Lab-AI — What artificial intelligence can do in your laboratory

Not a "smart lab" — concrete tasks an AI assistant does faster and more carefully than a tired person, and the line beyond which the decision stays with you. Click a stage on the sample’s path.

The sample’s path — where AI really helps

1. 📨 1. Client enquiry

   What AI does: Reads the quote request email, extracts parameters, matrices and standards, checks them against your scope of accreditation and drafts a reply with items to price.

   Time: 20 min → 3 min

   The line — this stays with a person: A person sets price and deadline; AI promises the client nothing on its own.

2. 🧪 2. Sample receipt

   What AI does: From a photo of the label and the sampling record it fills the register: identification, date, conditions, sampler. Raises an alarm when the holding time for the ordered test has already passed.

   Time: 5 min → 1 min

   The line — this stays with a person: The decision to accept a sample in doubtful condition is made by the receiving person — and agreed with the client.

3. 🔬 3. Testing

   What AI does: Watches the control chart in real time: a result outside the usual range for that matrix gets an "are you sure?" before it moves on. Spots method drift earlier than the monthly review.

   Time: continuous, in the background

   The line — this stays with a person: Repeating a test and judging whether a result is valid is a decision of the authorised analyst.

4. 📊 4. Calculations and uncertainty

   What AI does: Checks the calculation sheet: units, conversions, rounding, uncertainty budget components. Writes the first draft of a budget for a new method with a cause diagram and a component table.

   Time: 1 day → 2 hours

   The line — this stays with a person: Component values and the coverage factor are approved by the technical manager — AI does not know your instrument.

5. 📄 5. Report

   What AI does: Compares the report with the standard’s list of required elements: method with year, uncertainty where needed, subcontractor results marked, decision rule with a statement of conformity. Points out gaps before authorisation.

   Time: 15 min → 2 min

   The line — this stays with a person: Authorising a report is the signature of an authorised person. Always.

6. 💾 6. Subcontracts and documents

   What AI does: Reads the subcontractor’s PDF — header, table, footnotes — matches rows to order items, normalises units and the "<" sign, shows the proposal next to the original fragment. Drafts procedures and audit checklists.

   Time: 8 min → 1.5 min per report

   The line — this stays with a person: AI does not guess missing information — it reports the gap. Only a person’s confirmation saves the result.

Which assistant for what

Six tools laboratories choose most often. Green dot — can do it, yellow — partly or in a paid plan, grey — no. As of September 2026; check with the vendor before buying.

Tutorial: five steps, five minutes

For people who have never talked to an AI assistant about laboratory work. Every step has ready text — copy and paste.

1. 1. Say who you are and what you want

   The assistant answers a quality manager differently than a student. One sentence about role and goal changes everything.

   I am the quality manager of a water testing laboratory accredited to ISO/IEC 17025. Answer concretely, with clause numbers, no generalities.

2. 2. Give context, not just the question

   Matrix, method, instrument, range — without them you get a textbook answer, not one for your laboratory.

   We determine lead in drinking water by ICP-MS to EN ISO 17294-2, range 0.5–50 µg/l, instrument calibrated every 12 months.

3. 3. Ask for a form you can actually use

   A table, a checklist, a procedure with numbered points — the assistant can do any of them, but you have to say which.

   Prepare a table: uncertainty component | type A/B | data source | value | unit. At the end calculate the combined and expanded uncertainty for k=2.

4. 4. Make it check itself

   An assistant can invent a standard number that sounds real. Ask it to flag what it is not sure about — and check it yourself.

   Mark in your answer everything you are not sure about and tell me where to verify it: in the standard, with the manufacturer, in the accreditation body’s document.

5. 5. Record what came out of the conversation

   The assistant’s answer is not a quality system record. Before it becomes a procedure it goes through review and approval like any document.

   Summarise our conversation as a draft document for review: title, purpose, scope, steps, records. Add at the top: "Draft generated with AI assistance — requires review and approval".

Try it now — ready-made prompts

Eight prompts from real laboratory work. Copy, paste into ChatGPT, Claude or Gemini, add your details.

