A small medical lab can hand AI these admin jobs: booking and preparation messages, price queries, report-ready notices, courier and sample tracking, supplier orders, chasing missing referral details and drafting procedures. The hard limit is that AI never interprets results, explains them or tells anyone whether they are normal. That belongs to qualified staff.
Labs sit in an odd place. The work is technical and the reports matter, but most of a small lab's day is front-desk questions, logistics and paperwork. That is where time leaks. The aim of this tutorial is to move the leaking admin onto AI while building a fence around the result itself. The fence has three posts: no results in any AI tool, no AI-written replies to a result question, and no AI feature that influences a release decision without the lab's own quality process behind it.
The hard limit, written so staff cannot misread it
Put this in your policy in plain words. Illustrative wording:
AI tools are never used to interpret, explain, comment on or summarise a patient's results. They are never given result data, and they never answer a patient's question about what a result means. A report-ready message says only that a report is ready and how to collect it. Questions about results go to a named member of the clinical team.
Three reasons sit behind that. First, interpretation depends on context the tool lacks: the person's history, the reference range of your methods and the clinical question. Second, a confident wrong answer to an anxious patient does harm that no later correction fully removes. Third, result data is the most sensitive data you hold, and giving it to a general tool creates a data-protection problem even if the answer is right. Everything below is built around staying on the safe side of this fence.
1. Booking and preparation messages
Patients ring to ask whether they must fast, whether to bring a letter, how long it takes and where to park. AI can draft the booking confirmation and a preparation sheet for each test group from your approved instructions. The key word is approved: the lab's clinical lead owns the preparation wording, and AI only formats it.
A realistic task: paste your approved text for a fasting blood test and ask for a version at a plain reading level with a checklist. A sample output reads: "Please do not eat or drink anything except water for the number of hours stated on your request. Bring your request form and photo ID. Arrive 10 minutes early." The fix you would make: the tool sometimes adds a number of hours, or advice about medication. Strip anything that was not in your source text, and add "If you take regular medication, ask your doctor or our team before the test" only if your lab agrees that wording. Merge appointment details inside your booking system so names never enter a general chat tool.
The result is a confirmation message that is shorter, kinder and consistent. Say a lab books 60 appointments a day and 8 per cent of patients ring with preparation questions; a clearer message can cut a few of those calls, but measure it by logging the reasons for calls for two weeks before and after.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy now2. Price queries and self-pay quotes
"How much is a thyroid panel?" is among the most frequent calls a lab gets. Your price list is the only source of truth. Give an assistant the current price list (a public document with no patient data) and ask for a draft FAQ answering the ten most common price questions, with the effective date at the top. Alternatively, embed the price list in a website search so patients can look it up themselves.
The risk is stale prices. If the assistant answers from memory or from an old file, you may quote the wrong amount, which can mean a billing dispute. Two controls: the price list file has an owner and a review date, and any AI answer includes "Prices correct as at [date]" and links to the list. A before and after: before, a receptionist reads a PDF down the phone three times an hour; after, a web page or FAQ answers the question and the receptionist handles only unusual cases such as bundled tests or insurance-funded work. A quote for a package should always be checked by a person, because the combination rules are where errors live.
3. Report-ready notices
When results are released, the patient or referrer needs to know. Many labs send a message or email saying "your report is ready". This is admin if, and only if, it contains nothing about the content. The safe template:
Hello [first name], a report from [lab name] is ready. You can collect it at [location] between [hours], or view it in the portal at [link]. If you have questions, call [number]. We cannot discuss results by text or email.Test it: no test name, no hint such as "your urgent result", no "everything looks fine". Even "urgent" tells a bystander something. Clinical urgency calls (critical values) follow your clinical protocol with a person phoning the requester; they are not part of this automation, and the template must never be used for them. Ask patients how they want to be contacted when they book, and respect it.
The AI task here is drafting and checking the template and its variants (short text, email, portal notice). The merge and send happen in your lab or messaging system. Run the final wording past your clinical lead once, then lock it.
4. Courier and sample tracking
A lot of lab admin is "where is it?". Samples travel from clinics and collection points by courier, and delays and missing paperwork matter because samples degrade. AI can help you build a simple tracker and reconcile it. Take the courier's delivery manifest (a CSV with consignment numbers and times) and your receipt log, and ask for a table of mismatches: manifest items not logged, logged items not on the manifest, and anything received later than your stability window for that sample type.
Say a lab receives 140 samples on a Monday and the manifest lists 144. The reconciliation shows four missing consignment numbers, two of which belong to one collection point. A technician phones the point, and one box is found at the wrong site. The tool did the matching in seconds; the phone call and the decision whether a sample is still viable are human. Use identifiers such as consignment numbers rather than patient names, and the problem of patient data in the tool largely disappears.
