Candidate deviations: what AI can surface and what only you can judge
Short version: AI is good at finding the facts a deviation opinion rests on: an abnormal result with no response, an order with no matching task, a policy step that was never charted. It cannot tell you what a reasonable physician in that specialty would have done on that date. Ask it narrow, factual questions, check the answers against the pages, and form the judgment yourself with the outcome out of view.
Limits: We read each product's own web pages on 2026-11-12 and did not upload records to any of them, ours included. Prices, plans and features change, and nothing here is legal advice or a substitute for your own clinical judgment.
What "candidate deviation" means here
A deviation is a departure from what a reasonably careful physician in the same specialty would have done in the same circumstances. You decide that. A candidate deviation is a fact pattern worth your attention: a 3-hour gap, a missed repeat test, a consult that never happened. A record tool can list candidates in minutes. Many will turn out to be fine once you read the context, and that is the point: the tool widens the net, you do the sorting.
The AMA's ethics opinion on medical testimony says a physician expert should consider the standards that prevailed at the time of the event under review (AMA Code of Medical Ethics, Opinion 9.7.1). That one line shapes how you should use AI on this task, as the sections below show.
8 deviation patterns and how to ask a record tool about each
Ask for facts with page numbers, never for a verdict. "Was there a delay?" invites a narrative. "List every potassium above 6.0 and the next order or note after each, with times and pages" gives you something to check.
| Pattern | Ask the tool | What it can show | What only you decide |
|---|---|---|---|
| Abnormal result, late response | "Every value outside range from [start] to [end], with result time, and the first order, note or call after each." | The interval in minutes, with pages | Whether that interval was acceptable for that value, that unit, that day |
| Ordered but not done | "Every order in the window and the matching task, administration or result, if any." | Orders with no matching record | Whether the gap is a care failure or a charting gap |
| Monitoring frequency | "Neuro checks were ordered q1h at [time]. List every charted neuro check after that." | Missed intervals | Whether the frequency ordered was itself appropriate |
| Escalation | "Every page, call or message from nursing to a physician, and the physician's next documented action." | Who was told what, and when | Whether the response met the duty of that role |
| Missed follow-up | "Every recommendation for repeat imaging, referral or follow-up, and whether a later record shows it happened." | Recommendations with no later match | Who owned the follow-up |
| Medication | "Every dose of [drug] on the MAR with time given, and any held or missed doses with the reason charted." | Doses, holds and gaps | Whether the dosing met the standard for that patient |
| Informed consent | "Every consent form and every note describing a risk discussion for [procedure]." | What was documented | Whether the discussion met the standard; the chart rarely settles this alone |
| Policy or protocol | "Compare the charted steps in the window to the attached policy, step by step." (Upload the policy as its own document.) | Steps with no charted match | Whether the policy reflects the standard of care, which it may not |
The question patterns work in any record tool that answers with page citations. A general chatbot without your file can only guess at them.
The date-of-care trap in guideline AI
Clinical AI tools are built to answer with current evidence. UpToDate Expert AI says it is built on UpToDate content; ClinicalKey AI says its sources are updated daily; OpenEvidence answers from published literature. That is what a treating physician needs today. A standard-of-care reviewer needs the opposite: what the guidance said on the date of care.
- Ask the literature tool the clinical question to find which guideline and which society govern it.
- Ask it for the publication history: "When was this recommendation first made, and which earlier edition preceded it?" Treat the answer as a lead.
- Pull the edition in force on the date of care from the society's site, an archive or a library, and read the actual text.
- Record in your notes the edition, its date and where you got it, so you can answer the question at deposition.
A recommendation published 6 months after the event is hindsight, however persuasive it reads in a chatbot answer.
Keep the outcome out of the review
Reviewers who know the outcome judge the same care more harshly. In a 1991 study, 112 anesthesiologists rated 21 cases whose outcomes were swapped between temporary and permanent; the share of "appropriate care" ratings fell by 31 percentage points when the outcome was made permanent (Caplan, Posner and Cheney, JAMA 1991). Every med mal file arrives with a bad outcome attached, so the bias is built in.
AI can help you hold it off, or make it worse. A tool that read the whole file will mention the outcome in every answer unless you stop it. 3 habits that help:
- Cut-off questions. "Describe only what the record shows up to 10:30. Do not refer to later events." Write down your view of each decision point before you move the cut-off forward.
- Decision-point framing. Ask what the clinician could see ("labs resulted and visible at 10:30"), not what went wrong ("why was the leak missed?").
- Staged uploads. If your tool lets you add records to a case later, load the records up to the decision point first. This is the only way to keep the outcome fully out of the tool's answers.
Cut-off prompts leak. A tool may still cite a later page, or describe a test as "the CT that showed the leak". Treat any leak as a reason to re-ask, not to give up on the method.
Where AI makes deviation review worse
- Leading summaries. Tools sold to plaintiff firms are built to find problems, and tools sold to defense firms to explain them. Ask the same neutral question of the record and read the pages either way.
- Charting gaps read as care gaps. "No documentation of reassessment" may mean no reassessment, or a nurse who did it and charted it on paper. The tool cannot tell these apart.
- Policy treated as standard. A hospital policy can be stricter or looser than the standard of care. A tool comparing chart to policy shows non-compliance, not negligence.
- Specialty drift. A general model may describe what an internist would do when the defendant is an emergency physician. Anchor every question to the specialty and setting.
The tools for each step, with dated sources, are on the tools page. The timeline errors that feed bad intervals into this step are on the timeline page.