AI in life sciences: what CROs, medical device makers and consultancies can use it for
What CROs, medical device makers and life-sciences consultancies can use AI for, sorted by whose documents each task has to read.
At a glance
- Most writing about AI in life sciences describes drug discovery and trial design at the largest pharma companies, which is not the daily work of a 50-person CRO, medical device maker or consultancy.
- A more useful way to sort the work is by whose data it needs: public information, your own records, or a sponsor’s or client’s material.
- The third group holds the most valuable work for service firms, and it is the group that sponsor and client agreements often keep out of public AI tools.
- EU GMP Annex 11 says a computerized system used as part of GMP-regulated activities should be validated, and that access to it should be restricted to authorized persons.
- ICH E6(R3) says computerized systems used in clinical trials should be fit for purpose, and that sponsors should protect trial data from unauthorized access and disclosure.
- Where the model runs decides which of the three groups you can use AI on at all, so settle that before you compare tools.
The AI in life sciences you read about, and the AI most firms can use
Search for AI in life sciences and you get the same list on every page: molecule design, target identification, trial optimization, personalized medicine, diagnostics. Large pharma companies publicize their AI programs, and those announcements fill the results. Ask which pharma companies use AI and the answer is that the big ones all say they do, which tells a 40-person CRO very little.
Drug discovery AI is real, but it runs on proprietary chemistry and biology data, specialist teams and budgets that most of the industry does not have. If you run a contract research organization, a medtech company or a regulatory consultancy, almost none of that is your work.
Your work is documents: protocols, monitoring reports, SOPs, deviation records, CAPAs, audit reports, technical files, supplier questionnaires and submissions. Generative AI is good at reading and drafting documents, so this is where it could save the most time. It is also where the constraint lies, because a large share of those documents belong to someone else.
Three bands of work, sorted by whose data it needs
Sorting AI use cases by technology (machine learning, generative AI, agents) tells you little about what you can do on Monday. Sorting them by whose data the task needs tells you which tool is allowed to do it.
Public information
Literature, published guidance, regulations, standards summaries, conference abstracts, a competitor’s public filings. Nobody’s confidentiality is at stake, so any competent AI tool will do, including a free public chatbot. Typical tasks: summarizing a new guidance document, comparing two published versions of a text, drafting a training slide on a public topic, getting a first read of a scientific paper.
The value here is real but modest. Your staff can already read public documents, and the AI saves them reading time, not judgment.
Your own records
This band covers SOPs, templates, work instructions, deviation and CAPA history, audit findings, training records, past proposals and supplier files. It is your firm’s own material, so the question is not permission but access: the tool has to be able to read your file shares and systems, and it should respect who is allowed to see what.
Typical tasks: finding which version of an SOP was in force on a given date, drafting a CAPA from similar past deviations, preparing an audit plan from last year’s findings, answering a client’s due-diligence questionnaire from your quality records. A public chatbot can’t do these well, because it can’t see the records, and pasting them in one at a time defeats the purpose.
Your sponsor’s or client’s material
Here the documents are protocols, investigator brochures, study data listings, monitoring findings, client SOPs, client submissions, contracts, or a medical device maker’s design history file sent to you for review. For most service firms this is the bulk of billable work, and it is the band where AI would save the most hours.
It is also the band where you usually have the least freedom. The material is covered by a master services agreement, a confidentiality agreement or a quality agreement, and many of those restrict who you may share it with. Sending a sponsor’s protocol to a public AI service means sending it to a third party. Whether your agreements allow that is a question for whoever signed them. For many firms the answer is no, which is why so many life sciences teams don’t use ChatGPT on their actual work.
This is the part the overview articles skip. A large pharma company’s AI work runs largely on its own data. For a CRO or a consultancy, the most valuable AI work runs on data it holds on someone else’s behalf.
