Independent research · AEC risk & assurance

The 2026 E&O Shift

AI is being written into — and out of — the policies that stand behind stamped engineering and design work. Here is what the paperwork actually says, and what is being sold on top of it.

Something real is happening to professional-liability coverage in the built environment, and something overstated is being sold alongside it. This report separates the two, for the people who carry the liability: the licensed professionals who stamp the work, and the principals who buy the policies behind it.

1What actually changed

For most of the last decade, AI sat inside professional practice without being named in a policy. Insurance commentary calls this “silent AI”: the wording neither granted nor excluded it, so any claim fell to be decided on the policy’s existing terms, definitions, and the facts. Silence is not the same as coverage — a point worth holding onto, because it cuts both ways.

What changed in 2026 is that the silence is now being closed on purpose, and the first visible instrument is a set of Verisk/ISO endorsements filed with a January 2026 edition date: CG 40 47 (the broad form, on Coverage A and B), CG 40 48 (Coverage B only), and CG 35 08 (products and completed operations). Two facts about them matter more than the form numbers.

First, they are optional. They are tools a carrier may attach; they do not switch on across the market on a fixed date. Whether one reaches your policy depends on your carrier, your risk profile, and your renewal cycle.

Second, they are specific to generative AI — machine systems that produce content such as text, images, code, or drawings. That word does a lot of work. A finite-element solver or a code-checking routine is not generative AI; an LLM that drafts a specification or generates a plan is. For an engineer, the generative line is the whole question, and the ISO forms sit squarely on it.

These are ISO forms on the general-liability line, and for stamped design work that line is largely beside the point: most design-firm CGL policies already carry a professional-services exclusion (the CG 22 43 family), so a professional-error claim was never going to be paid there anyway. The general-liability AI endorsements bite hardest for contractors — estimating, scheduling, site-safety tools — whose exposure does sit in CGL. The line that carries a design professional’s core exposure is errors & omissions, and that market is moving on a separate, less standardized track.

The clearest evidence there is Berkley’s “absolute” AI exclusion, in force on its D&O, E&O, and fiduciary products. It is far broader than the ISO generative forms: it reaches any claim “based upon, arising out of, or attributable to” the use, deployment, or development of AI “by any person or entity” — which, read literally, extends to a subconsultant’s tools, not just the insured’s. But note the boundary of what that establishes. Berkley’s form is a general professional-liability instrument; whether an equivalent exclusion is being written onto architects-and-engineers professional-liability programs specifically — the policy most design firms actually buy — is not something public sources settle. The honest read is that the E&O market is turning, with one broad carrier form to point to, and the A&E-specific picture still forming.

2The rule that did not change

It is tempting to read all this as a new kind of liability arriving with AI. The more accurate reading is that the insurance market is catching up to a rule the engineering profession has held all along.

In 2024, NSPE’s Board of Ethical Review took up AI-assisted work directly in Case 24-2. The facts are ordinary: an engineer used AI to draft a report and to help produce design documents, reviewed the text carefully but the drawings only cursorily, and missing safety features and misaligned dimensions slipped through. The Board’s line is quotable and clear:

“AI-generated technical work requires at least the same level of scrutiny as human-created work.”

— NSPE Board of Ethical Review, Case 24-2 (July 18, 2024). The Board found AI use was not unethical in itself; the failure was not maintaining responsible charge over the output before sealing it.

The profession has been here before. Case 24-2 reaches back to Case 90-6, which worked through the same worry about computer-aided drafting in 1990: a tool can produce the geometry, but the licensed professional owns the judgment that lets it be sealed. AI is that question with the stakes raised, not a new one.

Two boundaries keep this precise. The Board’s opinion is professional ethics, and it says outright that it “does not attempt to address any of the potential legal considerations.” The instrument that actually binds is licensure law — the responsible-charge provisions in state statutes and board rules — which is where the accountability lives whatever a policy says. Read that way, the insurance movement is coherent: carriers are responding to an exposure that professional accountability already places on the licensed individual. (That the ethics came first is analysis, not a claim about any single insurer’s motive — exclusions get written for aggregation, uncertainty, and litigation cost too.)

