Some patients may welcome an agent that completes an agreed administrative task conveniently; others may find the same interaction impersonal or difficult to change. The studies considered here do not establish how patients will respond to Heidi Work. They suggest questions about usefulness, individual circumstances and access to a person that a real evaluation should answer.
This September 2026 analysis assumes substantial authorised delegation, not an agent making independent medical decisions. Heidi's 22 September announcement describes supervised action and excludes the coming capabilities from the UK and EU. The patient interactions below are fictional design scenarios, not documented Heidi Work functions or reported patient experiences.
Taking notes and contacting a patient are different relationships
A background documentation tool and an agent acting on behalf of the practice create different patient-facing situations. The latter might initiate contact, request a response or change the next step in an administrative process. Its behaviour would become part of the patient's experience of the service.
The purpose of that contact matters. Arranging an appointment already agreed with the clinician is different from reminding a patient about an established action. Both are different from communicating new clinical information that the patient has not yet discussed with a professional.
A useful acceptance study should keep those situations separate. A patient who welcomes an appointment reminder has not thereby accepted every possible form of automated clinical communication. Equally, discomfort with receiving significant new information from an agent does not prove rejection of all administrative assistance.
The question should therefore be specific: will patients accept this action, delivered in this way, with these options for changing course? Asking whether they support AI in healthcare in general would leave too much unspecified.
Convenience can support personal care rather than replace it
Consider a fictional patient who has agreed to a routine follow-up but cannot attend the first proposed appointment because of caring responsibilities. A helpful delegated workflow might offer suitable alternatives, recognise the patient's response and confirm a new arrangement without requiring another lengthy exchange.
The potential benefit is not that the patient has spoken to an AI. It is that the practice has understood the practical problem and completed an agreed action. A quick administrative resolution could make the wider care relationship feel more responsive.
Now change the scenario. The patient replies that the suggested appointments are unsuitable and receives the same options again. They ask for a telephone conversation but are sent back to a form. The wording remains warm, yet the interaction provides no practical way to influence what happens next.
That would be a different experience even if the agent sent every message promptly. The design question is whether the patient can redirect the process, not merely whether the system can produce reassuring language.
Neither scenario requires a false choice between immediate automated contact and an indefinitely unavailable person. An evaluation should compare realistic service arrangements, including a human pathway that actually exists, rather than an ideal agent against an artificially poor alternative.
What a small appointment-management study contributes
A qualitative study published on 17 June 2025 examined acceptance of AI-enabled outpatient appointment management through interviews conducted in 2023 with 12 patients and seven staff at an English NHS trust. Participants described benefits including convenience and reminders, alongside limitations involving interactivity, integration and usability.
The study concerns a particular appointment-management setting, not Heidi Work. Its small qualitative sample cannot establish population-wide preferences or predict uptake of a new agent. Its value here is in identifying questions that an evaluation should investigate: can people respond meaningfully, do changes reach the relevant service, and does the process fit how patients actually manage appointments?
Those questions are more useful than treating a successful message delivery as proof of acceptance. A delivered message may still be misunderstood, impossible to act on or disconnected from the booking process it is supposed to support.
Individual circumstances must be able to change the interaction
A second fictional patient receives a request to arrange an action they have already completed elsewhere. Their response introduces information that may not yet be visible in the practice's record.
An agent should not necessarily accept the response as sufficient evidence to close the clinical matter. But a proposed administrative design should provide a route to record the information, explain what happens next and obtain human review when required. Simply repeating the original request would not address the patient's concern.
Other variations might involve an inaccessible communication format, a shared telephone, uncertainty about the identity of the sender or a request for assistance from an authorised representative. These are distinct situations to test, not interchangeable examples of a patient failing to engage.
The relevant behavioural concept is sometimes called uniqueness neglect. In research published online in May 2019, Longoni, Bonezzi and Morewedge found that concern about medical AI overlooking individual circumstances helped explain resistance in their experimental settings. Those settings were not Heidi Work deployments. The concept should frame an inquiry into whether patients feel understood, not serve as a prediction that they will reject autonomous agents.
A practical test would ask whether a patient's relevant new information can alter the process appropriately. Personalised wording alone does not establish that it can.
Act for the practice without pretending to be the doctor
An appropriate introduction in a proposed interaction might explain that the practice's automated assistant is arranging the follow-up agreed with the clinician. That wording identifies the system, links its action to an existing plan and avoids implying that the doctor personally composed the message at that moment.
The introduction should also match the actual scope of the conversation. An assistant authorised to organise an appointment should not appear to offer an unrestricted clinical consultation. When a response introduces something outside its permitted task, the patient needs an intelligible account of the next step.
A human route should be more than an instruction to contact the practice without further help. An evaluation should establish where the request goes, whether the receiving team has the necessary context and whether the patient must repeat everything. The proposed design goal is continuity across the handover, not the appearance of endless availability.
These are patient-experience principles for evaluating a workflow. They are not claims that a particular script satisfies every legal, information-governance or clinical requirement in every setting.
Measure acceptance through behaviour and experience together
A useful pilot would assess whether the intended task was completed, whether the patient understood what happened and whether additional contact was needed. It should also record requests for human help, unresolved interactions, corrections and the reasons people chose not to continue.
Requests for a person are not necessarily failures. They may show that the escalation route is visible and being used appropriately. Conversely, a low request rate might reflect satisfaction, but it might also mean patients could not find the option. Interviews and observation are needed to distinguish those explanations.
The evaluation should include people who struggle with the process, not only those who complete it. Otherwise the most convenient users to survey could become the entire evidence base. It should also separate administrative completion from patient confidence: both matter, and neither is a substitute for the other.
For straightforward agreed arrangements, an agent may offer genuine convenience if patients can influence the outcome. For situations involving uncertainty, sensitive communication or a changed concern, a well-supported human interaction may be preferable. The important verdict is by scenario, not a declaration that automated or human contact always wins.
Frequently asked questions
Will patients judge Heidi Work mainly by whether it is autonomous?
That remains a hypothesis rather than an established product-specific finding. The purpose of the contact, its responsiveness to individual circumstances and the ability to obtain human help may also influence acceptance.
Does a reply to an automated message demonstrate patient acceptance?
No, a reply may reflect confusion, a request to stop or an attempt to reach a person. Acceptance should be assessed alongside task completion, understanding, repeated contact and the patient's account of the experience.
Should an agent identify itself when acting for a practice?
Clear identification is an appropriate design principle for these proposed workflows, together with an explanation of the delegated task and available human support. It should not imply that an automated message was personally written by the patient's doctor.
Explore iatroX perspectives on AI and patient-centred practice →