Proof

Inspect the work,not the demo

Customer names stay private — these systems run internal operations. What cannot stay private is the shape of the work: status, elapsed time, build effort, where the system stopped, and what was recorded.

Status
Production after 1 month
Build effort
2 person-weeks
Evidence
Definition + run records
In productionSystems integrator A (Japan)

Engineer skill management

Live after 1 month / 2 person-weeks

The AI interviews engineers about their work history. It pushes back when answers are thin, splits a manager's free-form comments into individual action items, and reads attrition signals out of past sessions. Judgment work, not a macro.

  • Inputs were uneven interviews and free-form comments, not clean forms
  • Publishing always waits for manager review
  • Every run left a record the agent cannot edit
In proposalLocal AI provider B (Japan)

Control plane for private AI

Policy and audit stay outside the model

A commercial private-AI deployment cannot let the model hold credentials or rewrite its own history. Perstack keeps policy checks, external calls, and the audit record outside the agent while the work still moves.

  • Deterministic policy checks before commit
  • Credentials held in the gateway, not the model
  • Tamper-evident record of every action

No screenshots, no demo video. On a headless system the evidence is the records themselves — and they can be shown as they are. Three of them below: the workflow's review gate as defined, one run as the system recorded it, and the surfaces you can demand from any vendor.

From the workflow's definition — the review gate

  • ProcessDraft → generated → owner review → manager review

    Every sheet walks the same path. The agent cannot reorder or skip it.

  • ProcessOnly manager review leads to publishthe only path

    A version that has not passed review cannot publish — the write is refused before commit.

  • ProcessFlagged work goes back with comments

    It is supplemented and reviewed again; nothing publishes from the flagged state.

A run, as the system recorded it

Workflow
Engineer skill review
Status
Production
Built in
2 person-weeks
  • 09:14Follow-up askedheld before write

    A thin answer hit the quality rule. The agent asked for project scope before touching the sheet.

  • 09:22Publish attempt rejectedblocked before commit

    Publishing was requested before review. The write never landed.

  • 09:31Approval openedwaiting on manager

    The pending change, its reason, and the affected record went to the manager.

  • 09:36Commit recordedsealed outside the agent

    The approved write was stored with who, what, why, and when.

For the skeptical reader

How to tell this is runtime control, not model self-review

Ask any vendor for three surfaces. Perstack exposes all three.

The definition

the workflow, as a written document

What data exists, what operations exist, what can fail and how — readable before anything runs.

The run

the run history, step by step

Status, messages, tool calls, retries, required approvals, blocked reasons — per run, as it happened.

The verdict

success, business stop, or system failure

A business-rule violation is returned before commit — not apologized for by the model afterwards.

If a demo cannot show a rejected write before commit, it has not reached commit control yet.

Getting started

Bring one workflowthat already hurts

Rough notes are enough. No requirements document, no process map, no company-wide AI plan.

What to send

  • The workflow

    What starts it, what finishes it, which handoffs make it slow.

  • One hard rule

    The thing the AI must never do.

  • The system edge

    The database, SaaS, or internal API it reads or writes.

What comes back

A short memo, before any commitment:

  • Fit or not — if a lighter tool is enough, we say so
  • A first definition: trigger, finish line, rules, approvals
  • A sample of the run record you would inspect

Before you buy

Fit

Pick one operation

We map it into a first definition together. If it does not fit, that is the answer.

Month 1

Production

Core workflow live

Schema, integration, core operations — about 2 person-weeks of build.

Month 2

Steady state

Exceptions and routine

External edges, exception paths, monitoring, and the runbook.

Give AI one workflow.It can't break your rules.

Rough notes are enough: what starts it, one rule that cannot break, the system it touches. A short memo comes back — fit or not, a first definition, a sample run record.