A paid discovery pilot. One package family. One operating unit. One reviewer in the loop. Four metrics measured on your real data. Decisions made together at the end with the receipts in front of us.
Not a procurement maze. Not a multi-vendor bake-off. A scoped engagement with a real reviewer, real fixtures, and a clear off-ramp at the end.
Eight to twelve weeks, scoped to one package family and one operating unit. A signed scope, not a multi-quarter committee.
Real-shaped fixtures so the parser, indexer, and mapper run against the structures your reviewers actually see. Redacted or synthetic — whatever your data team can release fastest.
One named human reviewer. The criteria they apply now — even if it's tribal knowledge written down for the first time during week one. The pilot encodes the existing rubric, not a theoretical one.
Four specific KPIs measured on your packages: time-to-review, missing-evidence catch rate, override rate, and receipt completeness. We agree on the targets before week one, not after week eight.
Expansion candidates are spelled out up front as generic capability lanes: worksite intelligence, governance dashboards, back-office workflow, customer operations, and live intake. The pilot's job is to earn the next conversation, not assume it.
Weeks 1–8 are the build. Weeks 9–12 are the shadow review on real-shaped packages, alongside the human workflow. The final meeting reads out the metrics and decides what comes next.
Package family selection, fixture approval, redaction, scope sign-off. The week-two checkpoint either kicks off the build or surfaces the blockers honestly.
Fail-closed missing-sheet detection, requirement extraction. If a workbook is missing required tabs, the pilot says so loudly — it does not silently approve a partial manifest.
YAM detection + VisionPAC pool working on the real fixtures. Local inference, no external clouds, no customer-data exfiltration.
Decision panel for the named reviewer. Every accept, every override, every escalation produces a sealed receipt with manifest, evidence, finding, and decision in one envelope.
Four weeks of real-shaped packages running alongside the existing human workflow. Final meeting reads out the four metrics with the reviewer present, then we decide expansion together.
The pilot doesn't import a vendor's accuracy claim from another customer's deployment. It measures these four metrics against the baseline your reviewers experience today.
| Metric | What it measures | What success looks like |
|---|---|---|
| Time-to-review | Wall-clock from package upload to human signature. | Faster than the current baseline. |
| Missing-evidence catch rate | Percentage of MOP rows where required photo absence is correctly flagged. | High recall on missing-evidence — the system does not let a gap slip past the reviewer. |
| Override rate | Percentage of machine recommendations the human overrides. | Low for routine cases; preserved for the ambiguous ones — that's the point of the human in the loop. |
| Receipt completeness | Percentage of decisions with a fully-formed sealed receipt — manifest, evidence, finding, decision. | Approaching 100%. Receipts are the artifact, not a side effect. |
We don't claim 95% accuracy or any other vendor's borrowed number. We measure on YOUR packages, with YOUR reviewer, against YOUR baseline. If the pilot doesn't move those four numbers in the right direction, the honest recommendation at the end is "stop here."
At the end of the pilot you have a running ATLAS workflow instance scoped to the agreed package family, a complete audit trail, and a recommendation packet that makes the next decision easy — including the option to stop.
Scoped to the agreed package family and operating unit. Running locally on the agreed footprint, integrated with the workbook + photo intake your reviewer already uses.
Manifest, evidence, finding, decision — bound into one provenance-stamped envelope per package. Replayable. Exportable. Yours, at rest, on your footprint.
Every machine recommendation the human accepted, modified, or overrode — with reason, timestamp, and the evidence in front of them when they decided.
The four metrics measured against the agreed baseline. Read out in a meeting with the reviewer present, not delivered as a one-way slide deck.
Expansion options scoped honestly — or an honest "stop here" if the metrics don't earn it. The pilot is the receipt either way.
These are the shapes pricing takes at each stage of the relationship. Actual fees are scoped on signature against your package volume, footprint, and BU count.
Eight to twelve weeks. One package family. One operating unit. Fixed fee, scoped on signature. The pilot is paid because the work is real, the deliverables are real, and free pilots get deprioritized inside both organizations.
Annual platform fee plus an implementation fee for one country or one BU. Engaged after the pilot proves out — the pilot's KPI readout is the precondition, not the marketing.
Base platform fee plus a per-package or per-site review tier. Engaged once volume justifies variable pricing — variable fees should follow volume, not promise it.
Add-on modules for governance dashboards, agent assist, live intake, customer operations, and back-office workflow. Each module is its own scoped engagement with its own success criteria.
We don't price the substrate first. We price the closeout workflow outcome first, then expand the account by attaching governance dashboards, workflow support, live intake, and customer-operations capabilities. The substrate is real and it carries the receipts — but it's not what you're buying on the first invoice.
A pilot earns trust by being honest about its boundaries. Anything outside this list is also not promised — but these are the six most often misread, so we name them on the page.
Closeout-review automation is becoming a buyer category. Carriers, EPCs, and infrastructure operators are starting to ask the same question — "why is the most expensive part of every project the part where humans line up photos against a workbook?" — and the answers they accept now will set the patterns the rest of the industry adopts.
The carriers that pilot now do three things at once. They move their own time-to-review on a workflow that bleeds margin. They earn a seat at the table when category standards crystallize — what a sealed receipt looks like, what override rate is acceptable, what fail-closed parsing means in practice. And they do it with a reviewer-in-the-loop posture that keeps regulator-facing accountability where it belongs: with the licensed human who signs the package.
The carriers that wait inherit whatever the early movers normalized. That's a worse position whether you're a buyer, an investor, or a senior operator who's seen this pattern in adjacent categories. Pilot now, set the pattern. Pilot later, accept it.
One reply. One named reviewer on your side. One package family. We come back with a scoped pilot agreement inside a week.
Machine recommends. Human decides. Plan-acceptance is not implementation-authorization. The pilot's job is to make the next decision easy — including, honestly, the decision to stop.