Your Receptionist Is Not the Problem. The Job Is.

Meet Michèle

Michèle works the front desk. Except she also works the inbox, the diary, the recall list, the payment conversation, the Instagram messages, the missed-call log, the nervous patient in chair two, the surgeon running forty minutes late, and the printer, which has picked this morning to stop cooperating.

She is not struggling with any of it, which is rather the point. Michèle is very good at her job — good in the way that becomes invisible, because nothing she prevents ever appears anywhere. She knows which patient needs the longer phone call and which one needs the shorter one. She knows the referrer who likes a fax, in the manner of a Victorian, and the one who answers only on WhatsApp. She holds the practice’s tone, its history and roughly four hundred people’s emotional states in her head, and she does it while being interrupted constantly — the kind of constant that only people who have sat in that chair fully understand.

The practice depends on her far more than any org chart admits. And here is the sentence this essay exists to establish, before anything else is said: the problem is not Michèle. She is a composite, incidentally — assembled from the coordinators, receptionists and practice managers of many practices, not a portrait of any one of them — but every principal reading this has met her, and most are lucky enough to employ her.

The problem is what her job has quietly become.

The job description nobody wrote

There was a job description once, presumably. It was written when the front door was a telephone. The role meant the phone, the diary, arrivals, letters and the small daily diplomacy of a waiting room — a demanding job, honestly done, with edges.

Nobody ever rewrote it. Instead, the edges dissolved, one reasonable addition at a time. The website went up, and its form became her inbox. The practice joined Instagram, and the direct messages became hers too. WhatsApp arrived because patients preferred it, then SMS reminders, then the missed-call log, then online reviews, each one somebody’s excellent idea and none of them anybody’s staffing decision. Fee questions grew payment-plan sub-questions. Recalls became reactivation. Quotes acquired follow-up schedules, which acquired the special status of work that exists only if somebody remembers it. And threaded through everything, the original job carried on in the foreground: the phone, the diary, the arrivals, the nervous patient in chair two.

No single addition was unreasonable. That is what makes the accumulation so hard to see. The role description was written for a telephone; the job now arrives through a dozen channels, each carrying a stopwatch, and the person carrying it is still described — on the payroll, in the org chart, at the Christmas speech — as the receptionist.

Why good people look inconsistent inside bad systems

Here is the part principals most need to hear, because it runs against a very human instinct. When follow-ups slip, when a message sits unanswered, when the same detail gets asked for twice, the natural reading is a performance problem. Somebody dropped it. And occasionally somebody did. But when capable, conscientious people produce inconsistent results repeatedly, across different individuals in the same seat, the honest place to look first is the design of the work, not the character of the workers.

The research here is not fashionable opinion; it is decades deep, and much of it comes from healthcare. Interruption is not a neutral event. In hospital medication administration studies, interruptions were associated with increased frequency and severity of errors.¹ Observational studies of interruption in clinical work associate task-switching with elevated mental workload and increased likelihood of error, and the literature’s practical recommendation is consistent: redesign the workflow rather than simply urging staff to try harder.¹ The organisational psychologists add a quieter cost called attention residue: every time a person switches tasks, part of her attention stays behind with the unfinished one, degrading the next task even after the switch.² Now count Michèle’s open loops. The quote awaiting a decision. The recall list. The patient who said call me Thursday. The message half-answered when the phone rang. Each one is a small claim on a finite working memory, and the job description asks her to hold all of them, indefinitely, under constant interruption.

This is why visible busyness and completed patient pathways are different measurements. A person can be flat-out all day — genuinely, virtuously flat-out — while the loops that matter stay open, because open loops are precisely the work that interruption defers. None of this diagnoses any individual, and not every missed follow-up is systemic. But a repeating pattern across capable staff is a reason to examine the design before defaulting to discipline — and a workflow that depends on memory is a workflow waiting to fail.

Why another hire does not redesign the job

The instinctive fix is another person, and sometimes another person is exactly right — where the workload genuinely exceeds human capacity, hiring is the honest answer and nothing in this essay argues otherwise.

The narrower point is what a hire does not do. Another person extends capacity; she does not unify the practice’s channels into one view. She does not create an approved knowledge base, so the practice’s answers now vary by who happens to reply. She does not remove duplicate data entry; the same details are typed again at another desk. She does not make follow-up less dependent on memory; the practice now depends on two memories and a handover between them, which is a new place for loops to open. She does not, on a standard roster, cover the evening and weekend hours when a meaningful share of enquiries arrives. And she brings her own entirely legitimate needs — training, management, leave, cover — which are the ordinary costs of employing humans and should be spent on work that needs a human.

