Tandem's widely cited 29% time-saving figure comes from retrospective clinician reports of time spent completing a typical note, not a controlled measurement of all administrative work. In the study published on 10 July 2026, respondents estimated 6.69 minutes without the scribe and 4.72 minutes with it. The findings are encouraging, but their scope matters.
The source is Sanmark and colleagues' evaluation in JMIR Medical Informatics, conducted within Capio Ramsay Sante in Sweden. It deserves closer reading than a promotional percentage, particularly following Tandem's September funding announcement. A study can provide useful evidence without supporting every interpretation attached to its headline.
There are two datasets, not one enormous clinical trial
The July 2026 paper combines operational metadata from 236,153 notes generated by 1,295 clinicians with paired survey responses from 177 sustained, fully onboarded users. The note dataset describes activity recorded within the software. The survey asks clinicians to recall their experience with and without it. Those are complementary sources, but they do not measure the same thing.
| Evidence in the July 2026 study | What it contributes | What it does not establish |
|---|---|---|
| Metadata from 236,153 notes involving 1,295 clinicians | Recorded editing activity inside the scribe | A controlled comparison of complete documentation workflows |
| Paired retrospective responses from 177 clinicians | Perceived change in note-completion time and experience | An objectively timed preimplementation baseline |
| A selected group of sustained users | Experience among clinicians who adopted the tool extensively | The experience of everyone offered it, including people who stopped |
All study figures and definitions in this article refer to the 10 July 2026 publication. Its observation window ran from 26 April 2024 to 27 October 2025. It should not be treated as an evaluation of every product function or software release subsequently introduced.
The large note count is valuable for describing recorded behaviour. It does not turn the survey result into a finding from hundreds of thousands of independently assessed clinicians or patients. Repeated notes from the same people are also not independent observations of clinical benefit.
What the 29% calculation actually means
The reported means differ by 1.97 minutes per typical note. Relative to the recalled 6.69-minute baseline, that is approximately 29%. The publication identifies these as estimated times, collected through paired retrospective ratings rather than a stopwatch comparison conducted before and after implementation.
The paper's limitations section makes the boundary narrower still: the survey question concerned finalising a typical clinical note. Respondents may not have included record navigation, prescriptions, referrals or other administration. It is therefore inaccurate to translate the result into a 29% reduction in all administrative work, all consultation time or the length of the working day. The authors discuss this explicitly.
That does not make perceived time worthless. Clinicians' experience of workload is important, and a recalled comparison may identify a meaningful change. It does mean that a practice should not multiply its entire administrative budget by the headline percentage and call the result a demonstrated saving.
There is another operational distinction. A shorter note-completion task may create time that is then used for another activity. That could be beneficial, but it is not automatically a cash saving, an extra appointment or an earlier finish. The destination of any released time needs to be observed separately.
Who was represented in the survey?
The survey was restricted to users who had worked with the scribe for at least three months, produced more than 100 notes, generated a document or certificate and used its conversational editing function. Of 451 invited clinicians, 177 responded, a response rate of 39.2%, as reported in the July 2026 paper.
Those criteria favour a particular question: what do established users report after making the product part of their work? They are less suited to estimating what happens when an organisation offers the software to everyone, including staff who never fully adopt it or stop using it.
This is selection, not an accusation of misconduct. A clinician who finds a product useful is more likely to continue using it. An analysis confined to continued users can consequently miss important implementation failures or reasons for abandonment. The paper acknowledges this limitation and the predominance of general practice in the survey cohort.
The authors also performed a sensitivity analysis addressing survey nonresponse using observed characteristics. That is useful, but it cannot recreate the experience of excluded non-adopters or supply the missing objective baseline. Statistical adjustment cannot answer a question for which the necessary observations were not collected.
Why the editing logs do not solve the baseline problem
The July 2026 evaluation reports a median in-tool editing time of 93 seconds in the analysed notes. This is a different measure from recalled total note-finalisation time. It cannot simply be subtracted from the survey's pre-scribe estimate to calculate a second, larger saving.
The paper also excluded 138,196 notes with no recorded editing time. A zero in that field did not reveal whether the note had been accepted unchanged, discarded, or edited later within the medical-record system. The researchers did not have the downstream information needed to settle that question. Source: the study's methods and limitations.
This is a useful lesson beyond Tandem. An activity log measures what the software can observe. Work done outside that software, interrupted activity and abandoned outputs may need different measurement. A fast final click is not necessarily a fast complete workflow.
The appropriate interpretation is that the metadata characterise editing behaviour within the tool. They do not provide an objective estimate of total time saved against the previous workflow.