3What is real versus what is being sold

This is where most of what circulates on the topic goes wrong. A recognizable narrative has formed: that named carriers are excluding AI on design-professional E&O today, that firms must stand up a defined four-part governance program to earn a coverage “carve-back” — an exception written back into an exclusion, restoring cover on stated conditions — and that underwriters are already demanding governance documentation at renewal. It is a tidy, alarming, action-forcing story, and most of it traces to a single source.

The distinction that keeps this fair: the broad AI-exclusion trend is real and better-sourced than the vendor narrative (Berkley’s form is public; insurance-law firms have written it up). What traces to the single placement-firm source is the narrower, more sellable layer — that this has specifically hardened into an A&E-E&O documentation requirement you can buy help satisfying today. The table sorts one from the other.

One data point does cut through, and it is A&E-specific: Ames & Gough’s 2026 architects-and-engineers professional-liability survey found that 80% of the insurers it surveyed view AI as a potential disruptor of that market. That tells you underwriters are paying attention. The same survey did not find carriers requiring AI-governance documentation at renewal — so “insurers are watching AI closely” is well-supported, while “insurers require a documentation package today” is not yet the same statement.

4What it means if you carry the liability

Strip out the marketing and a short list of no-regret implications is left. None of them requires buying anything.

The exposure is individual, and it always was. If your stamp goes on AI-assisted work, the accountability is yours under licensure law regardless of what a policy says. And keep the two apart: an insurance exclusion does not create liability — it changes whether the insurer pays for a liability that already exists. That is the most stable fact in this landscape.

Read your next renewal for AI language — on the E&O policy, and read the definitions and exclusions, not just the endorsement schedule. A&E professional liability is written on carriers’ own manuscript forms, not standard ISO forms, so a change can arrive as revised base wording or as new application questions rather than a labelled endorsement.

Being able to describe how AI is used in your practice is defensible on its own merits — which tools inform which decisions, how output is checked, who signs off. Treat that as sound risk management and good client hygiene, worth doing whether or not any carrier ever asks; it is a weaker claim, and a more honest one, than “this improves your position in a claim.”

If an exclusion does appear, you have moves before you have a problem. Ask whether the wording is generative-only or absolute; ask for the narrower CGL variant (CG 40 48 rather than CG 40 47) where relevant; and, because adoption varies by carrier, be willing to test another market. A broad “by any person or entity” trigger is also worth probing for how it treats a subconsultant’s AI use, which can reach you.

Do not buy a governance product to satisfy a requirement no one has shown exists. If a vendor says underwriters demand a specific documentation package today, ask them to name the carrier and show the endorsement language. The defensible answer right now is that this is where the market is heading, not where it verifiably is.

5The questions only your renewal can answer

Public sources take this story as far as this report has taken it and no further. What actually affects your firm turns on facts that are not published anywhere. Put these to your broker or underwriter before acting on anything in this space:

1. Has an AI-related exclusion been attached to our E&O policy, or is it likely at our next renewal?
2. Which line does it sit on, and what does it exclude — all AI-assisted work, or defined generative-AI exposures?
3. Does it reach AI used by our subconsultants, or only our own?
4. Is any carve-back or credit genuinely on offer for a documented AI-governance program — on what terms, in whose words?
5. If documentation is wanted, is it evidence of a workflow the carrier can inspect, or a signed attestation? These mean very different amounts of work, and this is the most consequential question for what a firm would actually have to build.
6. What is the pricing or capacity impact — of the exclusion, and of any credit?

A firm that walks into renewal already asking these is in a better position than one reacting to a form it did not expect. Most A&E firms have not yet hit their first renewal since the January 2026 forms appeared, so for much of the market this is still a question to prepare for rather than a change already received.