An AI agent and a clinic operating system are not directly comparable units of work. To assess Heidi and Tandem's September 2026 ambitions, define the activity the software is meant to complete, the approval it needs and the evidence that the work actually happened. A label cannot distinguish a helpful draft from a completed action with an audit trail.
This comparison is published by iatroX and includes iatroX as a clinical-reference and learning tool, not an operational agent. The task framework below is illustrative; it does not claim that both suppliers already deliver every stage.
What the announcements establish
Heidi's September 2026 announcement describes supervised activity around clinical care and specifies that the forthcoming capabilities in that announcement will not be available in the UK and EU. That restriction should not be applied to all existing Heidi products. Heidi's release provides the product-direction and regional context.
Tandem's 14 September 2026 announcement describes expansion towards an AI-native clinic operating system, including broader clinic activity such as patient flow, scheduling, triage and communications. It is a direction of travel, not evidence that every proposed operational task is live in every deployment. Tandem's roadmap is the starting point.
The common theme is a move beyond generating documentation. The meaningful difference will emerge in the specific work delivered, its integration and the way exceptions are handled.
Define one complete unit of work
Consider a fictional review appointment after which a clinician agrees that the patient needs a follow-up communication. The proposed sequence is to recognise that requirement, prepare the content, obtain approval, send it through the appropriate route and record the result.
Each stage needs a clear input and output. Recognition depends on what was agreed. Drafting depends on the relevant information. Approval belongs to an authorised reviewer. Sending requires the correct recipient and channel. Recording requires evidence about what happened, including whether communication failed.
This is a comparison framework rather than a description of an available Heidi or Tandem workflow. Its value is that a buyer can ask both suppliers the same concrete questions without assuming that similar marketing words describe the same behaviour.
Assistance and execution need different labels
A system that suggests a follow-up has not drafted it. A system that drafts a message has not obtained approval. A message placed in a queue has not necessarily been sent. A send action has not necessarily been delivered or acknowledged.
A useful evaluation should preserve these distinctions. Rather than a binary "follow-up supported" tick, the specification can state the furthest stage the product performs and the stage that remains manual. It should also identify where the user can stop or change the process.
This matters even when the final work is administrative. An apparently small ambiguity about whether a task was queued or completed can change what the next person does. A product's interface should make unresolved work visible rather than allow the reader to infer completion from a generated paragraph.
Review the exceptions before the happy path
A fictional evaluation can deliberately include missing information, an ambiguous instruction, an unavailable destination and a rejected proposal. It can also include a task that was correctly drafted but should not proceed after new information arrives.
Ask whether the system requests clarification, stops safely, routes the issue to an identified person or leaves an unresolved state that someone must notice. Ask what happens when the reviewer changes the content after another part of the workflow has already been prepared.
These are proposed test conditions, not allegations of specific supplier failures. They are useful because the polished demonstration usually shows only a straightforward encounter. Operational confidence requires understanding what happens when the original plan is incomplete or changes.
Measure completed work, not generated volume
A proposed scorecard should define the denominator as eligible tasks attempted, not merely outputs generated. It should then distinguish tasks completed correctly, tasks corrected before completion, tasks abandoned appropriately and tasks left unresolved. That makes restraint visible as well as success.
Review time should be measured separately from generation time. Interruptions, repeated data entry and work transferred to colleagues belong in the assessment. A faster draft can still be a net improvement, but the evidence should show the whole process rather than select the most flattering interval.
If a study reports user satisfaction, keep it as satisfaction. If it reports consultations processed, keep that unit. Neither should be reclassified as a completed-action rate. The metrics must match the claim being made.
No such comparison was run for this article. Results should be published from an actual authorised evaluation with the product versions, workflows, reviewers and limitations documented.
Integration must include the destination
As checked on 22 September 2026, Tandem publicly describes clinical decision support and an integration catalogue. Those product descriptions can inform the questions asked, but they do not prove that a particular deployment completes the illustrative sequence above. Tandem Clinical Decision Support and its integration catalogue provide the current published context.
The destination matters because a generated item can look complete while remaining outside the system where colleagues expect to find it. A buyer should ask how the product confirms that the approved version was saved, how later amendments are handled and what appears when a transfer is incomplete.
The same standard applies to Heidi's proposed supervised work. A broad agent description becomes useful procurement information only when it identifies the action boundary and shows what is actually delivered in the relevant market.
Appropriateness remains separate from execution
A workflow can carry out the wrong instruction efficiently. Clinical reasoning therefore remains a different assessment from technical completion. The reviewer needs to understand whether the action makes sense for the situation and whether the cited information supports it.
Per iatroX product information, September 2026, free Ask-iatroX provides linked clinical sources, and Socratic Tutor uses targeted follow-ups from an attempted question to explore a learner's reasoning. These functions can support reference and practice; they do not execute the workflow described above or prove competence to supervise it.
That separation helps organisations avoid a false choice between operational efficiency and professional learning. They can evaluate whether a system completes the approved work while separately supporting the knowledge needed to approve it appropriately.
Verdict by operational scenario
For a team needing better drafts, evaluate drafting and review without demanding an entire operating system. For a team seeking task completion, require evidence from approval through destination and recovery. For a buyer considering a roadmap, distinguish future scope from current entitlement. For an educator, practise the decisions at the review boundary rather than teaching product labels.
Heidi and Tandem should ultimately be compared through the work they deliver in those scenarios. An expansive name is not a substitute for a clearly defined and dependable task.
Frequently asked questions
Is an AI agent automatically more capable than a drafting tool?
Not for every task: the term alone does not establish what the product can do. Ask which actions are available, how they are approved and how completion is confirmed.
Does a clinic operating system mean all clinic tasks are automated?
No: Tandem's September 2026 description is an expansion ambition. Each workflow and local entitlement still needs separate confirmation.
What is the most useful comparison measure?
Use a defined task and measure correct completion alongside review, corrections and unresolved exceptions. Generated text volume cannot answer that question on its own.