A scheduled routine scales its selection rule as well as its actions. If the rule is wrong, a workflow can consistently produce the wrong cohort while every message is delivered successfully. For Heidi II, population-level evaluation therefore needs to examine who is included, who is excluded and what happens when the underlying assumptions change.
Heidi's 28 September 2026 announcement says Routines can be scheduled and applied across patient lists. The scenarios below are proposed tests of that capability, not reports of inappropriate recalls sent by Heidi. The new capabilities are not available in the UK or EU according to the launch materials.
Start with the rule, not the outgoing message
Imagine a fictional practice looking for patients who appear not to have completed an agreed review. The proposed routine selects people with no completion entry in a particular part of the record and prepares follow-up communications.
The message may be clear and the destination correct. The first question, however, is whether absence of that entry reliably identifies unfinished work in this practice. A completed review might be documented elsewhere, a record might await reconciliation or the patient's individual plan might differ from the default process.
These possibilities do not mean automation is unsuitable. They mean the selection rule needs testing against the way the service actually records and delivers care. A polished communication cannot repair an incorrect eligibility assumption.
Missing information is not a complete clinical answer
In the fictional cohort, one patient completed the review with another provider. Another has a documented decision to defer the action. A third has a coding mismatch, while a fourth genuinely has outstanding work. The same missing field could therefore represent several different situations.
A useful routine might separate clearly eligible cases from those needing clarification. It might also present the basis for selection so that a reviewer can inspect the uncertainty. The goal is not to make the system infer an answer where the record does not contain one.
These are proposed behaviours for evaluation. The public launch description does not establish that every deployment can reconcile all such circumstances automatically, and it would be equally unjustified to assume that none can.
Count false exclusions as well as false inclusions
A practice can inspect the generated list and still miss an important class of error: patients who should have appeared but did not. Reviewing only selected cases gives no direct information about those exclusions.
For a proposed evaluation, an independently reviewed reference cohort should include both selected and non-selected records. Assessors could then identify inappropriate inclusion, inappropriate exclusion and cases where the available information was insufficient for a firm classification.
The denominator should remain explicit. The proportion of selected patients who truly qualify is different from the proportion of all qualifying patients identified. Neither is interchangeable with message-delivery success or the proportion of recipients who respond. Combining those measures into a single success percentage would conceal the problem the routine is supposed to solve.
Stage deployment around what has been demonstrated
A proposed introduction would begin with a preview of the cohort and its selection rationale, without external action. Reviewers would examine exclusions as well as inclusions and resolve how uncertain cases should be handled.
A bounded pilot could then test the authorised workflow in the actual service, with clear responsibility for exceptions and an agreed stopping process. Expansion should depend on the results and the stability of the selection rule, not simply on the fact that the first batch completed without a technical error.
The pilot should also account for work created elsewhere. An apparently efficient recall process could increase calls, reconciliation tasks or duplicate correspondence. Those consequences should remain part of the assessment rather than being attributed to other teams and excluded from the product's workload claim.
A routine can become wrong after being right
The practice's process may change, an external service may alter its arrangements or new information may arrive after the cohort was prepared. A previously appropriate rule can therefore become unsuitable without any change to the agent's technical ability.
The evaluator should ask when eligibility is checked again. Does the routine use the original snapshot, the current record or some combination? Can the responsible clinician see what changed between selection and action? What happens to scheduled work after the governing instruction is withdrawn?
These are practical specification questions. The answer may legitimately differ by workflow, but the organisation should know which behaviour it has authorised. A schedule is not a standing justification to continue when the premise no longer applies.
Test stopping and containment at the same scale
If a problem is identified, the practice needs to know which cases have been selected, which actions are pending and which have already occurred. Stopping future work is different from managing the consequences of completed work.
A proposed simulation could change an eligibility rule partway through a test cohort. The system and team would then need to contain further action, identify affected cases and make the remaining state understandable. Another scenario could introduce an ambiguous external response for only part of the cohort.
The assessment should not reward a reassuring stop message if pending work continues elsewhere. Nor should it assume that already delivered communications can simply be erased. Recovery may require human review and appropriate follow-up rather than a technical reversal.
No population-level Heidi II test has been run for this article. Results should come from a documented evaluation with defined records, criteria and outcomes.
Compare with the real manual process
Manual recall systems can also select the wrong people, miss eligible patients or continue using outdated assumptions. The fair question is whether the proposed routine improves on the current process while making its limitations visible.
A useful comparison would measure selection quality, unresolved cases, staff effort and relevant patient experience in both workflows. It should account for the possibility that automation makes pre-existing data-quality problems more visible rather than creating them. It should also distinguish those cases from genuinely new errors introduced by the routine.
This approach avoids both idealising people and assuming that greater consistency means greater correctness. Consistent execution is valuable when the rule is appropriate. It can amplify a problem when the rule is not.
The population-level standard
For clinical leaders, the key decision is whether a bounded use has demonstrated adequate selection, review and containment. For operational teams, it is whether uncertainty and unfinished work remain manageable. For suppliers, it is whether the product can explain the cohort and respond coherently when a shared assumption changes.
A routine should not be judged only by how many patients it reaches. It should be assessed by whether it reaches the appropriate patients, preserves individual circumstances and stops when the reason for acting no longer holds. That is the difference between scaling activity and improving a clinical workflow.
Frequently asked questions
Has Heidi II been reported to send inappropriate population recalls?
This article does not report such an incident. It presents test cases prompted by the scheduled, patient-list workflows described in the September 2026 launch.
Why is reviewing the selected patient list insufficient?
It may reveal inappropriate inclusions but cannot, by itself, identify eligible patients who were omitted. Exclusions and uncertain cases need their own assessment.
What should happen before a routine expands to a larger cohort?
Its selection rule, exception handling and stopping behaviour should be tested in a bounded setting. Expansion should depend on demonstrated performance and the continuing validity of the assumptions.