Every professional engagement rests on an assumption that is rarely stated and almost never tested: that someone involved has the knowledge to verify whether the work is correct.
This assumption held for most of professional history because the barrier to producing expert-level output was the expertise itself. A structural engineer's calculations required structural engineering knowledge to produce. A regulatory submission required regulatory knowledge to navigate. The output and the competence to verify it arrived together, because they came from the same source.
AI has broken that coupling. The output no longer requires the competence. The verification still does.
What Verification Actually Requires
Verification is not proofreading. It is not checking whether a document is internally consistent, whether the grammar is correct, or whether the formatting matches the template. These are surface properties that any careful reader can assess.
Verification in a professional context means something specific: the ability to determine whether the substantive claims, judgments, and recommendations in a piece of work are correct given the actual conditions to which they will be applied.
A market entry analysis can be verified by someone who has executed market entries in that sector and knows where the standard frameworks fail to capture what actually happens. A compliance framework can be verified by someone who understands how the relevant regulator interprets the requirements, not just what the requirements say. A technical specification can be verified by someone who has built what is being specified and knows which tolerances matter and which are theoretical.
Verification requires domain depth. It requires the ability to identify not just what the document says but what it does not say, what it gets subtly wrong, and where the gap between the written analysis and the operational reality will create problems. These are not skills that come from familiarity with a field. They come from sustained engagement with the failure modes of a field.
The Proliferation Problem
AI has not created a world where verification is impossible. It has created a world where the volume of output requiring verification has increased faster than the supply of people capable of providing it.
This is the operational reality that most organizations have not confronted directly. When a compliance team that previously produced ten documents a month now produces forty, the verification requirement has quadrupled. The number of people with the domain depth to verify those documents has not changed. In most organizations, it has not increased at all.
The result is one of two outcomes. Either the same verification capacity is spread across four times the output, reducing the depth of review on each document to a quarter of what it was. Or the organization maintains review depth on a subset of documents and accepts, explicitly or implicitly, that the remainder are going out unverified.
Neither outcome is advertised. Both are common.
The downstream consequence is a body of work that looks complete, looks professional, and contains errors that no one in the organization has the knowledge to find. Those errors remain invisible until the work is applied to the real conditions it was meant to address.
Who Catches What AI Gets Wrong
The errors AI produces in professional output are not random. They follow a predictable pattern that reflects how large language models work.
AI systems generate outputs by predicting what text is likely to follow given the inputs provided. They are trained on the documented outputs of expertise, which means they are very good at producing text that resembles correct expert output. They are not good at the things that distinguish genuinely correct expert output from plausible but wrong expert output.
Specifically, AI systems are poor at accounting for context that is not in the text. The regulatory interpretation that is applied differently in practice than the written rule suggests. The market dynamic that every operator in the sector knows about but that does not appear in any document the model was trained on. The engineering constraint that is obvious to someone who has built the thing and invisible to someone who has only read about it.
These are precisely the errors that require genuine domain expertise to catch. A reviewer without that expertise will read the same output and find nothing wrong because they are checking the text against the text, not the text against reality.
This is not a hypothetical failure mode. It is the standard failure mode for AI-assisted professional work when verification is performed by someone with insufficient depth to do it properly.
The Client's Position
Organizations that engage outside expertise face a specific version of this problem. When a deliverable arrives, the client is typically not positioned to verify it independently. This is, in many cases, the reason the expertise was engaged in the first place.
The traditional assumption was that this was acceptable because the deliverable came from someone whose engagement with the subject was deep enough to have caught the errors themselves. The expertise was its own verification mechanism.
That assumption no longer holds universally. A deliverable produced with AI assistance by someone without genuine domain depth will arrive looking identical to one produced by someone with it. The client has no reliable way to distinguish between them using the surface properties of the document.
This creates a structural problem for how professional engagements are evaluated and trusted. The traditional proxies for quality — credentials, reputation, the apparent coherence of the work product — have always been imperfect. They are now significantly less reliable than they were. The gap between what a deliverable looks like and what it actually is has widened, and the client has no new tools to close it.
The Institutional Response That Has Not Yet Arrived
Regulated industries are beginning to address this. The FDA's framework for AI in manufacturing, emerging guidance on AI use in financial services, and the EU AI Act's high-risk classification for AI systems used in consequential decisions all reflect a recognition that output-based evaluation is insufficient when AI is producing the output.
These frameworks share a common logic: the question is not whether the output looks correct. The question is whether the process that produced the output is one that a qualified person can stand behind. This requires the qualified person to have been genuinely involved in the work, not merely to have reviewed a document at the end.
In unregulated contexts, no equivalent framework exists. The market for professional services outside regulated industries has not yet developed the verification standards that the shift in how work is produced requires.
The consequence is a period in which clients are systematically less protected than they were, without knowing it. The work they are receiving is produced under conditions that have changed materially. The standards by which they are evaluating it have not.
What Adequate Verification Looks Like
Verification that is adequate to the current environment has three properties that distinguish it from review that is merely formal.
It is performed by someone with genuine domain depth, not familiarity. The difference is the ability to identify errors that are plausible but wrong, not just errors that are obviously wrong.
It engages with the work at the level of judgment, not just content. A reviewer who can confirm that a document addresses all the required topics has not verified the document. A reviewer who can assess whether the recommendations are correct given the specific conditions to which they will be applied has.
It is documented as a substantive act, not a sign-off. In regulated contexts, the documentation of review is the evidence that verification occurred. A signature on a document does not establish what was reviewed or to what standard. The record of verification needs to reflect the actual engagement with the work.
These properties describe verification as it has always been done well, by experienced professionals who took the responsibility seriously. What has changed is that they now need to be applied to a larger volume of work, produced under conditions that make the need for them greater, in an environment where the incentives to shortcut them have increased.
The verification problem is not that AI produces unverifiable work. It is that the supply of people capable of verifying it has not kept pace with the demand that AI has created for their judgment.