Evidence, governance & intelligence

Anthropic publishes an open dataset that measures how usage of its AI assistant breaks down across the working world, mapped to the U.S. Department of Labor's occupation catalog. We pulled the latest release and tabulated the May 2026 window ourselves: 718 professions, ranked by share of global usage.

Here is where property work landed. Real estate appraisers and assessors came in at 0.04% of global usage. Construction managers and property managers sat at 0.01% each, and construction and building inspectors rounded to 0.00%. Computer and mathematical roles, for comparison, accounted for 23.8%.

Property due diligence sits at the very bottom of the board. The obvious explanation would be that the work does not fit AI. The obvious explanation is wrong. Assessment work is photographs, checklists, condition calls, cost tables, and public-record research. Reading images, drafting structured text, and pulling government data are among the things this technology already does well.

The real explanation is that the profession has a requirement most AI tools were never built to meet. Every finding in an assessment ends up under a professional's signature. A chat window that produces a confident paragraph with no chain back to the evidence is not a time-saver in this field. It is a liability generator. Assessors are not behind on AI. They are correctly refusing tools that cannot survive the question a lender, a buyer, or a courtroom will eventually ask: how do you know that?

That is the standard we build SITUS to. Every AI-drafted finding is bound to a checklist item and tied to the photograph it came from. Condition and cost proposals arrive in ASTM E2018 phrasing with their basis attached. Validation blocks an incomplete report from shipping. The reviewer confirms or overrides every call, every override is logged, and nothing reaches the report until a person approves it. The AI clears the gathering and the first pass. The professional keeps the signature, and the evidence chain proves the reasoning behind it.

This is why we think the debate about what AI should do in technical work misses the point when it stops at speed, and equally when it stops at searching what a firm wrote in the past. Precedent tells you how to phrase a finding. Evidence is what makes it true. The axis that decides adoption in a liability profession is neither speed nor recall. It is whether the AI's contribution to the report you are signing today is traceable, reviewable, and defensible the day someone challenges it. Speed is a consequence of getting that right. So is trust.

A profession at 0.04% is not a profession AI forgot. It is a profession still waiting for tools built to its standard. The firms that adopt on that standard first will compound the advantage for years while the rest of the industry waits for someone else to prove it out.

We make living intelligence.

Source: Anthropic Economic Index, open dataset, May 2026 window. Figures are our own tabulation of global usage share by occupation from the public release at anthropic.com/economic-index.

At a trade show this spring, a senior environmental professional told me she walked past a booth selling AI-generated Phase 1 reports for the price of one hour of human review. She wasn't worried about her job. She had a question: who signs that?

That question is the whole story.

AI is good at making cheap work cheaper. It summarizes a document, captions a photo, drafts a paragraph. The parts of due diligence that were already close to commodity will compress fast, and price will follow. None of that produces the answer a lender or a buyer pays for. That answer still comes from a professional who puts their name on it.

Speed is the easy part. The hard part is accountability, and it does not move. When an engineer or environmental professional signs an assessment, they vouch for the reasoning behind every finding. AI can do the gathering and the first pass. It cannot hold the liability. For AI to belong in this work, every claim it surfaces has to be auditable: tied to its source, reviewable, and overridable by the person who signs.

This is where general-purpose AI struggles. A chat tool hands you a confident paragraph with no chain back to the evidence. A tool built for the liability field treats the signature as the point, and everything upstream as support for it.

We built SITUS around one assumption: AI changes the speed of the work, not the accountability for it. Findings trace to the photo they came from. Validation blocks an incomplete report from shipping. The reviewer confirms or overrides, and every override is logged. The professional still owns the call. They spend their judgment where it matters instead of on the busywork.

The firms that thrive over the next few years will not be the ones that generate reports fastest. They will be the ones that can still prove their reasoning.

We make living intelligence.

Most AI tools for assessments are chat windows or PDF summarizers. You upload photos, ask questions, and get suggestions. What you don't get is structure, validation, or an audit trail. When the report has to hold up later, that gap is the whole problem.

SeaGoat works the other way. It is a governed workflow engine: evidence goes in, validated outputs come out, and every step stays on the record.

What That Means

Evidence becomes structured records. Photos and documents don't just sit in folders. They become evidence objects with IDs, timestamps, and lineage. When you attach a photo to a finding, that binding is permanent and traceable.

Validation gates enforce quality. You can't mark an item "Poor condition" without evidence. You can't create a "Repair" action without a corresponding cost entry. The system blocks outputs when required fields are missing.

Human review is mandatory. AI suggests findings based on evidence analysis. Humans approve, edit, or override. Overrides are logged with rationale. Nothing ships without explicit human approval.

Outputs are reproducible. Reports can be regenerated from saved workflow state to verify consistency. Every cost estimate traces back to its source finding. Every finding traces back to its source evidence.

Why This Matters

In liability-heavy workflows (property due diligence, industrial operations), decisions need to hold up under scrutiny. Chat AI can't provide that. You need:

• –Clear chains from evidence to conclusion
• –Validation that prevents incomplete data from shipping
• –Audit trails showing who approved what and when
• –Reproducible outputs that can be verified

That's governed workflow. Evidence → validation → review → outputs. With structure, not just suggestions.

We make living intelligence.