Review each AI-generated SOAP note against what actually happened in the visit, not against how tidy it reads. Check patient and date, subjective report, objective findings, assessment, treatment and plan; resolve missing or contradictory details; confirm recording consent; preserve the edit trail; then sign only after the chiropractor has verified the final record. Never let the tool sign automatically.
SOAP means Subjective, Objective, Assessment and Plan. AI can arrange dictated or recorded material into those headings, but a fluent draft can turn “not discussed” into “normal” or copy a previous visit's finding. The chiropractor remains responsible for the clinical record. A short repeatable review is more dependable than reading the note once and assuming polished prose means accurate care documentation.
Check identity and visit context before reading the prose
Start at the top of the record. Confirm the correct patient, appointment date, visit type, clinician and source material. Check whether the draft came from a live recording, a post-visit dictation or typed shorthand. Those inputs have different failure patterns: recording may mishear a term, dictation may omit a measurement, and copied shorthand may carry yesterday's plan. An initial assessment needs a different completeness check from a subsequent visit. One large public payer's documentation checklist for chiropractic doctors (April 2025 edition) separates the initial evaluation from subsequent visits: a later visit should show a review of the chief complaint, changes since the last visit, an assessment of change and the treatment given that day, including the specific areas and levels of the spine manipulated. Treat it as a reminder to match the note to the encounter, not as a universal template for every payer.
For an illustrative identity error, a clinician records two back-to-back visits. The note generator attaches a shoulder complaint from the second visit to the first patient's record because the recording was stopped late. The opening lines look plausible. The reviewer catches it by comparing the visit time and source recording identifier, quarantines both drafts, and corrects the record before signing. Patient matching is a stop rule, not an ordinary wording edit.
Set up a simple review header: patient identifier, date and time, encounter type, source capture method, draft version, reviewer and sign-off status. It can live in the practice system rather than a separate spreadsheet. If any header field is wrong, stop and correct the source linkage before checking the clinical text. A signed note with the wrong identity is much harder to repair cleanly than an unsigned draft.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowSubjective: verify what the patient actually reported
Read the S section against the encounter notes or recording. Check chief concern, onset or change, location, quality, aggravating and relieving factors, reported function, response to previous care and relevant negative statements. Distinguish the patient's words from the clinician's interpretation. If a symptom was not discussed, remove any AI-generated “denies” sentence. If the patient gave an uncertain date, do not turn it into an exact onset. Review any quoted speech for accuracy and necessity.
An illustrative draft reads, “Pain has resolved and the patient returned to unrestricted lifting.” The actual dictation said, “Pain is less frequent, but lifting at work still causes discomfort.” The reviewer replaces the sentence with the supported version and flags the discrepancy. The first wording might falsely imply a treatment goal has been achieved. This is why a note review should compare clinical meaning rather than only spelling and grammar.
In a similar case, a patient says that a leg symptom occurs “sometimes after sitting”. AI writes “constant radiating leg pain”. The clinician checks the source, corrects frequency and timing, and decides whether the change needs further assessment. The correction is not cosmetic; it affects the record of the condition. A general AI scribe error checklist for medical notes offers neighbouring failure modes, while this review remains tied to chiropractic visit details.
Objective: demand observations and measurements, not guesses
The O section should contain only findings actually observed, tested or measured in that encounter. Check examination method, side, level or region, measurement units, comparison with prior results where relevant, and whether a finding was performed at all. AI must not fill an empty field with “within normal limits”. If a template inserts standard text by default, the chiropractor must remove or confirm it. Verify that any mention of imaging or outside reports reflects a report actually reviewed, with the relevant date.
A common illustrative error is a range-of-motion result copied from the previous visit. The prior record says 45 degrees; the current visit documents 35 degrees. The draft repeats 45 and describes improvement. The clinician corrects the number and the interpretation, then checks whether other objective fields were also carried forward. A useful system highlights reused text so the reviewer can distinguish a deliberate unchanged finding from a silent copy.
Where the visit involved spinal manipulation, the documentation should accurately identify the areas treated and what was performed. Do not infer a procedure from a scheduled appointment type or a billing entry. Check any recorded contraindication or risk discussion when relevant, without using AI to declare the treatment safe. The clinician's judgement and local documentation requirements control the final wording. Ask a qualified adviser when payer or professional rules for a specific service are uncertain.
Another concrete example: a voice transcript hears “left” as “right” in a cervical finding. The AI draft carries the wrong side into both O and A. The reviewer checks the encounter record, corrects both sections and examines the rest of the note for propagated errors. A single transcription mistake can become several apparently independent statements if the model paraphrases it repeatedly.
Assessment: test the reasoning link
The A section should explain the clinician's current assessment using the subjective and objective information that is actually present. Look for an unsupported diagnosis, a confident improvement claim, a contradiction with the measured findings or a copied assessment that ignores a new complaint. AI may propose a tidy sentence, but it does not independently examine the patient. If the assessment has changed, the reviewer should make the reasoning visible in the record rather than simply replacing one label with another.
