When your company rules change, who updates the AI? 3 vendor clauses
You edit the rules, the vendor updates the AI, so the contract sets how changes reach it and what the AI does meanwhile.
Key points
- List every company rule the AI follows
- One master copy in-house; versions dated
- Name a notifier; the vendor also checks
- Until updated: go on, caveat or hand off
Contents
List the rules the AI follows before you set fix times
A&A perspective
If a vendor builds an AI that follows your company's rules, list every rule the AI follows before you write into the maintenance contract how many days the vendor has to fix an error once it is found. By an AI that follows your company's rules we mean one that answers or processes requests based on your own documents, such as your refund policy or HR rules.
Hypothetical example
Say a vendor built you an AI that answers refund questions, and a few months have passed. Your team changed the refund conditions, but nobody told the vendor. The AI keeps answering on the old conditions, and you first hear about it through a customer complaint.
From the sources
Anthropic, the company that makes the AI model Claude, has published a customer story about Decagon, a North American software company that provides customer-support AI to businesses. The page says Decagon's customers deploy AI agents that follow their own policies and practices.
From the sources
The page credits higher customer satisfaction to Claude's ability to follow complex instructions and policies exactly. On refunds, Decagon's co-founder says the agent does not just explain the steps: it checks whether the customer is eligible and processes the refund through the payment system.
A&A perspective
Automate & Augment (A&A), which writes this article, reads these two statements together. An AI that follows rules exactly will, once a rule is out of date, follow the old conditions exactly too. If it is built to complete the refund, the error is not a wrong answer but a wrong refund that has already gone through.
A&A perspective
A clause on days to fix only starts to matter once someone notices the error. Until then, the AI keeps completing work on the old conditions. That is why you first list the rules the AI follows: the refund policy, HR rules, style rules, and even the table of who may approve what.
A&A perspective
Decagon's page does not discuss the contract question of who fixes things when the rules change. The way the three clauses below are split is A&A's own framing.
Keep one master copy in-house and have versions dated
A&A perspective
The first clause states, for each rule you listed, where its master copy lives, which version the AI follows now, and what the AI's scope covers.
A&A perspective
The master copy is the one version your company treats as current. You decide its single home: a set file in a shared folder, or a set page on the company wiki. "Ask the person in charge" is not a place.
A&A perspective
Have the vendor keep the version the AI currently follows inside the deliverable. If you do not number your versions, the vendor stores the date you sent it and the file itself. If you later disagree over whether it was sent or received, you can show which version the AI was running on.
A&A perspective
For scope, write down what the AI follows and what it does not. For example, the refund policy is in, while one-off exceptions that staff grant on the spot are out. Without a scope, your staff cannot tell which document changes they need to report to the vendor.
Hypothetical example
For a refund-answering AI, for instance, put two sentences like these in a schedule to the contract. This is a made-up example, not a real engagement. "The rules this AI follows are limited to the Returns and Refunds Policy the Company provided on 1 October 2026 and revised versions received through the contact point named in this agreement. One-off exceptions granted by staff and operating notes shared only internally are excluded."
From the sources
Anthropic has also published a story about brand.ai, which offers a system that uses AI to keep a brand's rules consistent. One co-founder says that at a previous company they waited almost two years for an agency to produce a brand book. The other says today's brands cannot be contained in a static guideline document.
A&A perspective
Both pages are customer stories that Anthropic wrote from what the customers told it, and neither shows a publication date.
A&A perspective
A&A reads this as saying that rule documents take time to produce and keep changing after they are written. If your rules are not written down, there is no master copy to point to. In that case, before any maintenance, someone has to write up how things run today and get it approved inside the company. If the vendor does that work, have it quoted as part of the initial build.
| Rule the AI follows (made-up example) | How changes are passed on and checked | AI behavior until updated |
|---|---|---|
| Returns and refunds policy | Notice: the revised version goes to the vendor's contact point 7 days before it takes effect. Check: the vendor compares the current documents quarterly | Hand off |
| HR rules (leave and internal procedures) | Notice: the HR lead sends each revision to the contact point. Check: compare the whole rulebook once a year | Caveat |
| Style rules | Notice: whenever they change. Check: every six months | Go on |
| Approval authority table | Notice: whenever the organization changes. Check: quarterly | Hand off |
| Parts tied to the law | Notice: you or your legal adviser turn the change into rule wording and send it to the contact point. Check: none (the vendor does not interpret the law) | Hand off |
Notify the vendor of changes, and have it check too
A&A perspective
The second clause says that rule changes reach the vendor by two routes: you notify the vendor, and the vendor comes and checks.