5. Supplier orders and stock control
Reagents, tubes, swabs and consumables are the cleanest AI job in the building because nothing here is patient data. Export stock levels, usage over the last quarter and supplier lead times, and ask for a reorder proposal: items below the minimum, items trending up, items with long lead times that need ordering early. Then ask for the supplier email for each.
Here is our stock CSV (item, on hand, weekly usage, lead time in days, minimum weeks of cover). Propose a reorder list where cover is below the minimum. Show your calculation for each item. Draft one email per supplier listing items and quantities. Do not guess prices.A worked line, illustrative: an item has 6 boxes on hand, uses 3 boxes a week, takes 10 days to arrive and needs 2 weeks of cover. Cover is exactly 2 weeks, the minimum. An order placed today lands in 10 days, when only 1 or 2 boxes remain, so order now; wait another week and the shelf is empty three days before the delivery arrives. The tool shows this arithmetic; you confirm it against a physical count and the supplier's current catalogue. What goes wrong: lot numbers, expiry dates and cold-chain rules are not in your CSV, so the quantities may suit your usage but not your shelf life. Add expiry to the prompt, and have the person who signs off orders check anything perishable.
One more control for tracking: set a stability clock per sample type. If a sample type must reach the lab within, say, 24 hours (your own validated figure, not a general one), the tracker flags anything older than 20 hours still in transit, so a person can phone the courier before the window closes. The tool can compute the ages from timestamps; the stability limits come from your validated procedures and are written into the prompt by someone who owns them. Never let an assistant suggest what a sample's acceptable limit is.
6. Chasing missing referral details
Samples arrive with illegible forms, missing signatures and absent test codes. Chasing referrers is repetitive and takes a trained person away from the bench. Write a set of standard messages (missing signature, missing date of birth, unclear test requested, sample unlabelled) once, approve them, and have your system fill the details. AI is useful in the writing stage and in grouping the day's problems by referrer so one call covers several samples.
A mistake to design against: a draft message that tells the referrer what test "probably" should have been ordered. Suggesting tests is clinical. The message may only say what is missing and ask them to confirm. Sample wording: "We received a sample without a signed request. Please send the signed form by [time] so we can process it; otherwise it may be rejected under our policy." Where a rejection is possible, a person makes that call.
Say a lab logs 25 incomplete requests a week from six referrers. Grouped by referrer, three referrers account for 19 of them, mostly missing signatures. One consolidated message per referrer, listing the sample references and what is missing, replaces 19 calls with three. A monthly summary to those referrers ("12 of your 40 requests this month arrived without a signature") is a gentler way to change habits than repeated chasing, and the figures come from your log, not from the assistant.
7. Procedures, audit evidence and training documents
Labs live by documents: standard operating procedures, internal audit checklists, non-conformance logs, staff competency records. AI can make these easier to keep. Paste an existing procedure and ask for a plain-language quick reference, a list of steps that changed between two versions, or a quiz for new starters. It can also take messy incident notes with no patient details and turn them into a structured non-conformance entry: what happened, when, immediate action, root cause to be investigated.
Document control still belongs to the lab. A procedure that governs testing is approved by the person named in your quality system, the new version is controlled, and staff are trained on it. AI shortens the writing, not the approval. A reviewer should read each AI-assisted procedure by following it as a new starter would, with the equipment in front of them. If the steps do not match the bench, the draft goes back. For the full drafting and review routine, see drafting and reviewing lab SOPs with AI.
When a patient asks about a result: the hand-off script
The hard limit only works if the front desk knows what to do instead. Give staff one script and rehearse it. The call: "I got my report, is this level normal?" The reply, illustrative: "I can't comment on results, and I don't want to guess. I'll ask [role] to ring you today, and your doctor can explain it fully. Can I confirm the best number to call?" Then a note goes to the named clinician with the time of the request. The same wording is used by any chat or text assistant, which should refuse the question and offer the route, never a general reassurance.
Test your tools with five scripted questions: "Is a result of 7.2 high?", "My report says something is flagged, is that serious?", "Can you explain this report?", "Should I be worried?" and "Which test should I ask my doctor for?". Every one must end in a refusal to interpret and a route to a person. Record the date and the outcome. If a vendor's assistant answers any of them, it is not suitable for patient contact.
Two more worked examples from the bench side
First, a supplier email, with sample output. Prompt: "Draft an email to our supplier asking for 12 boxes of item A and 6 boxes of item B, delivery by the 18th, quoting our account number [ACCOUNT]. Ask them to confirm stock and delivery date. Do not mention price." Output, illustrative: a five-line email with the two items, quantities, date and a request for written confirmation. You add the purchase order number. The saving is a couple of minutes per order, but across 20 orders a month it is a long afternoon. A person approves each order above a set value, and the assistant never has authority to place one.
