Heidi II is Heidi's newly announced agentic platform, extending its proposition from documentation towards delegated work. Announced on 28 September 2026, with rollout starting on 29 September, it introduces Agents, Memory, scheduled Routines and research access. These announcements establish a product direction, not independent proof of reliable execution or improved patient outcomes. The new capabilities are not available in the UK or EU, according to Heidi's launch announcement.
For readers searching for "Heidi 2", the official branding is Heidi II. The useful question is not simply whether it has become more autonomous. It is which work an eligible clinician can delegate, what remains under review and how successful completion will be demonstrated.
What changed at launch?
The following separates the functions described in Heidi's September 2026 launch materials from the evidence a practice would need before relying on them. These are announced product designs, not results from an iatroX product test.
| Announced component | Intended role | What a buyer should establish |
|---|---|---|
| Agents | Plan and carry out instructed work across practice systems | Which actions work in the buyer's actual configuration |
| Memory | Retain preferences and learn from instructions or corrections | What is stored, its scope and how corrections change future behaviour |
| Routines | Save and schedule repeated workflows, including work across patient lists | Selection accuracy, exclusions, stopping conditions and supervision |
| Research access | Bring published evidence into the working context | Citation support, patient applicability and relevant access entitlements |
| Connections to practice systems | Work with records, calendars, inboxes and other interfaces | Available permissions, reliable completion signals and recovery after interruption |
These descriptions come from the 28 September announcement and 29 September product introduction. A staged rollout means that a published capability is not necessarily enabled for every account, region or connected system on the announcement date.
Customer testimony can explain why someone finds a workflow useful. It cannot establish its failure rate without information about eligible tasks, attempted tasks, exceptions and the work needed to rescue them. Similarly, an impressive demonstration establishes what happened in that demonstration, not the range of circumstances in which the same result will recur.
Approval is more specific than a general promise of supervision
Heidi's 29 September 2026 product introduction describes a starting Manual mode and a selectable Auto mode. It says actions affecting the chart, reaching patients or leaving the system wait for review by default, and identifies prescribing, deleting records and submitting claims as outside the permitted actions in either mode. These are important stated boundaries, rather than evidence that every consequential mistake will be detected. The same introduction should be read alongside the actual configuration offered to an organisation.
This distinction matters because supervision can mean reviewing an entire proposed workflow, approving individual steps or inspecting completed work. Those arrangements distribute attention differently. Before deployment, a practice should identify which arrangement applies to each authorised task rather than assuming that the presence of an approval control settles the matter.
Follow a bounded workflow, not a promotional label
Consider a fictional clinician arranging an administrative follow-up after an agreed hearing review. The clinician asks an agent to prepare the relevant correspondence and route it through the practice's approved process. This is an illustrative assessment scenario, not an observed Heidi II session.
Supplying information means making the agreed plan and relevant context available. Asking for a plan means allowing the system to propose the sequence of administrative steps. Authorising execution means accepting the content, recipient and scope of those steps. Checking completion means establishing that the correspondence reached the intended destination and that any outstanding work is visible.
A useful demonstration would introduce a complication: the recipient has changed, an attachment is missing or the destination returns an ambiguous acknowledgement. The assessment should establish whether the system stops appropriately, asks a focused question and leaves the clinician able to continue. A confident completion message is not enough where the receiving system has not confirmed the intended result.
What remains unproven?
In the public materials reviewed on 29 September 2026, we did not identify a longitudinal Heidi II evaluation measuring retained clinical reasoning or deskilling. That is an evidence gap, not evidence of deterioration. Nor should it erase Heidi's own description of pre-launch clinical testing in its agent trust article. Internal development work and a published, independently assessable outcome study answer different questions.
Operational evaluation should report the proportion of suitable, authorised tasks completed correctly, how often a person intervenes, what happens to unresolved work and whether effort moves to another team. Patient benefit requires its own appropriate measures. Neither a shorter interaction nor a larger number of executed actions demonstrates that care improved.
Cognitive effects also require separate assessment. A clinician may gain time for listening and interpretation, or may become less familiar with a case because its first interpretation arrives already prepared. Both are hypotheses worth investigating. Neither follows automatically from the word "agent".
Assistance can improve performance without answering every question
The evidence base does not justify treating clinical AI assistance as intrinsically harmful. In a randomised study published in Nature Medicine on 5 February 2025, 92 physicians using GPT-4 alongside conventional resources performed better on simulated patient-management tasks than physicians using conventional resources alone. The study assessed a different system and task, not Heidi II deployment or long-term competence. See the original trial.
Its practical relevance is that the clinician-system partnership deserves measurement. A model's isolated performance is not interchangeable with the performance of a busy clinician supervising its work. A product could improve immediate task quality while creating new recovery requirements, or save administrative effort without changing clinical reasoning at all.
What would justify stronger conclusions?
A stronger case would combine representative deployments, transparent task denominators, independent assessment and follow-up beyond launch enthusiasm. Reports should describe the systems connected, permissions enabled, clinician experience and excluded workflows. Results should show exceptions alongside successes, including appropriate refusals rather than counting every refusal as failure.
Practices should also test whether people can challenge a plausible but unsuitable proposal and resume an interrupted workflow. These exercises belong in approved test or simulated environments. They should not introduce deliberately misleading information into live patient care.
For UK and EU readers, the immediate distinction remains availability: the September 2026 announcement does not make these new capabilities locally deployable. For eligible organisations elsewhere, the priority is a bounded workflow with a recognisable endpoint. For everyone assessing the launch, the standard should be evidence proportionate to what the product is being asked to do, not enthusiasm or suspicion alone.
Frequently asked questions
Is Heidi II the same as Heidi 2?
Heidi II is the official branding in the company's September 2026 launch materials; "Heidi 2" is an informal way readers may search for it. It describes the newly announced agentic platform rather than a separately verified clinical outcome.
Is Heidi II available in the UK?
Heidi's launch materials checked on 29 September 2026 say the new capabilities are not available in the UK or EU. That restriction should not be extended to every existing Heidi product.
Has Heidi II been shown to prevent clinical errors?
The reviewed launch materials describe capabilities and safeguards, but they do not establish that Heidi II prevents errors across clinical settings. That claim would require appropriately designed outcome evidence for the specified use.