A hospital chatbot is only as accurate as the last time someone updated what it reads. Patients ask about three things more than anything else: when a doctor sits, whether the clinic is open, and what a procedure or health check costs. All three change constantly, and all three are the facts a bot is most likely to get wrong without anyone noticing.
This article is for hospital marketing, IT and patient-experience teams who already run, or are about to launch, a chatbot on the website or WhatsApp. It answers one narrow question: how do you design the system so the bot reads timings, rosters and prices from the place where they are actually maintained, and so that any fact that cannot be confirmed stops being stated? It assumes you know what retrieval-augmented generation (RAG) is. If you do not, start with the AI chatbot implementation guide.
The failure it prevents is ordinary and expensive. A patient travels to a Saturday OPD that was discontinued in September, or is quoted a package price that finance revised last month. The bot did not hallucinate in the usual sense. It faithfully repeated a stale document.
Why chatbots go stale
Most hospital bots are built once from a crawl of the website, a few PDFs and an FAQ spreadsheet. That content is copied into a search index. From that moment there are two versions of every fact: the one in the hospital's systems and the one in the index. Only the first is maintained.
Stale answers usually come from four places:
- Marketing copies of operational facts. A department page says "OPD Monday to Saturday, 9 to 5". The real schedule lives in the HIS or the appointment system, and someone updates the page when they remember.
- PDFs that outlive their validity. A price list from last financial year is still in the index because nobody removed it, and it still matches a query well.
- Duplicates that disagree. The knowledge base holds the old and new version of a roster, and retrieval returns whichever scores higher for the wording of the question.
- Time-relative content. "Camp on the 14th" or "offer valid this month" is correct on publication day and wrong afterwards.
Faithfulness metrics do not catch any of this. A metric such as Ragas faithfulness scores whether a response is consistent with the retrieved context. A stale document produces a perfectly faithful, perfectly wrong answer. Freshness has to be designed into the data, not tested out of the output.
Step 1: Classify facts by how fast they change
Do not treat the knowledge base as one pile. Sort every fact the bot may state into volatility classes, because each class gets a different retrieval method.
| Class | Examples | Changes | How the bot should get it |
|---|---|---|---|
| Stable | Department descriptions, how to prepare for an MRI, visiting policy | Quarterly or less | Indexed documents with a review date |
| Slow | Doctor profiles, specialties, address, facility list | Monthly | Indexed from the CMS, rebuilt on publish |
| Fast | OPD timings, doctor availability, holiday closures | Weekly or daily | Live lookup from the roster, never from a copied document |
| Priced | Consultation fees, health-check packages, procedure estimates | Per revision | Live lookup from the price master, with an effective date |
| Event | Camps, offers, webinars | Per event | Indexed with a hard expiry date |
The rule that follows: fast and priced facts must not be stored in the vector index at all. The model should call a function that returns the current value, and the index should hold only the stable and slow classes. This is the single biggest change most teams can make.
Step 2: Name a source of truth for each fact
For every fast and priced fact, write down one owning system and one owning person. A simple register is enough:
| Fact | System of record | Owner | Update trigger |
|---|---|---|---|
| OPD timings per department | HIS or appointment scheduler | Front-office head | Roster change |
| Doctor-wise availability | Appointment scheduler | Department coordinator | Leave or slot change |
| Consultation fee | Price master | Finance | Annual or ad hoc revision |
| Package price and inclusions | Package master | Billing and clinical lead | Package revision |
| Holiday closures | Admin calendar | Administration | Notice issued |
If a fact has no system of record, which is common for packages managed in a spreadsheet, fix that first. A chatbot cannot be more current than the least disciplined owner in the register. Where the owner is a person with a spreadsheet, at least make the spreadsheet a controlled file with a version history and a required "effective from" column.
Your records and the website should draw from the same place. If the department page says one thing and the bot says another, the patient sees an inconsistent hospital, not a technical problem.
Step 3: Build the lookup layer
The bot should not search for "Dr Rao timings". It should call something like get_opd_schedule(department, doctor_id, date_range) and receive structured data. A workable response shape has these fields:
value: the timings, fee or package price as structured dataeffective_fromandeffective_to: when the value appliessource_systemandrecord_id: where it came fromretrieved_at: when the lookup ranstatus: one ofconfirmed,unconfirmedorunavailable
The system prompt then states the rule plainly: use only values returned by a tool for timings, availability and prices; if status is not confirmed, do not state a figure; say the front desk will confirm and offer a call-back or a booking link.
Two details matter. First, the model never does arithmetic on prices. If a package has add-ons, the lookup returns the computed total, or the bot states the base price and says extras depend on the clinician's assessment. Second, estimates for procedures should be stated as the hospital's published ranges, not as a quotation. NMC's October 2026 ethical advertising guidelines say any disclosure of charges, packages or fees must be factual, transparent and not misleading. A bot that quotes an outdated or incomplete price is a compliance risk as well as a service failure. Have your compliance reviewer read the full guidelines and decide how they apply to your communications.




