A junior designer at your firm runs a generative tool over a tenant fit-out, the tool produces a clean set of egress and life-safety sheets, and the project architect, slammed on three deadlines, applies the firm's stamp and pushes the package to the plan checker. Two months later the building official calls with a question about the occupant-load assumption on a corridor, and the architect of record whose seal is on the sheet cannot answer it, because nobody at the firm actually formed the judgment the seal certifies; the AI did, and the AI is not a licensed professional. That single sheet is now a liability the firm's professional-liability carrier will scrutinize, a board-complaint exposure for the named architect, and a credibility wound with that jurisdiction. The stamp is binary: it is either backed by a responsible licensed professional who exercised judgment, or it is fraud. This lesson is for the design-firm principal, the AOR, and the EOR who must set policy before that sheet ships, and it ends in the named artifact your firm will actually adopt: the firm's design-practice AI policy that says, in writing, what professional judgment can and cannot be delegated to a tool.

The Stamp Is Binary and Cannot Be Delegated to AI

The program has carried one principle since L1: the stamp is binary. A sealed drawing is a licensed professional's certification that they exercised independent professional judgment over the work and take legal responsibility for it. There is no partial seal, no "AI-assisted" seal, no seal that means "mostly mine." The responsible licensed professional either formed the judgment the seal certifies or they did not, and if they did not, the seal is a misrepresentation regardless of how the drawing was produced. At the firm-policy level this becomes the load-bearing premise: the firm cannot write a policy that delegates professional judgment to AI, because professional judgment is precisely the thing the seal certifies and the thing no statute permits a non-licensee to perform.

What AI can do is enormous and worth capturing: it can draft, route, compare drawing sets, produce a code-narrative first pass, generate layout and routing options, populate a schedule, flag conflicts. What it cannot do is be the responsible professional. The architect of record and the engineer of record hold a statutory responsibility under their state practice acts that is non-transferable: it attaches to a licensed human, it is the basis on which the public relies on the seal, and it is the thing a board can revoke a license over. So the policy must draw a bright line that mirrors the binary seal: AI produces inputs and drafts, the licensed professional forms the judgment and owns the output through a real act of review, not a rubber stamp. Any workflow where the seal lands on work whose professional judgment was never formed by the sealing professional is the failure mode the entire policy exists to prevent.

This is why the policy cannot be a productivity memo. A firm can say "use AI to draft RFIs faster" and stop there, because an RFI draft is not a sealed instrument. But the moment the output flows toward a sealed sheet, a stamped calculation, a code interpretation, or a life-safety determination, the policy is no longer about productivity; it is about the integrity of the seal and the firm's license to practice. The principal setting this policy is not optimizing hours; they are protecting the firm's authority to put a seal on anything at all, which is the firm's entire reason to exist as a licensed entity.

What the AOR and EOR Can and Cannot Delegate

The practical core of the policy is a delegation map: which tasks the AOR and EOR may hand to AI as a draft-and-input engine, and which they may never delegate because they constitute the professional judgment the seal certifies. The delegable category is wide and is where the firm captures its efficiency: drafting (narratives, transmittals, specification first passes), document comparison (issued-for-permit versus issued-for-construction overlays), option generation (massing, layout, MEP routing alternates), extraction and population (submittal registers from specs, schedules from models), and first-pass checks that surface candidates for a human to verify. In each, the AI's output is an input the professional reviews, corrects, and adopts; the professional's judgment is still what ships.

The non-delegable category is narrow but absolute, and it is built from the program's five verification gates: design intent (does this embody what the designer actually intends), code compliance (does this satisfy the building code as the licensed professional interprets it), contract authority (does the firm have the authority to bind this), dollars (is this number a commitment that is correct), and life-safety (egress, structural adequacy, fire and smoke, the determinations on which people's lives depend). Tier-3 code interpretation, reading the code and deciding what it requires for this specific condition, is the licensed professional's stamped act and never the AI's; the AI may summarize the code and draft a narrative, but the interpretation the seal certifies is the human's and cannot be delegated. The policy must list these gates by name and state that no AI output crosses a gate without a licensed professional forming the judgment behind it.

