Method

How we build it, and how you can check it.

Software does the routine checking. Your people do the deciding. This page shows the work behind a Proof: the four stages, what a record looks like, and what your security and compliance team will want to read. Price and terms are on the Consulting page.

The 30-Day Proof

The four stages, in detail.

  1. Week 1

    Decide what gets checked.

    We score a few of your workflows on value, data and risk with you, and pick one. Then we write down the bar: what the checking has to get right, and at what cost, for it to be worth keeping.

    • Day 5: bar agreed in writing; your team signs off the past records we test against
    • Day 7: if we don't believe we can hit it, we say so and stop. You owe nothing more.
  2. Week 2

    We draft the workflow and the rules.

    We turn your procedures, forms and past records into a written workflow: the steps, who does each one, the rules each step is checked against, and who receives which exception. Where rules are enough, we use rules. Where an AI model is needed, we test up to three candidates on your records, against one leading hosted model where your data policy allows. Your team reviews a draft instead of writing one.

  3. Weeks 3–4

    Run it beside your current process.

    It runs on your systems, reading them but never changing them, on real work, with up to 25 of your people using it. Routine cases are checked against your rules, and exceptions go to the person you named, with the reason. Every check and decision leaves a record: what was checked, against which rule and version, and who decided.

  4. Day 30 onward

    Switch when your team trusts it.

    You get the records, the comparison against the bar and the evidence pack. If the bar was hit, you pay the balance. Switching over, writing back into your systems and adding workflows is a separate, scoped step on your timeline. Or you stop there.

What “live” means. Your people are using it on real work, on your systems, beside the old process. We haven't switched anything off.

The day the process doesn't fit.

Every team has a process it follows and days it has to bend it. Routine cases run as written. A case that doesn't fit goes to a named person, who decides and says why, and the reason stays with the record. When the same exception keeps coming back, you change the rule. The version moves on, and earlier decisions keep the version they were made under.

Try it in the demo 
What a record looks like

Every check leaves one of these.

What was checked, against which rule, who decided and when. This is an example from an invoice-against-contract check.

decision_idex-0001
workflowinvoice-vs-contract check
rule_applied"PO above threshold requires second approver" (v3)
outcomeexception, routed to Finance Ops
reviewed_byfinance-ops reviewer (example)
timestamp2026-03-14 09:12:41 UTC (example)
previous_recordex-0000
Illustrative Invented values, real shape. The questions an auditor asks, in the order they ask them.

“Just automate this. We don’t want AI.” We’ll take that brief.

If rules do the job, we use rules and tell you so. If a simple rule does the job, we'll use a rule. A model goes only where one is needed, and it's tested on your own records before anyone relies on it. If the leading hosted model beats the alternatives on your workflow, the report says so.

For your security and compliance team

Use hosted AI where it wins. Keep control where it matters.

We're not anti-API. We make sure you can leave one without a rewrite, and that the evidence stands up when someone else reads it.

How "better" is measured.

A test set is a set of past records your team signed off, so "better" is measured on your work, not ours. The leading hosted model is tested on your data too, and where it wins we say so. The benchmark format below is illustrative.

Rented
  • Data leaves your systems unless you contract otherwise.
  • Pricing and model versions are the vendor's call.
  • Your audit trail depends on their logs.
Owned
  • Runs on your systems, and your data stays with you.
  • You choose when models change, with test results as the evidence.
  • Every AI dependency has an exit plan.
  • The models, test sets, instructions and documents are yours when we leave.
  • The evidence pack is written so your compliance team can hand it to a regulator.

Illustrative Format only; these are not our results.

MetricLeading hosted modelYour own model, on your systems
Task accuracy
Hosted90%
Yours84%
Cost per 1,000 tasks
Hosted$16
Yours$7
Response time, slowest typical (p95)
Hosted2.8 s
Yours1.8 s
Data leaving your country
Hosted100%
Yours0%

If the leading hosted model wins on your workflow, we tell you. That is the job.

Tell me when Benchmark v1 publishes
Specified, not deployed

Reference design: what runs where.

Six parts, installed inside your environment, wired to one workflow, and left with you. This is the design we build to; the first build happens in your Proof.

Reference design: what runs where Six parts inside a boundary labelled Your environment: routing, models, retrieval, guardrails, tests and tracing, evidence and audit log. One arrow leaves the boundary to a leading hosted AI model, only where your data policy allows. Your environment RoutingSends each case to arule or to a model ModelsThe AI models youchoose, tested first RetrievalFinds the rightdocument or record GuardrailsLimits what the systemmay read and do Tests and tracingScored on your records;each step logged Evidence and logThe record of everycheck and decision Leading hostedAI model, only whereyour data policyallows. Reference design: what runs where Six parts inside a boundary labelled Your environment: routing, models, retrieval, guardrails, tests and tracing, evidence and audit log. One arrow leaves the boundary to a leading hosted AI model, only where your data policy allows. Your environment RoutingSends each case to a rule or a model ModelsThe AI models you choose, tested first RetrievalFinds the right document or record GuardrailsLimits what the system may read and do Tests and tracingScored on your records; each step logged Evidence and logThe record of every check and decision Leading hosted AI model, onlywhere your data policy allows.
In one line: the only path out of your environment is to a hosted AI model, and only where your data policy allows.

The exit plan.

Every AI dependency has a written way out. The three-year cost and exit map lists what each one costs to run and how to switch away from it. You own the test set, so you can re-run it on any replacement.

Score your own AI exit risk 

Where the evidence pack applies.

The pack shows how the workflow meets the rules in one country or region: India, Australia, the GCC, or the EU and UK. The US is planned.

  • India
  • Australia
  • GCC
  • EU and UK
  • US: planned

A map of the rules, not legal advice. Have your own counsel confirm what applies to you.

Pick the workflow. We'll tell you if it's a Proof.

A 30-minute call and six questions. You get green, amber or red.

Talk to us See it work