Checking an AI-generated note means comparing its clinical meaning with the encounter, not simply correcting its grammar. Tandem's transcript references can help locate supporting conversation, while edit highlighting can show what changed during revision. Neither feature removes the need to read the complete final record.
Tandem's March product update, published on 1 April 2026, describes inline references linking note sections to transcript segments. Its June update, published on 3 July 2026, describes highlighting text added and removed by Adjust. These are distinct review aids: one concerns source context, the other a change between drafts.
Read for meaning before polishing language
Begin with the parts that would alter another clinician's understanding: the reason for attendance, significant history, current findings, assessment, decisions and follow-up. A neatly formatted record can still be wrong about any of these.
Ask whether statements refer to the right person and time. A relative's history should not become the patient's diagnosis. A symptom that resolved before the appointment should not necessarily be recorded as ongoing. An option discussed should not appear as an action completed.
Only after those checks should the review concentrate on preferred wording. Otherwise it is possible to spend time perfecting a sentence whose underlying claim is false.
A fictional encounter for practising the review
Consider this original, simplified account. A patient describes dizziness that occurred earlier in the week but has not recurred today. They deny losing consciousness. An old referral letter mentions a previous medication change, while today's discussion considers whether a further review is needed. The clinician plans to request follow-up, but no appointment has yet been booked.
This is not an actual Tandem transcript or output. It is an editorial scenario designed to expose four common distinctions: previous versus current symptoms, negation, historical versus proposed medication changes, and planned versus completed follow-up.
A reviewer should be able to identify those distinctions before inspecting any generated note. Without a clear account of what matters, a convincing draft can become its own misleading reference standard.
Follow a claim back to its source
Where a note says that dizziness is ongoing, open the relevant transcript reference and inspect the surrounding exchange. A single sentence may be insufficient if the patient corrected an earlier statement immediately afterwards.
The task is to establish whether the source supports the final wording. A link to a real transcript segment proves that a segment exists; it does not automatically prove that the note interpreted it correctly.
If the transcript itself is unclear, do not resolve the uncertainty by choosing whichever version sounds more clinical. Use the approved local process to check the encounter record or clarify the information. Access to recordings and their retention may vary, so do not assume that replay will always be available.
Negation needs more than a keyword search
In the fictional case, "did not lose consciousness" should not become "loss of consciousness". But the review also needs to preserve scope. A denial of one symptom is not a denial of every related concern.
Read the whole phrase and identify who supplied the information. A clinician asking "Did you faint?" is not evidence that fainting occurred. A patient's reply may qualify the question rather than simply answer yes or no.
This is why scanning only for important medical terms can miss a substantial error. The relationship between those terms determines the meaning of the record.
Keep old records in their proper place
Additional context can help a clinician understand a visit, but it should retain its provenance. In the fictional example, the medication change comes from an earlier letter. It must not silently become an action taken today.
A practical review separates source date, event date and today's discussion. If a document was written later than the event it describes, those dates are not interchangeable. Where the source is uncertain, preserve that uncertainty rather than inventing a clean chronology.
No clinical dose is needed to test this distinction. The relevant question is whether the note correctly records what changed, when it changed and whether today's clinician confirmed it.
Use highlighted edits without narrowing the review too far
When Adjust changes a draft, inspect the added and removed text. An apparently minor rewrite might remove the word "possible", shift a historical event into the present or turn a proposed referral into a completed one.
Then reread the surrounding passage. A locally accurate edit can create inconsistency elsewhere, especially if the assessment and plan now describe different decisions.
Highlighting shows changes made during that adjustment. It is not a general error detector for the original note, and unchanged text should not be treated as previously verified merely because it lacks a highlight. Review the complete final version before approval.
Follow-up is an action with a status
In the fictional case, follow-up is to be requested. A note saying "follow-up booked" would make a stronger claim than the encounter supports.
The final record should distinguish the decision, the responsible person or service and the current state of the arrangement. Where information is missing, it should prompt clarification rather than confident completion by the writer or software.
This approach also makes handover more useful. A colleague can see what remains outstanding instead of assuming that a smooth sentence describes work already done.
Turn repeated corrections into a useful improvement
If the same type of correction recurs, record its category without exporting identifiable patient content into an informal spreadsheet. Was the problem unclear audio, speaker attribution, chronology, missing information or a template instruction?
A small review can guide local improvement, but it should not be described as the general error rate of Tandem. The result depends on the encounters, configuration, reviewers and software version assessed. No such audit was conducted here.
This guide is published by iatroX and includes its clinical-reference and learning role. Under September 2026 product information, Ask-iatroX provides free source-linked knowledge support; it does not establish what was said in a consultation. Its CPD tools can support a personal learning record after an improvement exercise, without claiming accredited CME or importing patient-identifiable detail.
For routine review, use source references to investigate meaning and highlights to inspect revision. For recurring problems, investigate the workflow rather than simply editing faster. The final sign-off should reflect a checked encounter record, not trust in the presence of an AI review feature.
Frequently asked questions
Does a transcript reference prove the note is correct?
No. It helps locate source context, which still needs to be checked for attribution, chronology and whether it supports the note's wording.
Do highlighted edits identify every error?
No. They show changes made during an adjustment, not all possible problems in the original or final draft.
Can Ask-iatroX verify what a patient said?
No. Clinical reference can support knowledge checking, but fidelity to an encounter must be established from the appropriate encounter sources and professional review.