A&A perspective
If you rely only on your own notices, some notices never go out. The refund policy is often revised by finance or customer support, and that team may not know that an AI reads it. It is a gap with no ill will or negligence behind it, so punishing someone will not make it smaller.
A&A perspective
For the notice route, write who notifies, by what means, and how many days before the change takes effect. You name the notifier by role, not by name. On the receiving side, use the vendor's contact point named in the contract rather than one person's email, so that the route survives when someone leaves.
A&A perspective
Also define what counts as notice. A passing remark in a meeting does not count; notice is given when the revised document itself reaches the contact point.
A&A perspective
On the checking route, the vendor gets your current documents on a schedule and compares them. Set the frequency by how often each rule changes: for example, quarterly for the refund policy and the price list, and once a year for HR rules. Have this check included in paid maintenance.
From the sources
Rippling, a back-office platform for small businesses, replaced its automated decision-tree logic with a Claude-powered agent, according to Decagon's page. The agent guides clients through HR and payroll processes while ensuring compliance with local regulations, the page says.
A&A perspective
As this example shows, what the AI follows can include both your own rules and the law. Deciding which change in the law affects a given process is legal interpretation. So agree that, for anything tied to the law, you or your legal adviser turn the change into rule wording and send it to the vendor's contact point. Also state that the vendor does not interpret the law.
Pick what the AI does until a change is built in
A&A perspective
The third clause states, rule by rule, what the AI does while a change has not yet been built in: go on, caveat, or hand off. Even with two routes, the gap between a rule change and the AI catching up never goes away.
A&A perspective
"Go on" means the AI keeps answering on the old version. Use it for rules where a lag only leaves some old wording behind, such as style rules.
A&A perspective
"Caveat" means the AI keeps answering but adds one line to answers that touch that rule: "Please check the latest conditions with our staff." Log the answers that carried the line, so that you can later count how many were affected.
A&A perspective
"Hand off" takes only the requests that touch that rule away from the AI and routes them to your staff. It suits refunds, where money moves, and approvals that cannot be undone.
A&A perspective
If you let the vendor build without choosing, the AI in practice runs on "go on". The refund case at the start was not the result of choosing that; it was the result of writing nothing.
A&A perspective
Also write what happens when the vendor's check does not happen on schedule. For example: if a quarter's comparison has still not been done when the next quarter begins, refund requests move to "hand off".
A&A perspective
The third clause is both a contract term and part of how the AI is built. If you write "hand off", the vendor has to build a branch into the AI that spots requests touching that rule and routes them to staff. You write the clause and the vendor builds the AI, so this is where the two tend to drift apart.
A&A perspective
Some rules, such as promotion terms that change month to month, move faster than either route can keep up with. For those, it can be quicker to keep the rule out of the AI and put it in a table your own staff can edit directly.
Summary
A&A perspective
Before you set fix times, list the company rules the AI follows, from the refund policy to the approval table. In the first clause, keep one master copy in-house and have the vendor keep each version the AI follows, dated.
A&A perspective
In the second clause, name who notifies changes and have the vendor also check the current documents on a schedule. In the third clause, choose for each rule what the AI does until a change is built in: go on, caveat, or hand off.
Errors in an AI that follows your company's rules happen while a rule has changed but the change is not yet in the AI. Before setting fix times, list the rules the AI follows. Decide the master copy's home and who notifies changes yourself, and in the maintenance contract ask the vendor for a scheduled check and for what the AI does until a change is built in.
FAQ
If we ask the vendor to check for changes too, will maintenance cost more?
The checking work becomes part of paid maintenance. A&A has not measured how much it changes the cost. Some work stays with you: listing the rules, setting where master copies live, naming the notifier, sending revised documents to the vendor's contact point, joining the scheduled checks, turning legal changes into rule wording, and choosing per rule what the AI does until a change is built in.
Does the same clause cover updates that the vendor makes to the AI?
Use a separate clause. There the change starts from the vendor's build, not from your documents. Whether to redo acceptance each time the AI is updated is covered in "Re-acceptance after AI updates: what your vendor tests, what you check".
Will these three clauses reduce complaints or disputes with the vendor?
We cannot say that they will. A&A has not measured how maintenance contracts that spell out an update route relate to disputes or cancellations. What the three clauses settle is who does what when a rule changes.
Sources & editorial note
Primary pages read for this article. Publication dates below belong to the sources; access dates record our research.
- Brand.ai uses AI to make brands more human with Claude
Anthropic · undated customer-story page; no publication or update date shown
Accessed 2026-10-22 - Decagon delivers white-glove customer service at scale with Claude
Anthropic · undated customer-story page; no publication or update date shown
Accessed 2026-10-22