Skip to main content
Kainotomia Innovation Technologies Contact us

Rules, limits and records for AI agents

An agent that can act needs limits. We set what it may touch, where it must stop for a person, and how every action is recorded.

It covers permissions, approval points, guardrails and action logs.

Standard +27001ISO/IEC certified delivery on every build

What you can do

Permissions

Each agent gets its own identity and the least access it needs.

Approval points

A person signs off anything large, unusual or impossible to undo.

Guardrails

Inputs and outputs are checked against your rules.

Action logs

What the agent did, with what, and why.

The flow, end to end

01List possibleactions02Rate each by risk03Give each agent anidentity04Set limits andapprovals05Filter inputs andoutputs06Log and review
  1. List possible actions
  2. Rate each by risk
  3. Give each agent an identity
  4. Set limits and approvals
  5. Filter inputs and outputs
  6. Log and review

What you receive

Named outputs. Each one is yours to keep.

01Action and risk register
02Permission design
03Approval workflow
04Guardrail rules
05Action log
06Review routine

What you decide, and what we need

You decide

  • Which actions need a person
  • Spending and rate limits
  • Who reviews the logs
  • When to switch an agent off

We need from you

  • The agent design or the running agent
  • Your risk policy
  • Named approvers

How we measure it

Agreed before the build. Reported after every change.

01Actions inside their limits
02Approval and override rates
03Blocked inputs
04Time to switch off

What buyers ask us

Banks, operators and public bodies ask the same six questions.

"Can we test your AI with our own data?"

Yes. Every system ships with a test set built from your cases, and the report is yours.

"Where does our data go?"

Nowhere. Models run on your servers, a private cloud or a national cloud that you choose.

"Who is accountable for a decision?"

A person on your side. The system prepares the work. People approve what affects rights or money.

"Will we see when the model changes?"

Yes. Updates are tested on your cases first, and recorded.

"What if it fails?"

A fallback procedure is agreed before launch, and rehearsed.

"Are you certified?"

Our delivery is certified to ISO/IEC 27001 and ISO 9001:2015. We do not hold ISO/IEC 42001.

What we will not promise

A vendor who promises these has not run a model in production.

  • An AI system that never makes a mistake
  • A system nobody can attack
  • Delivery in a fixed number of weeks before we have seen your data
  • A saving we have not measured
  • Bangla results equal to English without testing them
  • AI that needs no human oversight

How it works

  1. List the actionsEverything the agent could do, ranked by risk.
  2. Set limitsAccess, spending and rate limits per agent.
  3. Place approvalsWhere a person must sign off.
  4. Record and reviewLogs, alerts and a regular review of overrides.

What can go wrong, and what we do about it

Hidden instructions in a document

Inputs are filtered, and the agent cannot act outside its tools.

Approvals become a rubber stamp

Override and approval rates are tracked and reviewed.

Agents share one login

Each agent has its own identity.

How we run the project

Working software every two weeksIn your environment, not on a slide.
One delivery leadA certified lead who owns scope, time and quality.
Written scopeWhat is in, what is out and what it costs.
Security from the first sprintISO/IEC 27001 controls on every build.

Where it fits

We tell you before the project starts, not after.

Where it works well

Teams putting agents into real workflows with real systems.

Where it works badly

Chatbots that only answer questions and take no action.

What you'll get

  • A working system on your own data
  • An accuracy report on your own cases
  • Access rules and a full audit log
  • Training for the people who will use it
  • Documents your auditors can read

How it starts

Find out if it is worth building

A short, fixed-scope review of your data, systems and rules.

  • Interviews with the people who do the work
  • Review of data and systems
  • A ranked list of use cases with effort and cost
  • A written recommendation
About this stage

See it working on your own data

A working system for one use case, built on your real data.

  • One agreed use case
  • A working system, not slides
  • An accuracy or performance report
  • A costed plan for production
About this stage

Go live in your environment

The full build: integration, security testing, training and go-live.

  • Integration with your systems
  • Security testing before launch
  • Training and handover documents
  • Go-live support
About this stage

Keep it running and keep it safe

Monitoring, patches, model updates and reports for as long as the system runs.

  • Monitoring and alerting
  • Patches and updates
  • Model and data updates
  • Regular reports for audit
About this stage

Where we have done it

What you should know

Does our data leave our network?

No. The model runs on your servers, a private cloud or a national cloud. We agree the location in writing before work starts.

Which models do you use?

Open models and commercial ones. We pick per workload and explain the choice, the hardware and the running cost in writing.

How do you know the answers are right?

We test the system on your own cases before launch. You get the accuracy report. Unclear cases go to a person.

Who makes the final decision?

A person. The system prepares the work and shows its source. People approve anything that affects rights or money.

Tell us what the system has to do

A first call takes 30 minutes. You speak to an engineer, not a salesperson. We reply within one working day.

Book a consultationinfo@kainotomia.tech