Consider a note that says “good progress towards functional goal” while the patient reports no change in daily activity and the measurement is unchanged. The chiropractor can document a stable condition, revise the treatment plan or reassess, depending on the actual visit. The AI's upbeat phrase must not survive because it sounds professional. Compare the assessment with the goal and objective measure from the approved plan, then decide what the record supports.
Check for contradictory sections. If S says new numbness, O contains no relevant assessment and A says “no neurological concern”, the note is not ready for sign-off. The clinician decides what examination or referral action is appropriate; the AI draft should merely flag the mismatch. This is a useful reason to make the review checklist an active form with “resolved”, “amended” or “escalated” choices rather than a single tick saying “checked”.
Plan: match treatment, follow-up and advice to the visit
The P section often attracts template creep. Verify the treatment actually delivered, the region and method recorded, advice genuinely given, planned frequency or review date, goals and any referral or escalation. A plan from the previous visit must not be repeated without checking whether it still applies. If no exercise advice was given, remove a generated exercise paragraph. If follow-up was discussed as an option, do not write that it was booked.
Template creep can look like a draft that recommends three visits a week for four weeks because an older treatment plan used that schedule. The clinician discussed a review after two visits. The plan is corrected to the actual discussion, and the practice reviews why the model pulled old text into a current encounter. This is a different failure from speech recognition: it arises from the template and context window. Limit what prior notes the generator can see, or require that copied plans be highlighted for approval.
Another case concerns patient instructions. The AI writes, “Avoid all lifting until the next appointment,” although the clinician advised graded activity within tolerance. That change may alter what the patient does outside the clinic. Compare the note with any handout or after-visit message before signing. If they differ, a clinician decides which is correct and sends a corrected patient instruction through the appropriate channel. The physiotherapy treatment note tutorial covers a related review problem in another discipline.
A ten-note review trial with a visible error log
Consider an illustrative chiropractic practice testing AI on ten visits: two initial assessments and eight follow-ups. A clinician generates drafts from approved capture methods, then reviews each draft against the encounter source before signing. The practice records 120 field checks across the ten notes. It finds five factual corrections, three missing details and two unsupported template statements. These figures describe a hypothetical trial, not a product accuracy claim.
The five factual corrections are: wrong side on one finding, a copied range-of-motion value, an incorrect pain frequency, a wrong follow-up date and a treatment region imported from another visit. The missing details include a change since the last visit, a stated functional limit and an actual treatment performed. The two unsupported statements are “no new symptoms” and an exercise instruction that was not given. The clinician edits each note, marks the error source and checks whether the same error affected another section.
Suppose manual note preparation used 12 minutes per visit before the trial. AI drafting takes 3 minutes, but source comparison and correction takes 8 minutes, giving 11 minutes per note in this illustrative week. Across ten notes, the difference is ten minutes before allowing for setup, consent, quality review and vendor costs. If the quality review is rushed, that minute is not worth saving. A trial should judge record accuracy and clinician confidence alongside time, and stop or narrow the tool if high-risk errors recur.
Give each error a category: capture error, transcription error, unsupported generation, copied prior text, wrong patient linkage or human omission in the input. That makes action possible. A wrong side from transcription needs a different fix from a fabricated normal finding inserted by a template. Re-run corrected test cases after changing a prompt or workflow, and keep the original draft for the authorised audit trail.
Use a sign-off checklist with evidence beside each tick
The review can be a short form in the practice system. A tick alone is weak; add the source or correction for any disputed field. The following filled-in sample is illustrative. It shows how the clinician records what was checked without copying private session details into an uncontrolled spreadsheet.
| Review item | Illustrative finding | Action before signature |
|---|---|---|
| Identity and encounter | Correct patient; follow-up visit | Match draft to appointment and capture |
| Subjective change | Draft said constant; source says intermittent | Amend frequency and linked assessment |
| Objective measure | Prior value copied | Enter current measured value |
| Treatment and plan | Exercise advice not given | Remove sentence; check patient handout |
| Consent and record | Capture permission recorded | Confirm retention and sign-off route |
For every note, ask four final questions: does every clinical statement have a source in this encounter or a clearly identified prior record; are missing fields marked rather than guessed; are contradictions resolved; and does the signed version reflect the clinician's own judgement? If any answer is no, leave the draft unsigned. An administrator may chase a missing recording or attachment, but cannot substitute for the clinical review.
Check consent, edits and retention as part of the note
Patients should know when a visit is recorded or processed by an AI scribe, and the practice should follow its applicable consent, privacy and professional requirements. Record the choice and provide a workable alternative if the person declines the recording method. A broad consent form does not prove that a particular visit was captured appropriately. Check the practice's policy and ask a qualified adviser about the rules that apply to its setting.
