password expires. Each one can quietly stop an agent from doing a step it did yesterday.
Your work changes. You add a service. Your prices change. A new staff member does things a slightly different way. A new kind of request starts arriving that the rulebook never mentions.
The rules change. Regulators and bar associations keep writing guidance about AI. In February 2026, the Oregon State Bar issued Formal Opinion 2026-208 on chatbots and AI agents: a lawyer should be notified of every AI intake conversation, and a person should interview the prospect before the firm takes the case. An intake setup built before that opinion needed a fresh look afterward. Every regulated trade gets its own version of this.
What happens if nobody maintains it?
The agent drifts. It keeps working on the cases it knows and gets a little more wrong on the ones it does not. Nobody notices at first, because nobody is looking. Then a client complains, or a request sits unanswered for two days, and the team decides AI does not work.
A drifting intake agent can be worse than none, because you believe the phone is covered. For a law firm, the stakes are clear. In Clio's 2024 secret-shopper test, only 40% of law firms picked up when a prospect called. An agent that answers every call is a real change. An agent that answered every call until a form changed last Tuesday is not.
Who maintains AI for a small business?
Someone has to be the keeper: the person responsible for the agents after launch. The keeper's job is not exciting, and that is the point. It comes down to four habits.
- Rulebooks. Every agent works from a written rulebook: what to do, what never to do, when to hand off. When something changes, the rulebook changes first, then the agent.
- Test sets. A set of real past cases with the right outcome written next to each one. After any change, the agent runs through the whole set before it goes back to live work. If an old case now comes out wrong, the change waits.
- Logs. Every action an agent takes is written down. The keeper reads them for surprises: cases handed to a person, odd answers, steps that took longer than usual.
- A weekly check. Once a week, the keeper reviews the numbers, reads the hard cases, checks what changed in the business and fixes what drifted.
In a big company, this is someone's job title. In a firm of two to ten people, it is usually nobody's, because everyone already has a full week.
What does this look like at a small law firm?
Take the setup in our law firm case study, an example injury firm with one attorney. Agents answer after-hours calls, gather the facts into the intake system, check conflicts against the firm's records, write a one-page memo for the attorney, and send the fee agreement only once she approves it.
Now picture the changes a normal year brings. The intake form gets new questions. The firm starts taking a new kind of case, with different screening questions. The bar issues new guidance on AI intake. The attorney rewrites the decline letter. Each change means an updated rulebook, a few new cases in the test set, and a check before it goes live. Caught on a Friday, none of it touches Monday's calls.
The attorney's part does not change through any of it. She decides whether to take the case, approves the agreement and gives the advice. The keeper makes sure the agents keep handing her those decisions, and nothing else.
Can't the software vendor maintain it?
Vendors maintain their product. They do not maintain your rules. They do not know your intake questions, your price book, your carriers or the decline letter you rewrote last month. When their update changes a screen your agent relies on, nobody calls you to warn you.
The same goes for the AI models behind the agents. They get updated too, and a new version can answer the same question a little differently. The test set is how you catch that before a client does.
You can keep the agents yourself, if someone on your team has regular time each week and the habit of checking. Many small firms do not have that time. That is the honest reason the keeper's role exists as a service.
Why the monthly plan exists
This is why the setup is only the start of how we work. After the build, we stay on as your AI officer on a monthly or yearly plan: a morning note of what the agents did, a Friday review of the week, fixes when something drifts or changes, and one improvement a month chosen with you.
The Friday review is where you see the keeping happen: what the agents did that week, what they handed to you, and what was fixed, such as a form that changed and the agent that was updated to match. You read it in a few minutes and pick next month's improvement. The agents still never commit the business. Prices, advice, signatures, payments and anything regulated wait for a person. The full sequence is on our how it works page.
Case studySeven AI agents run a law firm's intake through the night
A crash at 11:40 PM: the call is answered, the facts and conflicts are checked, a one-page memo waits for the attorney, and the client signs by 8:41 AM.View case study
Questions owners ask
Can I maintain the agents myself?
You can, if someone on your team has regular time each week and the habit of checking. The work is reading the logs, updating the rulebooks when something changes, and running the test set before any change goes live. In most small firms, that time is the first thing to go in a busy week.
What happens when an agent makes a mistake?
Agents are built so mistakes are caught before they cost much: drafts wait for approval, and anything that commits the business goes to a person. The keeper reads every case the agent was unsure about, fixes the rulebook, and checks the fix against the test set before it goes live.
What does the monthly plan include?
A morning note of what the agents did, a Friday review of the week, fixes when something drifts or changes, and one improvement a month chosen with you. It starts after the setup and runs monthly or yearly.

