Sandbox first. A method.

Four stages, from proof to production.

1. 01

   Prove in sandbox

   See the governed run on synthetic or sanitized data — before any credential or file of yours is touched.

2. 02

   Map the real workflow

   We watch the real workflow and time the manual baseline — the map is built from the work, not a questionnaire.

3. 03

   Shadow the live process

   The system runs alongside your team on real work with zero external effect, measured against your own baseline.

4. 04

   Pilot with human approval

   Live work with a person approving every output — the pilot ends with an assist log and a clean boundary audit.

The nine-step protocol below is how each stage is actually run.

1. 01 Discovery

   Fit decision — go or no-go

   A listening call, not a pitch: where the time actually goes, what breaks under deadline, and what you'd never let software do alone. If we don't see a fit, we say so and stop here.

2. 02 Workflow Map

   Timed manual baseline

   We observe the real workflow and time the manual baseline — how inputs arrive, who touches them, and where the minutes live. The map is built from watching the work, not from a questionnaire.

3. 03 Data Inventory

   Data catalog + retention terms

   Every document and system the workflow touches gets cataloged, with a representative sample de-identified for the work ahead. Retention and deletion terms go in writing before any real file moves.

4. 04 Risk Classification

   Processing-boundary map

   Each kind of data is classified, and that class decides where it may be processed — what stays inside your boundary and what never leaves. The hard lines are drawn here, before anything is built.

5. 05 Pilot SOW

   Scoped statement of work

   One bounded slice of the workflow, in writing: what the system will do, what it will never do, and success metrics measured against your own baseline — not our claims. Nothing expands past this scope without a new agreement.

6. 06 Sandbox

   System built on de-identified data

   The system is built and exercised on the de-identified sample — never your live operation. It has to work here before it earns anything more.

7. 07 Shadow Mode

   Output-vs-baseline comparison

   The system runs alongside your team on real work with zero external effect: it drafts and flags in parallel, and we compare its output to what your people actually did. Misses cost nothing and teach us everything.

   Most vendors demo on curated examples and go live on hope. Shadow mode means you see real performance on your real work before the system is trusted with any of it.

8. 08 Human-Reviewed Pilot

   Assist log + boundary audit

   Live work, with a person approving every output — the system drafts, your team decides, and nothing external ever sends itself. Every assist is logged, so the pilot ends with evidence instead of anecdotes.

   Autonomy is earned, not assumed. The pilot has to end with a clean boundary audit — zero autonomous sends, zero machine determinations — or it doesn't pass.

9. 09 ROI Review

   Measured result · shared go/no-go

   Measured results against the baseline from the workflow map, plus the full assist log, reviewed together. We expand deliberately or we stop — go/no-go is a shared decision, not a renewal default.