How we work

AI-native speed. Production proof.

In short

We use AI across research, specifications, implementation, testing and operations, but we don't outsource judgment to it. Every release has a source, an owner, a verification path and a way back, and anything consequential stops at a named human gate before it ships.

Four steps, every project

  1. Map the operation

    Follow the work as it actually happens: inputs, systems, decisions, exceptions, owners and failure modes. The output is a source-linked operational model, not a generic process diagram.

  2. Name the boundary

    Choose one consequential path with a measurable outcome. Define what the system may automate, what requires human authority, and what would stop the build.

  3. Build the vertical slice

    Connect the real systems, implement the narrow path end to end, and expose the operating state. AI is used where it can be tested and bounded, not where confidence must be invented.

  4. Prove the release

    Attach each requirement to its change set, tests, review, security controls, release record, rollback and production evidence. The build is complete when the operation can trust it.

A visible evidence chain

The build receipt is part of the product.

Delivery control room / Build 047

Synthetic demonstration

Requirement / REF-014

Mirror a refund into operations without triggering a second money movement.

Money-touching Human authority

Source
Operator interview + system event
Failure mode
Duplicate financial action
Acceptance
Ledger parity + zero side effect
  1. 01RequirementREF-014 · source retainedSpec
  2. 02Implementationchange set linkedIn review
  3. 03Verificationcontract + regression suitesPass
  4. 04Releaserollback preparedGated
  5. 05Production proofoperating evidence attachedVerified

Release authority · AI → H

Prepared by the system. Released by a person. Consequential actions stop at an explicit human gate. The approval and its evidence remain attached to the release.

Awaiting approval Release 047 is held

  • Requirement linked
  • Test evidence attached
  • Rollback ready
  • Human-gated

What the method protects

Fast does not mean loose.

Least privilege

People, services and models receive only the data and actions required for the task.

Human gates

Money movement, clinical decisions, material claims and releases stop at an accountable person.

Source retention

Extracted facts and model outputs remain connected to the records that support them.

Failure design

Retries, dead-letter paths, monitoring, incident ownership and rollback are specified before launch.

Questions buyers ask

How long does a pilot take?

Most production pilots take three to six weeks after the audit or specification is approved.

Do websites follow the same process?

Yes, in a lighter form: a concept preview, a fixed-price proposal, a staging site you review, an accessibility check on every page, and launch with redirects and training.

What do you need from us?

A named decision-maker and a short weekly review. For workflow projects, a few hours of your operations lead's time in the first week.

See it applied in the telehealth platform, read human-gated AI, or browse capabilities.

Bring the workflow that keeps breaking.

We'll tell you whether it's a fit for a one-week workflow audit, and what evidence we'd need to begin.

Email ian@kotohlabs.com