The hardest cases sit at the boundary, and the policy must address them explicitly. A generative tool that produces an egress layout is not merely drafting; it is proposing a life-safety determination, so the policy must require the AOR to independently form the egress judgment rather than adopt the tool's layout as the determination. An AI code-narrative first pass is delegable as a draft, but the citations and interpretation are non-delegable, so the policy must require citation-by-citation verification against the published code, the same discipline the program taught against fabricated IBC sections. The principle resolving every boundary case is the same: if the output is the thing the seal certifies, the judgment behind it cannot be delegated, no matter how good the draft looks.

The seal is the firm's certification that a licensed human exercised professional judgment. AI can draft the inputs; it cannot be the judgment. A policy that delegates judgment to AI is not a productivity policy, it is a plan to seal work no one is responsible for.

The Cross-Firm Question: Who Is Responsible on an AI-Generated Drawing

The most dangerous gap in 2026 practice is not within a firm; it is across firms. Modern projects are multi-party: the AOR is at the architecture firm, the EOR is at the structural or MEP engineering firm, a design-assist trade contractor brings AI-generated routing options into the federated model, and the owner's rep coordinates the whole. When an AI tool generates a drawing or routing that flows across these firm boundaries, the question that must have a single named answer is: who is the responsible licensed professional on this AI-generated drawing? If the answer is ambiguous, the seal is at risk, because a seal with no clearly responsible human behind it is the exact failure the design-assist meeting in the program's Insider Brief produced, where three Augmenta-generated routes sat on the table and nobody on the call could answer who stamps them or where the basis-of-design memo lives.

The policy must resolve this with a rule that does not depend on who ran the tool: the responsible professional is the licensed individual who adopts the output into their sealed instrument by exercising judgment over it, regardless of which firm's software generated it. If a trade contractor's AI generates an MEP route, that route is an input until the EOR reviews it, forms the judgment that it is correct and code-compliant, and incorporates it into the sealed design; until then, no one is responsible for it as a design and it cannot be relied on as one. The policy must require that for any AI-generated content crossing a firm boundary, the receiving firm's licensed professional either adopts it through real review or rejects it, and that the basis-of-design memo names the professional who owns the resulting determination. This is the responsible-charge doctrine applied across the seam between firms, where ambiguity is most expensive.

This cross-firm clarity also protects the firm contractually. The program's contract lessons established that authority to bind and responsibility for design are governed by the AIA documents and the practice acts, not by who operated a piece of software. The policy must align the AI workflow with those instruments so the responsible professional named in the contract is the same professional who exercises judgment over the AI-generated content, with no gap where a tool's output is treated as a design no licensee has adopted. When the AHJ, the owner, or eventually an arbitrator asks who is responsible for a given AI-generated sheet, the firm must point to a named licensed human who reviewed it and owns it, and the policy is what guarantees that name exists.

AXP Supervisors and the Competence-Not-Dependence Problem

There is a second, slower failure mode the policy must address, and it is generational. The Architectural Experience Program (AXP), administered by NCARB, is how intern architects accrue the supervised experience required for licensure, and the AXP supervisor certifies that the intern developed real competence across the experience areas. AI introduces a quiet hazard here: if the intern uses AI to produce code narratives, drawings, and analyses without ever forming the underlying judgment themselves, they can log the hours and check the experience areas while developing dependence on the tool instead of the competence the program exists to build. A licensed architect who never learned to interpret the code without AI, who cannot read a set and form an egress judgment unaided, is exactly the professional who cannot answer for the sealed sheet when the AHJ calls.

So the policy must give AXP supervisors an explicit obligation: ensure interns develop genuine professional competence and not AI dependence, which means requiring the intern to form and defend the judgment behind AI-assisted work, not merely accept and route the tool's output. The supervisor's certification of the intern's experience is itself a professional judgment, and a supervisor who certifies competence the intern does not have misrepresents the AXP record the same way a hollow seal misrepresents the design. For the experience areas where judgment is the point, code analysis, life-safety, design development, the policy should require that the intern's AI-assisted work is reviewed for whether the intern can defend it, not just whether the output is correct, because correctness from the tool proves nothing about the intern's competence.

