The checks arethe product.
Getting an AI model to say something useful once is easy. The work is making it do the right thing every day, show its sources, hand over when unsure and leave a record. This is how we do that, taken from systems we run.
- Comes inA document, a client question or a supplier feed arrives.
- SortedThe agent works out what kind of job it is and sends it down a fixed, tested route.
- PreparedIt drafts the work from your own records and approved sources, and notes where every figure came from.
- CheckedCode checks the figures and the rules. Anything odd is held for a person, never guessed.
- Signed offSomeone on your team approves, edits or rejects it.
- LoggedWhat came in, what the agent did and who approved it are all recorded.
The model does the reading and the drafting. Fixed code does the routing and the sums. A person makes the decision.
Who does the work
A senior builder directs AI coding agents from Anthropic and OpenAI. One agent builds, a different agent reviews, and a person approves every release. It lets a small team build carefully and quickly, and every step leaves a trail.
Design
Decide what the agent may do before it does anything.
- We map the approval points first: what the agent may prepare and what only a person may decide.
- Safety rules are written as exact thresholds. For example, a new product whose price is more than three times the cost from a supplier with stock stays unpublished until a person checks it.
- Personal details and requests for personal advice are stopped by code before any AI model sees them.
- Bulk changes have hard limits, and records are archived rather than deleted.
- Before client data goes in, we write a data protection impact assessment that lists every field, provider and risk.
Build
AI reads and explains. Code calculates.
- We design agents so that figures come from your records or from fixed, tested code rather than a model’s memory.
- We limit what each step of an agent can see to the data that step needs.
- Sensitive fields are encrypted inside the database, on top of the storage encryption.
- Audit logs are append-only: the database refuses any change or deletion.
- Website enquiries are saved before they are emailed, and deleted automatically after 12 months.
Test
Test sets written to trip the agent up.
- Each agent has a set of real-style cases, run through the same path the public uses and graded by hand.
- The set includes attempts to trick it, personal questions it must refuse and questions outside its job.
- Weak answers are written down and fixed, not hidden. We say plainly what a test does and does not prove.
- Our rule is to test on made-up data, not live client data. Changes to live systems run first as a dry run that lists what would change.
Review
The agent that builds a change does not approve it.
- A second AI agent reviews each change. Findings are fixed, usually with a test that proves the fix, and reviewed again until they pass.
- Reviewers check results on the live system, not just the code. One review recomputed 9,661 prices against the rule and found no differences.
- A blocking finding stops the release.
Release
Every release can be undone.
- Staging first. Production only after a person says go.
- Before a release we record the live version and the version to roll back to.
- We deploy exactly the build that was reviewed, and check its file fingerprints to prove it.
- AI features have an off switch that stops them without deleting anything.
- Rollbacks are a person’s decision, not an agent’s.
Watch
Watched after release.
- After a release, read-only checks run against the live system for several minutes.
- Live test questions catch what automatic checks miss. They caught one faulty release, rolled back in five minutes.
- Daily checks compare live data with the rules, such as every price against the pricing rule.
- Incidents are written down with a rule to stop them happening again, and the log is read before releases.
Cost
Cost limits from day one.
- Public AI features have per-visitor and daily limits.
- The cheap path comes first: code and reviewed answers handle what they can before any model is called.
- Budgets are agreed before anything scales. We learnt this when an early pilot used up its AI credit in testing.
Data
Where your data lives is agreed before we build.
- We write down which services store and process your data, and where. Our recent client systems keep their data in Cloudflare’s EU region, with nightly backups.
- Some AI providers process requests in the United States. We tell you which, and send them only what each step needs.
- Passwords and keys are kept out of code and logs, and checked by name only.
Sign-off
A person decides.
- Our rule: nothing goes live, and nothing is sent outside the business, without a person’s go.
- What an agent flags goes to a person to decide, rather than being fixed automatically.
- Specialist wording, such as tax guidance, is for the client’s own experts to review.
Want to see the working?
In a 30-minute call we can walk through how an agent for your firm would be tested, approved and watched.