Headcount can solve a capacity shortage. It cannot, by itself, redesign a badly designed flow of work. Without the redesign, the practice simply distributes the same fragmentation across more people, and then wonders why the new hire looks inconsistent too.

The false choice about AI

Which brings us to the argument currently being conducted at conference volume: replace the receptionist with AI, or forbid the technology and preserve the chaos. Both positions are wrong, and both are wrong in the same way — they treat the question as being about people, when it is about tasks.

The replacement rhetoric fails on contact with the actual job. None of the judgement described in the opening — which patient needs the longer call, the reading of a referrer, the reassurance calibrated to one anxious person — should be delegated to automation alone. Technology may support the repetition surrounding that judgement. It cannot own the judgement itself, and a vendor who says otherwise is selling something other than what Michèle does. The prohibition position fails differently. A 2025 UK survey commissioned by Microsoft and conducted by Censuswide found a large majority of employees reporting use of unapproved consumer AI tools at work, many of them weekly — including in health and social care settings.³ Unapproved AI may already be in the building. A practice that declines to govern AI cannot safely assume that nobody inside it is using AI; it merely ensures the version in use is the ungoverned one — often through personal accounts and outside any practice policy.

The useful question is neither which person can this software remove nor how do we keep it out. It is: which tasks should no longer depend on a person remembering them under interruption? And whatever carries those tasks must be bounded and honest about itself — answering only from practice-approved content, escalating to a human at anything sensitive, clinical, ambiguous or out of scope, and with patients told, clearly and proportionately, when AI is assisting the conversation. Trust comes from clarity, boundedness and useful escalation, never from impersonation.⁴

The Human Work Test

So run the test — slowly, with the team in the room. Take any task that currently crosses the front desk and ask ten questions: Does it require clinical judgement? Does it require empathy in an unusual or sensitive circumstance? Is it repetitive and rules-bound? Does it need an immediate acknowledgement? Does it depend on memory or a timed follow-up? Can it be answered safely from approved content? Does uncertainty require escalation? Does a human need context before taking over? Who remains accountable? And should the patient be told AI is assisting?

The answers sort the front desk’s actual workload into four honest categories.

Category	What belongs here
Infrastructure can carry	Acknowledging an enquiry through a configured channel; retrieving practice-approved general information; collecting structured, non-clinical context; prompting a consented reminder; preparing a handoff summary; surfacing overdue follow-up; keeping the audit trail
Infrastructure can assist, a human approves	Fee-range and finance-option information; reactivation wording; unusual scheduling requests; any drafted response involving sensitivity or uncertainty
A human must decide	Clinical questions and suitability; patient-specific risk; treatment recommendations; consent; complaints; distress; financial hardship discussions; privacy and records requests; exceptions; safeguarding; consequential financial or contractual conversations
Stop and escalate	Urgent clinical concern; self-harm or acute distress; ambiguity outside approved content; privacy, identity or suspected fraud concerns; images or records beyond the configured scope; anything where the system cannot safely find the boundary

Two things about this table. First, it is useful whether or not a practice ever buys anything — it is a job-design instrument, and a principal could spend a team meeting on it tomorrow with a whiteboard. Second, notice how much of Michèle’s day sits in the first row: the repetitive, rules-bound, immediate, memory-dependent work — the work most exposed to the interruption and task-switching costs described above — currently occupying one of the most emotionally skilled people in the practice. That allocation is not a personnel insight anybody chose. It is an accident of accretion, and it is reversible.

The morning briefing

Redesigned, the day starts differently — not utopian, just coherent.

Michèle arrives, and the overnight enquiries that came through configured channels have been acknowledged or safely escalated, with the relevant context waiting for her rather than a queue of cold messages. She can see who was anxious about what, and what each person preferred as a next step. The follow-ups due today have surfaced themselves; nothing depends on her remembering Thursday. The exceptions — the message the system could not safely interpret, the conversation that touched a boundary — sit clearly separated at the top, flagged as needing exactly what she uniquely supplies. She opens a briefing, not a backlog.

Then she does the human work, which was always the actual job: the judgement calls, the reassurance, the exception nobody could script, the difficult call she is better at than anyone in the building. And when the system gets something wrong — it will, being infrastructure rather than magic — she corrects it, which is precisely why the oversight is operational rather than ceremonial. The design assumes fallibility on both sides and puts a human in charge of the joins.