Concrete tasks by firm type
The table sorts common tasks at CROs, medical device makers and consultancies into the three bands. The band, not the difficulty, decides the kind of tool you need.
| Firm type | Task | Whose data | Band |
|---|---|---|---|
| CRO | Summarize a new or revised guidance document for the team | Public | 1 |
| CRO | Find the SOP version in force when a visit took place | Yours | 2 |
| CRO | Draft an answer to a sponsor’s vendor qualification questionnaire from your quality records | Yours | 2 |
| CRO | Draft a protocol section in the sponsor’s template from the synopsis | Sponsor’s | 3 |
| CRO | Write up a monitoring visit report from the visit notes | Sponsor’s | 3 |
| CRO | Check a site’s consent form against the protocol’s requirements | Sponsor’s | 3 |
| Device maker | Compare two published versions of a standard or guidance | Public | 1 |
| Device maker | Draft a CAPA from similar past complaints and deviations | Yours | 2 |
| Device maker | Summarize complaint trends for a management review | Yours | 2 |
| Device maker | Review a supplier’s quality agreement against your template | Supplier’s | 3 |
| Device maker | Answer a distributor’s or notified body’s questions from the technical file | Yours | 2 |
| Consultancy | Scan the literature for a client proposal | Public | 1 |
| Consultancy | Draft a proposal from your past engagements | Yours | 2 |
| Consultancy | Run a gap assessment of a client’s SOPs | Client’s | 3 |
| Consultancy | Draft an audit report from a client’s audit evidence | Client’s | 3 |
| Consultancy | Review a client’s submission draft before it goes out | Client’s | 3 |
Two patterns stand out. Band 1 is the only band a public chatbot handles well, and it is the least valuable to a service firm. Bands 2 and 3 need a tool that can read your systems, and band 3 adds the condition that the material must not leave your control.
The table stays at overview level. Clinical research tasks, AI across the trial stages and clinical data management each need more depth than this article can give, and so do the specifics of pharma companies and a full list of generative AI use cases.
What GxP expects of a tool used on regulated records
Much of the band 2 and band 3 material is not only confidential but regulated, such as records supporting batch release under GMP or monitoring reports under GCP. When an AI tool reads, drafts or changes them, two texts matter for medicinal products. (Device makers answer to their own quality-system and clinical-investigation rules, which this article does not cover.)
EU GMP Annex 11: the tool is a computerized system
Annex 11 applies to all forms of computerized systems used as part of GMP-regulated activities, and it defines a computerized system as software and hardware components that together fulfill certain functionalities. An AI agent that drafts a deviation investigation is software doing exactly that.
The rest of the Annex sets out what follows from that:
- The application should be validated and the IT infrastructure qualified, and the extent of validation and data integrity controls should rest on a justified, documented risk assessment.
- Validation documentation and reports should cover the relevant steps of the system’s life cycle.
- When a third party provides, maintains or retains a computerized system, or processes data, formal agreements must exist that state the third party’s responsibilities. A cloud AI service that processes GMP data fits that description.
- Data should be secured by physical and electronic means against damage, and checked for accessibility, readability and accuracy.
- Based on a risk assessment, consideration should be given to building in a system-generated audit trail of all GMP-relevant changes and deletions, with the reason for each documented.
- Physical or logical controls should restrict access to the system to authorized persons.
The Annex also says that where a computerized system replaces a manual operation, there should be no resulting decrease in product quality, process control or quality assurance. That gives a QA reviewer a usable test for an investigation drafted with AI: is it at least as good, and as traceable, as the one a person would have written?
Whether a particular AI use falls inside Annex 11’s scope depends on what the tool touches and whether its output becomes a GMP record. That is a judgment for your QA function. A tool that drafts text a qualified person then reviews, corrects and signs is a different case from one that writes directly into the quality system.
ICH E6(R3): computerized systems and trial data
ICH E6(R3) applies to interventional clinical trials of investigational products that are intended to be submitted to regulatory authorities. Two parts of it bear on AI tools.
On computerized systems, the principles say that computerized systems used in clinical trials should be fit for purpose, for example through risk-based validation if appropriate. The guideline adds that the responsible party is responsible for a system’s validation status throughout its life cycle, and should ensure that systems are validated as fit for purpose, including those developed by other parties. A tool bought from a vendor is still the responsible party’s to validate.
On confidentiality, the principles say that the confidentiality of information that could identify participants should be protected in accordance with applicable privacy and data protection requirements. The guideline also asks the sponsor to ensure that trial data are protected from unauthorized access, disclosure, dissemination or alteration, and from inappropriate destruction or accidental loss.