• 📋 Checklist before a PCA assessment

  In two weeks I have a surveillance assessment by the accreditation body in a water testing laboratory. Prepare a checklist for clause 7 of ISO/IEC 17025: documents to lay out, questions the assessor will ask, and five things to do this week.

• 📐 Uncertainty budget for pH

  Help me build an uncertainty budget for pH measurement of water. pH meter calibrated with buffers 4.00, 7.00, 10.00, sample temperature 20±2 °C, repeatability from 10 measurements: 0.02. Give uncertainty sources, type A/B, values and the expanded uncertainty for k=2.

• 📝 Nonconforming work procedure

  Write a draft procedure for nonconforming work (clause 7.10 of ISO/IEC 17025) for a chemical laboratory: purpose, scope, responsibilities, steps from identification to resuming work, a record form.

• 🧪 Method validation plan

  I need to validate the determination of Pb, Cd, As and Hg in drinking water by ICP-MS. Propose a validation plan: parameters, acceptance criteria, number of replicates, reference materials and a 6-week schedule.

• 📑 Reading a subcontractor’s report

  I attach a test report from a subcontractor. List in a table: determination, result, unit, uncertainty, limit of quantification, method with standard year, remarks from footnotes. Mark everything not in the document as MISSING — do not guess.

• ✅ Decision rule for a client

  Explain to a client in three paragraphs, without jargon, how simple acceptance differs from a guard band when stating conformity of a result with a limit, and which rule you recommend for drinking water.

• 📅 Internal audit plan

  Draw up a twelve-month internal audit plan for a laboratory with three sections (physico-chemistry, microbiology, sampling) covering every clause of ISO/IEC 17025, with the rule that nobody audits their own section.

• 🎓 Training for a new technician

  Prepare an onboarding plan for a new technician for COD determination by spectrophotometry: stages from observation to independent work, competence criteria, number of control samples, how to document the authorisation.

Five red lines

What is never handed to AI in an accredited laboratory — no matter how good the tool is.

• ✍️ Authorising a report

  The signature of an authorised person. A tool that saves a result without confirmation shifts responsibility to the vendor — no assessor will accept that.

• 🔒 Client data in a third-party tool

  Client names, results and samples do not go into a free chat. First a data processing agreement, data in the EU, no training on your prompts.

• ❓ Guessing missing data

  Standard year, unit, limit of quantification — if it is not in the document, AI must report the gap. "Because it usually is so" disqualifies a tool in an accredited laboratory.

• ⚖️ Conformity assessment and the decision to repeat

  AI can show that a result lies at the limit. Whether it "passes" or the test is repeated is decided by the analyst authorised for the method.

• 🧾 Reading without an audit trail

  Every accepted reading has recorded: source document, time, model version, proposed value next to the approved one, accepting person. Without that nobody can trace a number a year later.

Test: is your laboratory ready for AI?

Eight yes-or-no questions. Instant result, no account.

1. We have written down who may use AI tools and for what.
2. We know where the tool vendor stores our data and whether it trains a model on it.
3. No result reaches a report without confirmation by an authorised person.
4. Our LIMS has a change log: who, when, what and the value before the change.
5. Spreadsheets and software that calculate results are validated and we have the record.
6. We have one repeatable task to start with — not "everything at once".
7. We can verify an AI answer against the source: the standard, the manufacturer’s document, the accreditation body’s guidance.
8. Management understands that AI changes the distribution of risk, not just hours — and wants that described in the management system.

Your result

• 7+ points: Ready. Start with one task.

  Pick the subcontractor who sends you the most reports and run AI reading alongside manual retyping for a month. The differences are worth more than the matches.

• 4+ points: Almost. Two or three things to close.

  What is usually missing is the data agreement with the vendor and a record of who accepts a reading. Without those two the assessor stops at the first question.

• 0+ points: Not yet — and good that you know.

  Start with clause 7.11 of the standard: a change log in the LIMS and validation of spreadsheets. That is the foundation AI can later stand on safely.