The factory is a delivery system, not a sovereign decision-maker
Organizations increasingly describe an AI Factory or Agent Factory as the place where use cases are shaped, prototypes are built, controls are applied and capabilities are prepared for production. That can be useful. The governance problem begins when the same mechanism that builds the capability is also treated as the authority that decides whether the organization should continue funding it, accept its residual risk or scale its use.
L1 should keep those decisions explicit. Strategy and portfolio governance decide what the organization is willing to commit to. The factory turns that commitment into an engineered capability and produces the evidence needed for the next decision. Those are connected responsibilities, not the same responsibility.
Four decisions are often collapsed into one word: “promotion”
Promotion sounds simple, but it hides several different authorities:
- Portfolio progression: should the organization keep investing in this initiative?
- Capability promotion: has this implementation met the defined qualification envelope for reuse or a higher environment?
- Deployment authorization: may this exact capability run with these permissions, data, dependencies and controls?
- Assurance conclusion: does the evidence support the stated level of confidence, and what gaps remain?
These decisions can share evidence, but they should not share authority by accident. A strong benchmark result can support capability qualification without proving business value. A security approval can allow deployment without proving the initiative deserves more funding. A successful pilot can justify further experimentation without proving organization-wide scale is appropriate.
A practical gate model
I use six portfolio states for the L1 interface. They are an architecture pattern, not a prescribed NIST, ISO or government process.
Gate 0 — Admit the experiment
The organization decides that the question is worth testing. Minimum evidence should include the business problem, intended beneficiary, accountable owner, experiment scope, initial risk context, cost envelope and an explicit statement of what the experiment is expected to learn.
The outcome is not “approved AI”. It is permission to spend bounded resources to reduce uncertainty.
Gate 1 — Commit to development
After early evidence, the initiative either becomes a funded portfolio commitment or stops being one. The decision packet should show the value hypothesis, material risks and obligations, ownership, expected operating model, dependencies, total-cost assumptions and evidence required before any later progression.
This is where a demo becomes an organizational commitment. It still does not become a production capability.
Gate 2 — Accept a qualification target
The factory now needs a defined target instead of a vague instruction to “make it production ready”. Required capability, quality, security, privacy, reliability, latency, cost, human-oversight and evidence expectations should be explicit enough that technical teams can test them and portfolio owners can understand what they are funding.
This gate hands work toward L2 and downstream technical layers. It does not predetermine their result.
Gate 3 — Decide whether qualified capability deserves production commitment
The factory returns a qualification packet: what was tested, what passed, what failed, which assumptions remain, what residual risks exist, what the expected operating cost is and what controls depend on human or external systems. Portfolio governance then decides whether the business case still holds.
A technically qualified capability can still be rejected because the value case weakened, cost grew, obligations changed or a better alternative appeared.
Gate 4 — Continue, scale or constrain
Production is not the end of the portfolio lifecycle. Actual value, operating cost, incident history, control performance, user adoption, model or supplier changes and new obligations should feed the next decision. Scale is another commitment, not the default reward for surviving deployment.
Gate 5 — Pause, redesign or retire
A governed factory needs a normal exit path. An initiative may stop because value did not materialize, risk exceeds appetite, a dependency became unacceptable, evidence quality is too weak, economics deteriorated or the capability is no longer strategically relevant.
Stopping should not require organizational embarrassment. A portfolio that cannot stop weak initiatives is not governed; it is accumulating sunk cost.
The evidence model behind every gate
The Tuesday L1 article defined four conditions for a portfolio commitment: value, risk, ownership and evidence. The same four conditions should be refreshed at every gate rather than copied forward unchanged.
- Value: what outcome has been observed, against which baseline, and how much uncertainty remains?
- Risk: what changed in exposure, assumptions, obligations, suppliers or operating context?
- Ownership: who has authority for the next commitment, and who owns the consequences if it proceeds?
- Evidence: which observations support the decision, which evidence is independent, and which gaps remain explicit?
A gate is useful only when it can produce more than “approved” or “rejected”. It should preserve the decision, conditions, owner, evidence basis, expiry or review trigger and any constraints that the next layer must enforce.
What current guidance supports
NIST AI RMF 1.0 treats governance as a cross-cutting function across the AI lifecycle, calls for clearly defined roles and responsibilities, and places responsibility for AI development and deployment risk decisions with executive leadership. It also includes ongoing review and safe decommissioning as governance outcomes. That supports explicit decision rights and lifecycle gates, but NIST does not prescribe the six-gate factory model used here.
The NIST AI RMF Playbook's Manage function explicitly asks whether an AI system achieves its intended purpose and stated objectives and whether development or deployment should proceed. That is strong support for repeated proceed / change / stop decisions rather than treating delivery as a one-way pipeline.
The UK DSIT AI Risk Management Toolkit, published 8 September 2026, emphasizes risk ownership, ongoing lifecycle management, monitoring, scaling responsibility and continuous reassessment. The toolkit is government guidance, not a mandatory universal portfolio framework.
Gartner's September 2026 AI budgeting webinar warns that visible AI initiatives can be overfunded while foundational capabilities required to scale and sustain them remain underfunded. That is analyst guidance, but it reinforces a practical portfolio point: factory throughput is not a substitute for investment discipline.
The minimum promotion packet
Before an initiative advances through a material L1 gate, the decision packet should contain enough information for an accountable owner to make a real decision:
- current lifecycle state and requested next state;
- business objective, value hypothesis and observed value evidence;
- accountable business owner and decision authority;
- material risks, obligations and unresolved exceptions;
- qualification evidence and known test boundaries;
- operating-cost and dependency assumptions;
- security, privacy and resilience constraints that must remain true;
- requested scope, users, data, permissions and environment;
- conditions for pause, rollback, reassessment or retirement;
- next review trigger or review date.
One control flow, from idea to scale
A useful end-to-end pattern is:
The factory owns much of the work in the middle. It should not own every decision around it.
Claim status
- Established: authoritative guidance supports lifecycle risk management, explicit roles, accountability, monitoring, reassessment and proceed/stop decisions.
- Supporting analyst evidence: current Gartner research highlights structural AI investment and scaling risks.
- Architecture synthesis: the six L1 gates, minimum promotion packet and separation between portfolio progression, capability promotion and deployment authorization are the implementation model proposed here.
- Not claimed: that “AI Factory” is a standardized governance function, or that NIST, DSIT, ISO or Gartner prescribe this exact gate sequence.
Next in the series
Week 3 moves into L2: The Governed Continual Harness: How Agent Systems Are Allowed to Change. The question shifts from “should the organization continue this initiative?” to “under what evidence and authority may the capability itself change?”