What staying in responsible charge of AI-assisted work looks like.

The situation

A structural PE uses an AI-assisted tool on steel connection calculations feeding a stamped set. The outputs look reasonable — which is the risk, because a bad shared assumption can look reasonable all the way through.

Staying in charge

The PE sets the inputs, assumptions, and acceptance criteria before running the tool, so the output is measured against the engineer’s criteria rather than the machine’s answer. Verification is risk-based, not uniform — governing and unusual connections get an independent check (a hand calculation, a different established package, a limiting case) capable of failing on its own terms; routine repeats within a checked range get proportionate scrutiny. The check is strongest when it doesn’t inherit the same assumption that drove the AI run.

What the record documents

Which AI touched the connections and its role, the criteria set beforehand, the independent verification performed and by what method, and the named PE — with the licence held at signing — who adopted the work.

The situation

A firm principal opens the professional-liability renewal and finds new questions: does the firm use AI on professional work, and how is its output reviewed? “We’re careful” is not something a carrier can evaluate.

Staying in charge

The firm doesn’t reconstruct a story at renewal — it has recorded responsible-charge evidence as the work was done, and can show, deliverable by deliverable, what a licensed professional directed, verified, and adopted before sealing AI-assisted work.

What the record documents

A firm-prepared evidence summary the broker can read — which deliverables the firm declared AI touched, and what each record contains. Not an underwriting score, and not a story — the firm’s own contemporaneous records.

The honest two-way point

This is evidence the firm produced about itself — useful, but self-declared, not an independent audit. And it cuts both ways: a complete record on a deliverable is stronger than none, but an incomplete record on a deliverable the firm declared AI touched is equally visible. That candour is the point.

The situation

A firm turned a validated design spreadsheet into a small web app with AI’s help, and has reused it across projects since. A different PE now picks it up for a new project — different loads, a newer code edition, different geometry. The AI touched the code, not the engineering, so the live danger is the undocumented translation step: a formula that may have shifted between spreadsheet and app, with no version control and nobody able to read what’s inside.

Staying in charge

The new PE does not inherit the original validation. This PE documents why this use of the tool is appropriate for this project — checking the embedded assumptions, standards, equations, and conditions against the new requirements, and verifying outputs against an independent method on the project’s own cases. Prior validation is useful evidence, not a substitute for the current engineer’s judgment. (In practice, a validated-tool register with defined re-validation triggers — code-edition change, use outside the validated envelope, version change — is how firms scale this without redoing everything every time.)

What the record documents

A project-specific record tied to this PE — why this use was appropriate here, what was checked, and the independent verification on this project’s cases, distinct from the firm’s original baseline validation.

The situation

A licensed architect uses an AI assistant to locate and summarize a code provision for a permitting narrative. The answer is fluent and comes with a citation. But AI produces exactly this kind of plausible authority unreliably: the cited section may be from a superseded edition, may omit an exception, may apply only under a dropped condition — or may not say what the AI claims at all.

Staying in charge

The professional treats the AI’s citation as a lead, not a source — verifying the provision against the actual code (the right jurisdiction, the current edition, the full text including exceptions) before it enters the design basis or the permitting package. The AI pointed; the professional confirmed against the authoritative text.

What the record documents

What source the professional relied on, that the cited provision was verified against the actual code (jurisdiction and edition), any correction to the AI’s summary, and who adopted the final language.

The situation

An AI tool produces a confident, well-formatted result. On independent review, the checker identifies a subtly bad assumption. The engineer in responsible charge decides to reject the AI output, and the corrected work carries forward.

Staying in charge

A reviewer flags it; the responsible professional decides. Recording the rejection matters — a record that only ever shows “adopted” hides the moments that most demonstrate engagement. This absorbs the back-check / comment-resolution log a competent firm already keeps; it doesn’t replace it.

What the record documents

The disposition, distinctly — the AI output was reviewed, the technical basis for rejection recorded, it was not adopted, and corrected work carried forward, with who decided.

The situation

A work product reaches the responsible PE for sealing. Late, the PE learns an unlicensed staff member used an AI tool during its preparation — and edited the output afterward — without declaring it. Responsible charge is, at its core, about direction and control over other people’s work; this is where it is most often lost.

Staying in charge

Reviewing finished work one had no hand in developing is not responsible charge. So the record has to capture the chain of custody — that AI was used and by whom, what a supervised staff member changed after the AI step, and that the responsible professional directed and reviewed the modified product, not just the original AI output, before sealing. The fix for the “found out at seal time” problem is the declaration made at the point of work, by the person who used the tool.

What the record documents

Who used AI and in what role, what staff under supervision changed after the AI step, and the responsible professional’s directed review of the final, modified product.