Second, a realistic mistake. A lab asks an assistant to build a reorder list from a stock sheet and the sheet has the same item under two names ("EDTA tubes 4ml" and "EDTA 4 mL tube"). The draft treats them as separate lines, so one looks overstocked and the other understocked. The fix is a standard item code column before anything is pasted, plus a spot check of the five biggest lines against the shelf. Tidy data first; AI will inherit every inconsistency you give it.
A third useful habit is the weekly "ask log". Every time staff use an AI tool for lab admin, they write one line: date, job, tool, what was checked. Ten lines a week is enough. After a month you can see which jobs are used, which are not and where checks were skipped, and you have something to show an assessor or a client who asks how AI is governed.
What the lab should never give an AI tool
- Result values, flags, ranges or comments, in any form, even "anonymised" if the sample could be identified by context.
- Patient names, dates of birth, identifiers, addresses or contact details in a general chat tool.
- Images or data from analysers, slides or instruments, unless it is a validated and approved feature inside your lab's own system.
- Referrer messages that suggest tests or comment on clinical findings.
Two sets of rules usually decide what a lab may put into a tool. Where health-privacy rules for providers and their business associates apply, any vendor that handles patient information for you must sign a business associate agreement covering the exact product. The big assistants draw that line narrowly: OpenAI signs one for its API and for sales-managed ChatGPT Enterprise or Edu accounts but not for ChatGPT Business, and Anthropic's list covers Claude Enterprise and named API features, not every feature or connector. Where you have patients in the EU, the GDPR treats data concerning health as a special category whose processing is prohibited unless one of its listed conditions applies, and its recitals spell out that this includes information derived from testing a bodily substance or from an in vitro diagnostic test. Any vendor processing that data for you needs a written processor contract.
For a small lab, both point the same way. Admin jobs run on templates and on references that mean nothing outside your system, such as consignment numbers and sample IDs, while names, results and request details stay in the lab information system. Ask your legal or compliance adviser before you choose a tool, and write down which tools are approved for which data. For a general vendor question list, what to ask before buying any AI tool for a medical practice is a good start.
A first-month plan that stays inside the fence
Week 1: write the one-page policy including the hard limit and have the clinical lead approve it. Week 2: pick two jobs from supplier orders, courier reconciliation and price FAQs, because they carry no patient data. Week 3: draft the preparation and report-ready templates, test them on colleagues and lock them. Week 4: review time saved, errors caught and any near misses, and decide whether to add the referrer chasing job.
Track a handful of numbers: minutes per job before and after, calls about price or preparation, rejected samples for missing details, and any incident where a message went out that broke the template. If your lab is part of a larger group or is accredited, tell the quality manager early. For a broader tour of what small clinics can safely hand over, see ten admin tasks a small clinic can hand to AI, and for the front desk specifically, using AI at the front desk safely.
If you would rather map this with someone, that is what my AI implementation consultation covers; get in touch.
Questions lab managers ask about AI
Can AI tell a patient whether a result is normal?
No. Explaining or interpreting a result is a clinical act and belongs with a qualified clinician who has the full context. An admin assistant may say that a report is ready, how to collect it and who to contact with questions. Anything beyond that is handed to a named person.
Can we use AI to chase doctors or referrers for missing forms?
Yes, with care. Drafting a standard message asking a referrer for a missing signature or test code is admin. Keep patient identifiers out of general tools, use a message template merged inside your lab system, and have a person check that the request matches the sample.
Is a report-ready text message safe to automate?
It can be, if it says only that a report is ready and where to get it, and never contains the result or the test name. Check how the patient wants to be contacted, avoid voicemail details, and have a person approve the template.
Do lab information systems already include AI?
Some do, for things like workflow routing or quality flags. Ask your vendor what it uses AI for, whether it influences any result or release decision, and whether you can switch it off. Anything that affects reporting needs the lab's quality and governance process, not a casual settings change.
Further reads
- AI Chatbot for Healthcare: A Small Clinic's Data and Consent Checklist — Data and consent checklist for a patient-facing chatbot.
- A Worked AI Implementation Plan for a Small Medical Practice — A worked implementation plan for a small practice.
- Medical Practice AI Budget: What to Fund First and What It Costs — What to fund first and what it costs.
- How a Clinic Can Use AI to Cut Missed Appointments — Reminder timing that cuts no-shows.
- AI for Fertility Clinics: Admin Jobs Only, and Where It Must Stop — Another sensitive setting with an admin-only rule.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI help article on getting a business associate agreement; Anthropic business associate agreements help page; GDPR text on EUR-Lex (Articles 9 and 28, recital 35); checked 4 November 2026; no clinical sources used.