Karvark Tech LabsBook a free Leak ScanFree Scan

How we build

Documents before code, and constraints instead of good intentions.

Most of the difficulty in an AI system is not getting a model to produce something impressive. It is making the result something you can inspect, correct and be accountable for. These are the five practices that follow from taking that seriously.

  • RECEIPTS1,000+ automated tests on every changeA real-model check every night
  • START HERE

    8-Point Leak Scan: we check all eight places your enquiries can leak, free.

  • THE RECORDverified every night

    Every step is kept in a tamper-evident record.

One enquiry, step by step

What happens inside the system, in order. The model drafts; rules in code decide what is sent.

One enquiry, end to end

An example shape, not a client engagement.

  1. TriggerAn enquiry arrives by form, email or WhatsApp
  2. ToolThe system reads it and the CRM record
  3. AI stepIt drafts a reply that says it is from an AI
  4. DecisionRules in code decide: send, or route to a person
  5. ActionThe reply goes out; the owner gets the deadline
  6. System updatedCRM, calendar and the record all written

The record

A complete record of what arrived, what was said, and who did what and when, which you can export and hand to anyone.

The five practices

  1. 01

    Documents before code

    The architecture, the data model and the test strategy are written and version-controlled before implementation starts. The specification is the authority: when the code and the document disagree, that is a defect to be resolved rather than a detail to be absorbed. It also means you can read what is going to be built, and say no, while saying no is still cheap.

  2. 02

    The model narrates, code decides

    A language model is used to draft, explain and propose. It is not the thing that decides whether an action is permitted. Authorization, validation and business rules live in ordinary code and in database constraints, where they can be read, tested and proven. A model that is asked to be careful is a hope; a constraint that rejects the write is a guarantee.

  3. 03

    Human takeover on every agent

    Anything that acts on your behalf ships with a route for a person to see what it is doing and take the controls. That path is built in from the start rather than added after an incident, and it is part of what gets tested before release.

    What a takeover actually looks like, step by step.
    1. TriggerA request arrives and the process starts, exactly as it does on any other day.
    2. AI stepThe model reads the request, works out what is being asked, and proposes what to do next.
    3. DecisionThe proposed action is checked against the rules in code. This one falls below the confidence required to act alone.
    4. HandoverThe agent stops rather than guessing, and escalates to a person with everything it has: the request, what it proposed, and why it stopped.
    5. Person decidesA human makes the call, corrects it if the proposal was wrong, and hands control back.
    6. Resume, and recordThe process continues from where it stopped, and the audit entry is written: what changed, when, on whose authority, and what the system believed at the time.
  4. 04

    Eval-gated releases

    Model-driven behavior is held to a suite of evaluations that has to pass before a release goes out, in the same way a test suite gates ordinary code. Changing a prompt is a change to the system, so it is reviewed and measured like one rather than edited in production.

  5. 05

    An audit trail on every mutation

    Every change the system makes to your data is recorded: what changed, when, on whose authority, and what the system believed at the time. This is what makes an automated process reviewable after the fact, which is the difference between a system an institution can adopt and one it cannot.

Why this is the whole argument

Gartner has projected that more than 40% of agentic AI projects will be canceled by the end of 2027, attributing this largely to weak governance and unclear value rather than to the models themselves. That is a projection rather than a measurement, and it is a forecast we are quoting rather than a finding of our own.

Gartner press release, 25 June 2025

We think it points at something real, and it is the reason this page is not an afterthought. The practices above are what governance looks like when it is built into a system instead of written into a policy afterwards: a specification you can read, a decision path you can audit, a person who can take over, and a record of everything that changed.

The 8-Point Leak Scan is where this starts. See what it covers, or read how an enquiry is handled.