Written procedures used to be a reference that experienced staff consulted and quietly corrected in their heads. Once an AI agent does the work, they are what it acts on. Every gap, contradiction and stale step becomes a decision made without a person's context. Fix the procedures and you fix a large share of agent errors before they happen.
Why were procedure defects tolerable before?
Because people compensated. A procedure says to check a document "is recent"; staff know that means within three months. Two procedures give different cut-off times for a payment; staff know which one the team actually follows. A step refers to a form that was replaced two years ago; staff use the new one. None of this is written down, and none of it causes problems, because the people doing the work carry the corrections.
New starters learn the corrections from colleagues. Auditors occasionally find a gap between procedure and practice and raise a finding, which gets fixed or accepted. The system works, imperfectly, because judgement fills the holes.
What happens when an agent meets the same defects?
It does not have the corrections. It has the text. Each kind of defect turns into a characteristic kind of error.
| Defect in the procedure | What a person does | What an agent may do |
|---|---|---|
| Gap: a case the procedure does not cover | Asks a colleague or uses judgement | Improvises, or refuses every such case |
| Contradiction between two documents | Follows the one the team uses | Picks one inconsistently, case by case |
| Stale step: references a retired form or system | Uses the current one | Follows the stale step, or fails |
| Vague term: "recent", "significant", "where appropriate" | Applies local convention | Applies its own reading, which may vary |
| "Use judgement" with no criteria | Uses experience | Has no basis; outcomes vary |
| Unwritten rule known to everyone | Follows it | Does not know it exists |
The errors are not random. They repeat across every case that hits the same defect. That makes them both more damaging and easier to find, if you look.
How do you find the defects before the agent does?
Read the procedures as if you had no other knowledge of the process, because that is the agent's position. For each step, ask whether someone with only this document could carry it out correctly on every case. Mark every vague term, every reference to another document, every "use judgement" and every place where you know the team does something different.
Then test the procedures against policy. A procedure can be internally consistent and still conflict with the policy it is meant to implement. Agents can help with this comparison, as described in our guide on checking procedures against policy with an AI agent, though a person must decide how each conflict is resolved.
Finally, use shadowing. Run the agent alongside people on live work, compare outcomes, and trace every difference back to its cause. Many will turn out to be procedure defects rather than agent faults.
What should the agent do when instructions conflict?
Some conflicts will survive any review. The agent needs a rule for them, and the rule should usually be to stop. If two sources the agent relies on say different things about an action, the safe behaviour is to hold the case for a person and record which sources conflicted, rather than guess. Our piece on how an AI agent should handle conflicting instructions sets out the options.
Each such hold is also a signal. A cluster of holds caused by the same conflict tells the procedure owner exactly which documents to reconcile.
What does this look like on a real process?
Take the repair of rejected payment instructions. A payment comes back from the receiving bank with a reason code. The procedure says to correct the instruction and resubmit, or return it to the originator. Experienced staff know which reason codes can be repaired from the customer record, which need the customer's confirmation, and which should never be repaired because they suggest fraud. Very little of that is in the written procedure.
An agent given that procedure as written will repair some instructions it should have returned and return some it could have repaired. Writing down the reason codes, the source each repair may draw on and the codes that must always go to a person turns the procedure into something an agent can follow. Our piece on whether an AI agent can repair rejected payment instructions works through the details.
Who owns the procedures now?
The process owner, with real time allocated. Procedures are no longer occasional reference; they are live inputs to an agent acting at volume. They need version control so you know which version the agent used on any given day, an owner who approves changes, and a review whenever a product, policy or system changes.
There is a related but distinct idea in regulation. Under the EU AI Act, Article 13 requires providers of high-risk systems to give deployers clear instructions for use. Those are the provider's instructions about the system, not your procedures for the work. Deployers need both: the provider's instructions for operating the system, and their own procedures for the process it runs. Our explainer on Article 13 transparency and instructions for deployers covers the provider side. This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What is the trade-off?
Rewriting procedures is unglamorous and slow. It surfaces disagreements that teams have lived with for years, forces decisions about which unwritten rules are real, and produces longer, more precise documents that some staff will find harder to read.
The alternative is to let the agent discover each defect by making the same mistake many times. Before an agent goes live on any process, have the process owner read its procedures as the agent will, fix the gaps and contradictions, write down the unwritten rules and put the documents under version control. The deployment approach that builds this into the plan is set out in our guide to deploying an AI agent: train, shadow, certify, widen.
Questions people ask
Should procedures be rewritten specifically for the agent?
Better to rewrite them so they serve both people and the agent. A single precise procedure avoids two versions drifting apart, and people benefit from the clarity as much as the agent does.
Can an AI agent help rewrite procedures?
Yes. It can draft clearer versions, flag vague terms and find contradictions between documents. A person who knows the process must review and approve every change, because the agent cannot know which unwritten rules are real.
How often should procedures an agent relies on be reviewed?
On a set schedule, and immediately after any change to the product, policy, system or regulation they describe. An agent acting on an outdated procedure repeats the outdated step on every case.