The two workflows are chosen. What you stand up next is not a chatbot. It is inventory, leases, evaluation, rollback, and a named gate—so the agents can run.
The first two workflows in an engineering firm can be named on one page. Offer factory or aftermarket factory. Input, output, gate, system of record.
That page is not a project.
A project is what makes those two chains runnable as a twin: the same live files, a five-layer stack, humans above the gates, and a band around every layer that an underwriter, a director, and a court can read from the same record.
Without that band, the agents stay locked, leak into shadow tools, or run once and get decommissioned after the first wrong part number. Gartner’s 2027 number already forecasts that failure mode. Governance treated as a workshop after go-live is how you get it.
This note is the project design. Not a transformation office. Not a platform program. Twelve weeks, two adjoining workflows, the guardrail product shipped with the first draft — not after it.
What this project is for
It exists so the legal person can say yes.
An engineering company cannot let an agent assemble an offer, a manufacturing-relevant BOM, or a spare-parts list unless four conditions hold at once:
- Every running agent has a named human owner, a purpose, and a lease with an expiration date.
- Every material act hits a gate a human can refuse.
- Every act leaves a reconstructable record: who authorized it, on whose behalf, with what minimum privilege, which similar job or rule fired, and what was written where.
- Every act can be rolled back without a war room.
Those four are not an ethics appendix. They are the condition under which the CEO is allowed to run a parallel. They are also the condition under which the twin is insurable, auditable, and able to learn.
The project ships that condition as software. Then it runs the two paths. Then it reads four numbers. Then it deprecates what loses.
Mandate, not a steering committee
The twin sits inside the firewall and reports to the CEO. That is not poetry. It is the only reporting line that survives the immune system of a machine builder — calculation, construction, IT, the works council, the digitalization board.
Write the mandate on one page before anyone opens a model:
- Object. Pair A (complete RFQ pack → released offer, then accepted offer → released order-engineering pack) or Pair B (identified machine + parts request → confirmed parts list, then fault report → first-action service pack). Not both. Not a compromise pair for two vice presidents.
- Authority. The two named owners — Bid Management and Order Engineering, or Spare Parts and Service Operations — change the path. They do not convene a standing committee to ask permission.
- Constraint. No write-back to the customer, to purchasing, or to the laser cutter without the human gate, for the whole parallel.
- Success. After four weeks of live dual-run, the twin is better on cycle time, first-pass completeness, unit cost, and risk events — on the same inputs. Risk is a first-class number. Without it, “deprecate the loser” is a slogan.
- Stop rule. If the queue is empty because the workflow happens twice a month, the project chose prestige. Stop. Rename the object. Do not expand scope to hide the miss.
If that page needs six signatures, you do not have a mandate. You have a pilot.
Twelve weeks, five work packages
Do not start with the model. Start with the band. The stack is unusable until the band exists; the band is untestable until the stack has something to wrap. Build them as one product.
Weeks 1–2: Inventory, leases, and the data fork
Inventory. List every agent, tool, copilot, and script that already touches inquiries, order folders, parts identification, or service cases. Include the ones a Konstrukteur stood up last Tuesday. No owner, no agent. The ones without a name go dark or get an owner this week.
Leases. Each agent that will run in the twin receives a passport, not a career: purpose in one sentence; allowed tools and named fields; forbidden acts (send to customer; release to production; accept an ATEX, safety, or liquidated-damages deviation; invent a module that is not in the rule table); time-bound; minimum privilege. An agent may persist as an identity. Its right to act expires.
Data fork. Copy what the current path already uses — inquiry files, similar-job library, product rules, cost and lead-time tables, serial-number master, parts catalog and supersession chain, service history. Do not cleanse thirty years of Teamcenter. Instrument it. Version it. The ERP and PDM remain consumers, not owners of the work.
Weeks 3–6: Stack and band, layer by layer
Each layer of Sense → Interpret → Decide → Orchestrate → Learn carries its own evaluation, log, gate, and way back.
Sense classifies every ingested object (in-scope or not; personal data or not; complete file or not). Incomplete files are not a workflow; they are a human queue.
Interpret logs the objects it retrieved. An evaluation suite scores extraction against held-out real RFQs and parts requests before live files. Hallucinated part numbers and silent drops of a constraint are failing tests, not “model behavior.”
Decide emits a proposal plus an exception log. No exception log, no proposal. If the lease does not cover the deviation class, the agent cannot close. It can only escalate.
Orchestrate writes only into the review queue. Circuit breakers sit on volume, on class (safety, ATEX, pressure equipment, food-contact), and on confidence. Rollback is the default.
Learn is owned by the legal person. Overrides and “why we refused” do not leave the firewall to train a vendor’s general model. Continuous learning under your control is a product-liability fact after 9 December 2026.
Humans stay above the loop. They approve or refuse at the gate. They do not search, copy, and align.
Weeks 7–8: Evaluation suites and the review queue
A suite that only scores fluency is decoration. A suite that scores commitment error is the product: a missed constraint on page four of the RFQ; a module not in the rule table; a superseded part recommended without saying so; a safety note omitted from a first-action pack.
The review queue shows the proposal, the exception log, the similar objects used, the rule hits, the lease, and the one-click refuse. If overrides sit above twenty percent after two weeks, you do not yet have a workflow. You have a craft wearing a passport.
Weeks 9–12: Parallel run, board pack, deprecation rule
Same live inputs. Two paths. The legacy path remains official. Once a week, the owners and the CEO read four operating numbers and three governance numbers: cycle time, quality, cost per instance, risk events; plus override rate and time-to-contain; lease inventory; and a reconstruction test — can the record explain last week’s worst miss in an afternoon?
If you cannot explain the agent to an underwriter, a director, and a court from the same record, the twin is not ready for write-back. Four weeks is enough to know whether you chose a road. It is not enough to give the junior tenure. Keep the lease short.
Classify the agents. Do not govern them as one species.
Class R — read and retrieve. Class D — draft, write only to the queue. Class E — escalate; the only legal act is to stop and name the gate. Class W — write-back; shortest lease, dual control, not in week one.
A read-only retriever and a write-back agent do not share a policy. Uniform governance is how you get both shadow copilots and an unsupervised executive.
What the legal person still holds
The interface can move. The accountability boundary does not: residual claim, the data loop, purpose as protocol, the veto that has no playbook. Agents execute inside a lease. They do not hold the evidence, the loop, or the veto.
What this project is not
Not a Copilot rollout. Not a PLM migration. Not shop-floor execution. Not generative design. Not an ethics committee on Thursdays. Not five workflows. Two adjoining chains, one stack, one band.
The one-page charter at the end of the draft is the start condition: if you can’t fill it out without a workshop, the project isn’t ready.
Evolut_ing, as the series uses the word, is taking ownership of the picture. Here, the picture is the project: two workflows, one band, humans above the loop. Initiate an evolution before the revolution occurs.