Before uploading audio or records, verify the supplier's agreement, permitted processing, subcontractors, access controls, retention, deletion and export. Where health-privacy law requires a signed agreement with any vendor that processes patient information, such as a business associate agreement, confirm that it covers this tool before live use. The chiropractic note consent and data checks gives the pre-recording questions in more detail. This checklist starts once a draft exists but should still verify that the draft came from an authorised capture.
Keep the original AI draft, clinician edits, final signed note and timestamps according to the practice's record policy and system capabilities. If the system overwrites the draft, ask how an amendment and edit history can be reconstructed. Do not change a signed note silently; use the approved amendment process and preserve who changed what and when. Recording every keystroke is unnecessary, but the practice needs enough evidence to explain the final record.
An illustrative records problem appears when a patient asks for a copy of a note. The practice can produce the signed version but cannot tell whether an AI-generated sentence was corrected before signature because drafts were discarded and no edit history exists. The owner should investigate the system's audit functions before wider use. A proper record route should also allow the practice to export its information if it leaves the vendor, without relying on an indefinite audio archive.
Set a weekly sample review during rollout. Check a few signed notes against their source and error logs, with particular attention to copied findings, laterality, treatment actually performed and plans that appear unchanged across several visits. Ask clinicians which checks are taking time and whether a template should be narrowed. If the same unsupported phrase keeps appearing, remove it at the source rather than expecting reviewers to delete it forever.
Decide which missing fields block a signature
Not every blank has the same consequence. A missing punctuation mark can be corrected during ordinary editing; a missing treatment description, absent clinician identity or unresolved new symptom needs the chiropractor's attention before signature. Create a short severity list for the practice. “Stop” includes wrong patient, wrong encounter, invented examination, contradictory clinical information or a plan that may change patient behaviour. “Clarify” includes incomplete visit history or an unverified outside report. “Edit” covers wording that is accurate but hard to read. The reviewer chooses the category, not the note generator.
Picture a follow-up note with a clear subjective report and treatment entry but no record of whether the planned objective measure was checked. The AI fills the gap with last week's value. The reviewer removes that value, checks whether the measure was taken and documents the true outcome. If it was not taken, the note should say so where appropriate rather than pretend it was. The clinician also decides whether the missing measure affects the assessment or the next visit's plan.
Outside imaging reports raise a related problem. The patient mentions that a report exists, but the clinic has not received it. The draft states that imaging “confirmed” a finding. The chiropractor changes this to a factual statement about what the patient reported, records that the report is pending and avoids claiming to have reviewed it. That distinction is easy to miss when AI turns conversational material into formal prose. It also tells the next clinician exactly what needs to be obtained before relying on the report.
Put the severity category in the error log during the first month. If “stop” errors recur, reduce the input scope or pause the tool while the cause is found. If only minor edits recur, change the template or style instruction. This gives the practice a decision based on error type rather than a vague impression that the notes “usually look fine”.
A SOAP draft is useful only when it makes the chiropractor's record clearer without changing the clinical facts. The sign-off sequence is simple but strict: confirm the encounter, compare each section with evidence, resolve contradictions, check consent and storage, preserve edits, then sign. If a tool repeatedly invents findings or obscures its source, step back to dictation or manual notes until the process is repaired.
Further reads
- AI Scribe for Chiropractors: Chiro Tools vs Heidi and Jane — Compare tools before choosing a note workflow.
- AI Scribe Security: Audio, Retention and BAA Vendor Questions — Ask vendors about audio, retention and access.
- Chiropractic Software With AI: Which Features Are Genuine? — Check what the practice system actually automates.
- AI Tools for Chiropractors: Report of Findings, Care Plans and Handouts — Keep patient-facing explanations separate from visit records.
- AI Scribe for Medical Practices: Pick by Country and EHR — Assess the wider scribe buying decision.
- AI Software for Chiropractors: A Data Export Checklist Before You Switch — A chiropractic clinic's checklist for leaving or joining a practice system: export formats, notes ownership, exit terms, a migration test and a cut-over plan.
- Jane App for Chiropractors: The AI Scribe and Online Booking — Step-by-step set-up for Jane's AI Scribe (five free notes a month, unlimited add-on) and online booking in a chiropractic clinic.
- AI Scribe Patient Consent: Implied or Explicit for UK Clinics? — Implied or explicit consent for an AI scribe: what NHS England's 2026 guidance says, what Healthwatch found, and a decision rule for private clinics.
- Chiropractic AI for Clinic Owners: Which Admin Jobs to Start With — A clinic owner's order of play for AI: bookings and recalls first, notes with review, marketing within claim limits, no AI diagnosis.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: CMS, documentation checklist and guidelines for chiropractic doctors (MLN1232664), April 2025; HHS, cloud computing guidance for business associates. Consulted November 2026. Payer and professional requirements must be checked for each practice.