The owner’s obligation

None of this can be delegated to a vendor, and any vendor who implies otherwise is misdescribing where accountability sits. Introducing infrastructure into a front desk is a leadership decision, and the decisions that matter are all the owner’s: what the approved content actually says, where the boundaries sit, what escalates and to whom, who is accountable for each category of the table above, how staff are consulted and trained rather than surprised, how privacy arrangements are set, how patients are told AI is assisting, which tasks are permanently reserved for humans, how errors get corrected, and how the team’s concerns get heard after go-live rather than only before it.

Technology does not relieve the owner of judgement. It makes the owner’s judgement visible in the system — or exposes its absence. A practice that installs infrastructure without doing this thinking has not modernised its front desk. It has laminated its confusion.

CareConcierge

This is the redesign CareConcierge exists to serve. It is AI-assisted patient communication for high-trust private healthcare, and in this article’s terms it does a deliberately narrow set of things. Depending upon the agreed configuration, CareConcierge can carry selected tasks in the first row of the table — acknowledgement through configured channels, answers drawn from the practice’s approved knowledge base, structured non-clinical context, authorised reminders within the practice’s approved workflow, surfaced follow-ups, an audit trail — support appropriate tasks in the second row under human oversight, and route the remaining work to the practice with context attached, escalating at sensitive, clinical or out-of-scope boundaries. AI assistance is transparent to patients. The practice retains responsibility for clinical and regulatory approval, and the service runs on reasonable commercial efforts across configured channels — infrastructure honestly described, with the limitations infrastructure has, not a promise of omniscience.⁵

The measure of it is not a dashboard. It is the shape of Michèle’s morning. It does not make her unnecessary. It makes her judgement less interruptible. Where it cannot do something safely, it is designed to say so and hand the moment to a person.

Michèle, again

The front desk was never meant to become a call centre, a treatment encyclopaedia, a CRM, an after-hours service and the practice’s commercial memory in one chair. Michèle was not failing the job. The job had stopped being one job — and the kindest, most commercially serious thing a principal can do is separate it again.

Give the infrastructure the repetitions. Give the human the exceptions, the judgement and the relationship. She decides; the system remembers. And the patient walking in this morning meets a person who has time to be present — which was, all along, the thing no channel could carry.

Gregory William Gray is a co-founder of CareConcierge and the author of The Exit Discipline (Summit Communication Group, 2026). CareConcierge provides AI-assisted patient communication for private healthcare practices.

Source notes

1. On interruption and error in healthcare settings: Westbrook J. et al., “Association of interruptions with an increased risk and severity of medication administration errors”, Archives of Internal Medicine (2010) — error frequency and severity rising with interruptions in Australian hospital observation; Shan Y. et al., “Workflow interruption and nurses’ mental workload in electronic health record tasks: an observational study”, BMC Nursing (2023) — interruption-driven task switching associated with elevated mental workload and negative outcomes; the study’s practical recommendation, paraphrased here, is workflow redesign rather than urging staff to try harder. Clinical-setting evidence, applied here by analogy to front-desk work: the mechanism described — interruption raising cognitive load and error likelihood — is general, and no front-desk-specific error rate is claimed. ↩ ↩
2. Leroy S., “Why is it so hard to do my work? The challenge of attention residue when switching between work tasks”, Organizational Behavior and Human Decision Processes (2009). ↩
3. Research commissioned by Microsoft and conducted by Censuswide (October 2025; n=2,003 UK employees, including health and social care respondents): a large majority of UK employees reported having used unapproved consumer AI tools at work, many at least weekly. Cited directionally as evidence of ungoverned AI use in workplaces — a single commissioned survey, not a healthcare-specific rate. ↩
4. On transparency: the Australian Government’s Guidance for AI Adoption, the UK Information Commissioner’s Office guidance on transparency in AI and personal-data processing, and the OAIC’s guidance on privacy and the use of commercially available AI products, as cited in the previous essay in this series. General good-practice guidance, not bespoke legal advice. ↩
5. Product descriptions in this section reflect CareConcierge’s standard service terms: the configured, practice-approved Knowledge Base from which responses draw; guardrails preventing clinical claims, diagnostic statements, outcome guarantees or treatment recommendations; escalation logic routing enquiries requiring clinical judgement to the practice for human review by appropriately qualified staff; service availability on reasonable commercial efforts, subject to maintenance and third-party platform limitations; and the practice’s retained responsibility for clinical and regulatory approval. ↩

Michèle is a composite drawn from front-desk professionals across many practices, not an identifiable employee. CareConcierge does not diagnose, recommend treatment for an individual patient, guarantee outcomes or replace clinical judgement; it does not remove human accountability; patients should be told clearly and proportionately when AI assists an interaction; sensitive, clinical, ambiguous or out-of-scope matters escalate to a human; and configured channels, integrations and third-party services may have limitations or outages.