For a CRO, one more principle matters. Where a sponsor has transferred activities to a service provider, the responsibility for the conduct of the trial, including the quality and integrity of the trial data, resides with the sponsor. It is one reason sponsors limit where a CRO may send their data: the data stays the sponsor’s responsibility wherever it goes.
Neither text mentions AI by name. Both mean that an AI tool used on trial data or GMP records is a computerized system like any other, and the questions an auditor asks of your EDC or eQMS (who can access it, what it does with the data, how you know it works) apply to it too.
Where the model runs decides which bands you can use
Put the three bands next to the two constraints and the decision becomes simple.
| Where the model runs | Band 1: public | Band 2: your records | Band 3: sponsor or client material |
|---|---|---|---|
| Public AI service, personal accounts | Yes | Poor fit: it can’t read your systems, and records leave the firm | Often ruled out by the agreement |
| Business tier of a cloud AI service | Yes | Possible, under that vendor’s contract | Only if every affected agreement allows that vendor |
| A model running on hardware in your office | Yes | Yes, with your existing permissions | Yes, because the material doesn’t leave your premises |
A business cloud tier improves the contract but not the geography: the sponsor’s protocol still goes to a third party’s servers, and the question for band 3 becomes whether your sponsor agreements allow that particular vendor. Some will. Many service firms find it easier to keep the model where the documents already are.
If band 1 is all you need, a public tool is fine and cheaper than anything else. If bands 2 and 3 are where your hours go, the deciding question is where the model runs, not which model. Our guide to on-premise AI explains what it means to have the model on a device in your own office, and private AI for business compares private deployments with the cloud. If your band 3 material includes personal data of EU participants or patients, GDPR-compliant AI covers the data protection side.
Where IntraMuros fits
IntraMuros puts the model where band 3 needs it: on a device in your office, run by a vendor. Mayet, a British company that builds software for pharma, biotech, medical devices, clinical research and other regulated industries, ships the device in a couple of days. You plug it in, Mayet configures it remotely in a few hours and connects it to your documents and email, and can connect it to software such as your eQMS, eTMF, CTMS or ERP, whether those sit in the office or in the cloud. It is running within a week of signing. Mayet then patches, monitors and backs it up, and replaces it if it fails.
The model runs on the device itself, so a sponsor’s protocol or a client’s SOPs do not go to an AI company, and your documents and conversations are never used to train any AI model, by Mayet or anyone else. Staff use the agent, MURO, from a browser, signing in with the Microsoft or Google work account they already have. MURO works with that person’s permissions, so someone who can’t open a client’s folder or a sponsor’s study in connected software today can’t ask questions about it either. Every answer shows the document it came from. Mayet connects to the device to update, monitor and back it up. As the one outside party that administers it, Mayet belongs on your vendor list.
That makes it a fit for bands 2 and 3: the SOP lookup, the CAPA draft, the protocol section in the sponsor’s template, the gap assessment of a client’s SOPs. The how it works page explains the set-up, and the medical writing and quality and audit pages describe those two uses.
Everyone in the firm can use it, and maintenance is £1,000 a month or £10,000 a year for one device. The device is a one-off charge with set-up included: a consultancy of about ten fits Base at £4,500 or Pro at £7,500, and a 40-person CRO or medical device maker fits Enterprise at £12,000. A 50-person firm takes Enterprise plus more devices as needed, each at its one-off charge plus £200 a month or £2,000 a year. When it is busy, questions wait their turn rather than being refused. See pricing.
It is the wrong choice in three cases. If your AI work is band 1, a public tool does it for less. If you need drug discovery or molecular modelling, IntraMuros is a document agent and does no molecular work. And it does not validate itself: if your QA decides a use falls under Annex 11 or E6(R3), the validation and qualification work is yours, as it would be for any computerized system.
Sources
- ICH E6(R3) Good Clinical Practice (r3, adopted 6 January 2025; Annex 1)
- EudraLex Volume 4, Annex 11: Computerised Systems (2011; Principle, §1, §3.1, §4.1, §7.1, §9 and §12.1)