What the study did not test
The study was single-arm and observational, without a randomised control group. Retrospective recall can be influenced by experience, expectations and the difficulty of remembering fragmented tasks. Changes in workload or other aspects of practice cannot be separated from the tool's effect in the way a well-designed comparative study might allow.
The authors reported improvements in average ratings of administrative stress and perceived presence with patients. They also reported a median change of zero on each of those scales, showing why a favourable average should not be described as an improvement experienced by every respondent. These were clinician-reported experiences, not patient assessments.
Crucially, the paper says that it did not assess note accuracy, error types, clinical quality or patient safety outcomes. It therefore does not establish that the scribe improved diagnosis, reduced harm or produced more accurate records. An efficiency signal and a safety finding are different claims. These boundaries are stated in the publication.
How to read the declared relationships
The study arose through cooperation between Tandem and Capio. Its disclosures identify Tandem employees with stock options, a cofounder with equity, a Capio employee and a spousal relationship between an academic author and a Tandem employee.
The paper also states that the academic coinvestigators independently handled the statistical analysis, interpretation and submission decision, and that the analyst had no vendor affiliation or compensation for the work. Both parts of that account belong in a fair appraisal. Describing the evaluation simply as independent would omit relevant relationships; dismissing its findings solely because those relationships exist would avoid examining the methods. The funding and conflict-of-interest statements are available in full.
The sensible response is to inspect the definitions, exclusions, analysis and scope of the claims. Transparency makes that possible. It does not remove the need for further evaluation in other settings.
A practical local evaluation, proposed rather than reported
A practice considering a scribe can use this study as a starting point without importing its percentage into a business case. The following is an original evaluation framework, not an account of tests performed here or results reported by Tandem.
Start by defining the task boundary. A usable clinical record may require drafting, checking, correcting, transferring and completing associated work. Record which activities are included, where timing starts and ends, and how interruptions are handled. Measure the same boundary before and after introduction.
Then retain the full implementation denominator. Record who was offered access, who tried it, who continued and who stopped. Ask discontinuers why. A satisfactory experience among frequent users and a poor overall adoption rate can coexist; averaging only the former conceals the latter.
Review quality separately from speed. An appropriate local sample could examine unsupported statements, omitted information, negation, attribution and follow-up. Define how disagreements are resolved and how material problems are escalated. These are proposed review categories, not claims about observed Tandem errors.
Finally, look for displaced work. Ask whether an apparently shorter note task creates more checking for another colleague, whether patient messages require additional correction, and whether the time released is actually usable. Report clinician experience alongside these measures rather than using satisfaction to stand in for accuracy.
Where feasible, plan a prospective comparison, retain the product version and record concurrent workflow changes. A modest but transparent evaluation is more useful than a precise-looking return-on-investment number assembled from incompatible measures.
The same standard applies to reference and learning tools
For a referenced answer, useful evaluation includes whether the citation supports the claim and whether the response addresses the intended context. For a tutor or simulation, it includes what a learner can subsequently explain or perform, not simply whether they enjoyed the session. These are proposed assessment questions, not reported iatroX outcome results.
As described in iatroX's methodology, checked on 14 September 2026, its retrieval, citation-grounding and output-checking processes are product design features. They are not proof that every response is correct. The same evidential discipline should apply when iatroX discusses itself and when it reviews another platform.
Tandem's July study provides an encouraging account of perceived efficiency among sustained adopters. The appropriate next step is neither to dismiss it nor to enlarge it into a universal productivity claim. It is to measure the complete task, retain the people who did not benefit, and examine quality alongside time.
Study source: Sanmark E, Vartiainen V, Sanmark J, Wettin K, Saari L, Entezarjou A. Evaluation of an AI Medical Scribe After 236,153 Notes Generated Across Care Levels in a European Health System: Mixed Methods Retrospective Observational Study. JMIR Medical Informatics. 2026;14:e90052. doi:10.2196/90052. Published 10 July 2026. Copyright 2026 the named authors; originally published in JMIR Medical Informatics under CC BY 4.0. This article provides original commentary and a proposed local evaluation framework.
Frequently asked questions
Was the Tandem documentation-time study randomised?
No: the July 2026 study was an observational evaluation without a control group. Its main comparison used retrospective reports from a selected group of sustained users.
Does the 29% reduction mean all administrative work?
No: it refers to the reported reduction in recalled time for completing a typical note, from 6.69 to 4.72 minutes. It does not establish the same reduction in referrals, prescriptions, other administration or the working day.
Did the study demonstrate improved patient outcomes?
No: the paper states that it did not assess clinical quality, note accuracy or patient safety outcomes. Perceived presence with patients was reported by clinicians and is not equivalent to a measured patient benefit.
Explore clinical AI evidence and evaluation with iatroX Insights →