An empty room does not necessarily mean a clinic can offer another appointment. The same interval may require available equipment, an appropriate staff role and preparation time between visits. If one resource is missing, an attractive calendar is still only a proposal.
AI can be used to draft administrative scheduling options and explain constraints. Calculations need independent checking, and a person must approve changes. The following is an original Nadali training workflow using entirely fictional inputs. It does not cover clinical triage, treatment decisions, employee scoring or legal advice.
1. Separate availability from individual appointments
Start with a narrow question: how many blocks of a defined type fit within the stated resource constraints? Do not give the model a patient list or ask it to decide who should receive care first.
In HL7 FHIR R5, Schedule describes availability without containing actual appointment information. Booking that needs several resources may require checking several schedules. This is a useful conceptual distinction, not a claim about any particular software product. HL7: Schedule.
A training file needs only fictional resource codes, available intervals, blocked intervals, block types, durations and required quantities. Use labels such as “role A”, rather than employee names. Exclude diagnoses, contact details, reception notes and attendance histories.
Aggregation alone is not permission to share real information. For this exercise, invent the numbers from scratch, without a route back to anyone’s actual appointment. Reviewing real data and authorisation to use it is a separate task outside this template.
2. Define the planning unit
Equal numbers of appointments can represent different resource requirements. Before counting, define each block type and the requirements approved by the responsible professionals. The model must not shorten appointment or preparation times to improve its result.
Do not equate completed appointments with total demand. NHS demand-and-capacity guidance explains that observed activity reflects what a service can deliver and is not the same as demand. NHS: aggregation module guidance.
For the exercise, set a synthetic request for eight identical blocks. Any blocks that cannot fit remain a visible capacity shortfall. AI must not hide them by changing units or inventing resources.
3. Check a small example by hand
Every condition below is fictional. These durations are not standards for any healthcare service:
- Planning window: an imaginary day, 09:00–12:00.
- Room R1: available throughout.
- Device D1: available except 10:00–10:20.
- Role A: one available unit, 09:00–11:30.
- Each block occupies all three resources for 20 minutes: an illustrative 15-minute appointment and five-minute preparation allowance.
- Blocks start at 09:00, 09:20, 09:40 and subsequent 20-minute grid points. Double booking is prohibited.
The room alone has nine positions. Only six satisfy all resource constraints: 09:00, 09:20, 09:40, 10:20, 10:40 and 11:00. The 10:00 block conflicts with device unavailability. Blocks starting at 11:20 and 11:40 finish after role A becomes unavailable.
Eight requested blocks therefore leave a shortfall of two. A second room alone would not solve it. A separate scenario with role A’s availability already authorised until 12:00 would allow eight blocks. That is a conditional calculation, not a recommendation to extend anyone’s shift.
4. Write down hard constraints
Hard constraints cannot be traded away for a better-looking total. Here they include each resource’s availability and quantity, the complete block duration, preparation time, blocked intervals and the prohibition on overlaps.
Preferences, such as keeping spare blocks together, come after feasibility. Do not introduce scores for individual workers’ “speed” or “reliability”. Role availability is an input, not a judgement about a person.
Check the entire interval, not just its starting time. A block ending at 11:40 cannot use a resource available only until 11:30. State the boundary convention too: this example includes the start but excludes the end of each interval, so adjacent blocks may share a boundary minute.
Record the date, time zone and input version. The exercise uses one time scale; ambiguous dates or clock changes in a real calendar require separate technical checking.
5. Give AI a bounded task
This is an editorial prompt template, not a verified feature of a named product:
Use only the supplied synthetic resources. Draft available blocks without creating bookings. Do not change durations, availability or blocked intervals. For every block, show its start, end, resources and checked constraints. Separately list conflicts, unplaced blocks and missing inputs. If the requirements are incompatible, report the shortfall. Make no clinical or employment decisions.
Ask for a short change log against the previous version. For example: “Two possible blocks added only because confirmed role A availability changed.” This helps expose an unnoticed change to another assumption.
Do not rely on the model’s own reassurance. NIST describes the risk of confidently presented incorrect outputs, including misleading explanations. Persuasive wording is not an interval check. NIST: Generative AI Profile.
6. Validate independently
Use an ordinary table or a tested script to recalculate durations and check overlaps for each resource. Check small examples manually. Another paragraph from the same model is not sufficient independent evidence.
A useful test deliberately proposes two simultaneous blocks using the same device. Validation should reject the conflict. Also test a missing duration, a duplicate resource identifier and a block extending beyond availability. An unknown value must not silently become “available”.
The output can have five sections: input version, feasible blocks, detected conflicts, unplaced remainder and questions for the responsible administrator. Keep rejected options with a short reason.
7. Approve before changing the calendar
The exercise ends with a checked draft, not a populated appointment book. In an operational process, an authorised person separately confirms resources, required constraints and permission to change the schedule. Recheck current availability immediately before saving: it may have changed since the draft was prepared.
Protection against simultaneous booking belongs in the booking system, not in a chatbot’s promise. If its checking or saving mechanism is unknown, do not transfer the draft automatically. Retain the previous version and a log of approved changes.
Evaluate the exercise by counting conflicts, missed constraints and review time alongside drafting time. Do not promise savings or additional appointments before measuring them. The practical result is a clear account of what fits, what is missing and which decision still belongs to a person.
Nadali. Prepared with AI assistance (ChatGPT). The example and template are editorial, not a trial in a real clinic. Sources checked on 2 November 2026.
The cover is a reused conceptual AI illustration of checking records, not medical documentation or a clinic photograph.