This connects directly to the firm's long-term ability to seal anything at all. The responsible professionals of 2035 are the AXP interns of 2026, and a firm that lets its interns substitute AI for judgment is training a generation that cannot back a seal. The policy's AXP provisions are not a compliance footnote; they are the firm's investment in its own future capacity to practice, protecting not only today's seals but the firm's ability to produce professionals who can responsibly form the judgment the seal will require a decade from now.

The Evolving NCARB, AIA, and NSPE Guidance

The policy must cite and track the professional bodies' evolving guidance, because the firm's standard of care will be measured against it. NCARB, which administers licensure and the AXP, is surfacing AI-awareness content for intern architects and is the body whose position on AI in the path to licensure most directly shapes the AXP provisions above; the firm should track NCARB's guidance on what AI-assisted experience counts and how supervisors should evaluate it. The AIA, through its 2026 Practice Points series, including the Practice Point on AI tools in professional practice, is the most credible AOR-side guidance in the United States on AI ethics, code-review limits, and stamp liability; the firm's policy should align its delegation map and disclosure language with the current AIA Practice Point and update when AIA revises it.

On the engineering side, NSPE has begun publishing AI-ethics guidance for licensed professional engineers that is rigorous on the engineer's non-delegable responsibility and the limits of relying on tools for engineering judgment; the EOR provisions should align with NSPE's guidance on responsible charge and the engineer's duty to hold paramount public safety. The firm should treat all three bodies' guidance as the evolving floor, not the ceiling: it is directional rather than a bright-line rule set, so the policy must translate it into the firm's own operational rules and name an owner who tracks revisions, because guidance issued in 2026 will be revised, and a policy that cites a superseded position is itself a defect.

A note the policy should state plainly: this guidance, and the policy built on it, is informational and operational, not legal advice. The firm's counsel and its professional-liability carrier must review the policy, because the practice acts vary by state, the carrier's 2026 renewal policies are introducing AI-related exclusions, and the disclosure obligations on sealed work are jurisdiction-specific. The policy names the guidance it tracks, translates it into firm rules, and routes the result through counsel and the carrier, so that the firm's standard of care is documented against the professional bodies' positions and defensible if it is ever tested.

Disclosure, Documentation, and the Defensible Record

A policy that draws the line but leaves no record is half a policy, because when the AHJ or the carrier asks how a sealed sheet was produced, the firm needs a defensible record showing the responsible professional formed the judgment. The policy must require documentation that survives the question: which deliverables were AI-assisted, what the AI produced, what the licensed professional reviewed and changed, and the named professional who owns the determination. This is the AI-touched-deliverable register the program built at L1 and L2, now a firm-wide instrument tying every AI-assisted sealed deliverable to the responsible human and the review they performed.

Disclosure is the second half. The policy must specify the firm's disclosure language: what is added to a transmittal cover sheet when a deliverable was AI-assisted, and critically, what is and is not said on a sealed or stamped sheet, which is jurisdiction-specific and must be cleared with counsel because some boards have begun considering AI-disclosure-on-stamped-work rules. The disclosure is not an admission of weakness; it is the firm stating accurately how the work was produced while affirming that the responsible professional formed the judgment the seal certifies, the posture that protects the firm if the deliverable is ever questioned.

The record and the disclosure together convert the policy from a statement of intent into an operational reality. A principal can declare that AI will not seal work no one is responsible for, but only the register and the disclosure make that checkable: the register proves a named professional reviewed each AI-assisted sealed deliverable, and the disclosure communicates accurately to the AHJ, the owner, and the carrier how the work was produced. This is the same verify-before-stamp discipline the program has taught throughout, now instrumented at the firm level so the firm can prove, not just assert, that every seal is backed by a responsible licensed professional.

