Doctors will let AI read a scan but not decide a treatment. Estimators should draw the same line
A new healthcare survey shows clinicians accept AI for imaging but reject it for treatment calls and documentation, citing thin real-world evidence. Construction has the identical split between measurable AI and judgment AI — and almost nobody is asking vendors for the same proof.
Clinicians will let AI read their scans. They won't let it decide a patient's treatment — and the reason isn't fear of the technology, it's a lack of proof it works outside a narrow, measurable task. A survey released this month found that even as physicians' daily AI use nearly quadrupled year over year, 74% of clinicians distrust AI outputs because of hallucinations, and the same doctors using AI the most are the ones most resistant to letting it touch clinical judgment. Construction has the identical two-tier problem with estimating and scheduling AI, and almost nobody in the industry is demanding the same proof.
What did the survey find?
Wolters Kluwer's 2026 Future Ready Healthcare survey, conducted with Ipsos among 355 US doctors and nurses in March, tracked a sharp jump in AI use alongside a sharp jump in specific, named worries:
| Metric | 2025 | 2026 |
|---|---|---|
| Physicians using AI multiple times daily | 10% | 38% |
| Nurses using AI multiple times daily | 16% | 32% |
| Aware of a formal AI governance policy | 21% | 27% |
Layered on top of that adoption curve: 74% of clinicians worry AI is deskilling them, 74% distrust its outputs due to hallucinations, 72% worry advertiser-driven AI business models distort medical recommendations, and 77% say they double-check AI answers against a primary source before acting on them.
Why do doctors trust AI for scans but not for treatment?
Imaging and diagnostic AI earned its place the hard way — years of peer-reviewed studies and clinical trials scored against a clear, physical answer: the scan is read correctly or it isn't. A published review of radiology AI adoption found clinicians resist AI most when it's handed full autonomy over their core judgment calls, and accept it most when it's a supportive tool checked against a verifiable result. Documentation AI, treatment-recommendation AI, and patient-communication AI haven't cleared that same bar of published, large-scale, outcome-checked evidence. So clinicians use them daily and trust them conditionally — reviewing, correcting, and refusing to hand over anything that can't be double-checked.
What's construction's version of "imaging AI"?
The tools that come closest to clearing the same bar are the ones with a physical fact to check the output against: vision-based progress tracking that compares a jobsite photo to a schedule, or safety-monitoring cameras that flag a missing guardrail or hard hat. Right or wrong is checkable by walking the floor. That's the same structure that let imaging AI earn trust — a narrow task, a verifiable output, and years of field use building the evidence base.
What's construction's version of "treatment decisions"?
Everything that requires judgment and has no agreed ground truth: AI-generated cost estimates, AI-resequenced schedules, AI-drafted change order pricing, AI risk-scoring of subcontractors. Vendors selling these tools rarely publish the equivalent of a clinical trial — real variance data from completed projects, measured against what a job actually cost or how long it actually took. Marketing decks show demo accuracy on cherry-picked examples, not a track record across real bids.
What should an estimator or PM actually ask a vendor?
Before trusting an AI tool with anything that lands on a bid form or a schedule, ask for what clinicians are now demanding from their vendors:
- How many completed, real projects was the output checked against — not demo projects, real awarded jobs?
- What was the variance between the AI's number and the final actual cost or duration?
- Is that data published anywhere, or is it a claim made in a sales call?
- How much human review sits between the AI's output and the number that goes out the door?
If a vendor can't answer the first two with a number, treat the tool as a second opinion, not a decision-maker.
The takeaway
Split your AI tools into two buckets this week. Bucket one — takeoff counts, spec extraction, bid-leveling comparisons, photo-based progress and safety checks — has a physical fact to check against, so trust it more as the vendor's track record grows. Bucket two — full estimate generation, automated schedule resequencing, AI-drafted contract language — doesn't have that yet, so keep a human owning every number that leaves the office until the vendor can show real project variance data.
Construction already floated the fix for this gap — a certification standard for AI vendors, the same role UL and ICC-ES already play for products on a jobsite.
Forward this to whoever owns your next estimate.
Friday one chart. Every week, one piece of data that should change a decision on your project. Subscribe at constructionaibrief.com.
- What did the new healthcare AI survey actually find?
- Wolters Kluwer's 2026 Future Ready Healthcare survey, run with Ipsos among 355 US doctors and nurses in March, found physicians' daily AI use jumped from 10% to 38% year over year — but 74% still fear the technology is eroding their own skills, 74% distrust its outputs because of hallucinations, and 72% worry advertiser-driven business models are steering its recommendations.
- Why do clinicians trust AI for imaging but not for treatment decisions?
- Imaging and diagnostic AI passed through years of peer-reviewed studies and clinical trials measured against a clear answer — the scan is read correctly or it isn't. Newer AI tools for documentation, treatment recommendations, and patient communication haven't been validated at that same scale, so clinicians treat them as unproven even while using them daily.
- Does construction have an AI tool that's already cleared this kind of bar?
- The closest equivalent is vision-based progress and safety tracking — cameras that flag a missing guardrail or measure percent-complete against a schedule. The output can be checked against a physical fact on site, the same way a scan can be checked against a diagnosis. Estimating, scheduling, and bid-pricing AI haven't cleared that bar because there's no equivalent verifiable ground truth published by vendors.
- What should an estimator ask an AI estimating or scheduling vendor before buying?
- Ask for real completed-project data: how many actual bids or schedules the tool's output was compared against, the variance from final awarded cost or as-built duration, and how much human review sits between the AI's number and the number that goes on the bid form. A demo video or a single case study is not that data.
- Is this a reason to avoid AI estimating tools entirely?
- No. It's a reason to price the risk correctly. Use AI for the narrow, checkable tasks first — quantity takeoff counts, spec extraction, bid-leveling comparisons — and keep a human owning any number or schedule call that can't be verified against a physical fact before it leaves the office.