The Applied Problem: Produce the Firm's Design-Practice AI Policy

Here is the exercise. Produce your firm's design-practice AI policy: the named artifact that states in writing what professional judgment can and cannot be delegated to AI, who is the responsible licensed professional on AI-generated work, how AXP supervisors protect intern competence, and how the firm documents and discloses AI-assisted sealed work. This is the L4 deliverable for the design-practice principal: not a productivity memo, but the firm's governing instrument for keeping every seal backed by a responsible human.

Build it in the sections this lesson developed. First, the binary-seal premise and the delegation map: the wide delegable category (drafting, comparison, option generation, extraction, first-pass checks) and the narrow non-delegable category built from the five gates (design intent, code, contract authority, dollars, life-safety), with tier-3 code interpretation and life-safety determinations named explicitly as the AOR's and EOR's stamped acts. Second, the cross-firm rule: the responsible professional is the licensee who adopts AI-generated content into their sealed instrument through real review, regardless of which firm's tool produced it, with the basis-of-design memo naming that professional. Third, the AXP provision: the supervisor's obligation to ensure interns develop competence not dependence, with the requirement that interns form and defend the judgment behind AI-assisted work. Fourth, the guidance alignment: cite the evolving NCARB, AIA 2026 Practice Points, and NSPE positions, name an owner who tracks revisions, and route the policy through counsel and the carrier. Fifth, the record and disclosure: the AI-touched-deliverable register tied to the responsible professional, and the disclosure language for transmittals and sealed sheets, cleared with counsel.

The deliverable is the firm's design-practice AI policy, and the lasting product is a firm that can put a seal on AI-assisted work and answer for it, because every seal is backed by a named licensed professional who formed the judgment the seal certifies. This is the capstone of the firm-strategy level for the design practice: the principal who masters it captures the AI efficiency in the wide delegable category while protecting the firm's license to practice in the narrow non-delegable one, resolves the cross-firm responsibility question before the AHJ asks it, builds the next generation's competence instead of its dependence, and documents the whole so the firm can prove every seal is responsible. The stamp is binary, and this policy is how a firm keeps it backed by a human at scale.

Key Takeaways

• The stamp is binary at the firm-policy level: a seal is a licensed professional's certification of independent professional judgment, and a policy that delegates that judgment to AI is not a productivity policy but a plan to seal work no one is responsible for. The AOR's and EOR's statutory responsibility is non-transferable to a tool.
• The delegation map is the policy's core: a wide delegable category (drafting, document comparison, option generation, extraction, first-pass checks, where AI output is an input the professional reviews and adopts) and a narrow non-delegable category built from the five gates (design intent, code, contract authority, dollars, life-safety), with tier-3 code interpretation and life-safety determinations named as the stamped acts AI can draft but never own.
• The cross-firm question must have a single named answer: the responsible licensed professional is the licensee who adopts AI-generated content into their sealed instrument through real review, regardless of which firm's tool produced it, with the basis-of-design memo naming that professional, so a seal never lands on work no licensee has adopted.
• AXP supervisors carry an explicit obligation to ensure interns develop genuine competence and not AI dependence: interns must form and defend the judgment behind AI-assisted work, because the responsible professionals of the next decade are this decade's interns, and a hollow AXP certification misrepresents competence the same way a hollow seal misrepresents design.
• The policy must cite and track the evolving NCARB (licensure and AXP), AIA (2026 Practice Points on AI in practice), and NSPE (engineer's non-delegable responsibility) guidance as a directional floor, translate it into firm operational rules, name an owner who tracks revisions, and route the result through counsel and the professional-liability carrier, since guidance and state practice acts evolve and 2026 policies are adding AI exclusions.
• Documentation makes the policy checkable: the AI-touched-deliverable register ties every AI-assisted sealed deliverable to the named professional and the review they performed, and the disclosure language (transmittal cover sheets and the jurisdiction-specific question of what is said on a sealed sheet) communicates accurately how the work was produced while affirming the responsible professional formed the judgment.
• The policy is informational and operational, not legal advice: it names the guidance it tracks, translates it into firm rules, and is reviewed by counsel and the carrier, so the firm's standard of care is documented against the professional bodies' positions and defensible if it is ever tested by an AHJ, an owner, a board, or an arbitrator.