ADOPT AI PARTNERS
Insights / pharma

AI for Procurement in Pharma: Qualified Supply, Not Just Cheaper Supply

6 October 2026  ·  8 minute read  ·  Adopt AI Partners
clinician walking a hospital corridor
Photo via Unsplash

In most industries procurement is measured on price, lead time and risk. In pharmaceuticals, biotech and contract manufacturing there is a fourth constraint that outranks the other three: the supplier has to be qualified, and has to stay qualified, before anything they sell can touch a product that reaches a patient.

That single fact reshapes the whole function. A category manager who has run indirect spend elsewhere arrives expecting to negotiate. What they find is that the negotiation is often the short part, and the long part is evidence. Who audited this supplier, when, against what. Which quality agreement is in force and does it still describe what the supplier actually does. Does the certificate of analysis format match the specification we approved. Which of our filings name this material, this grade, this site. Procurement here is entangled with quality assurance in a way that cannot be organised away, because the two functions are deciding different halves of the same question.

Agents are useful in that entanglement, and useful in a narrow way. They are good at assembling, comparing and chasing evidence across systems that do not talk to each other. They do not make a qualification decision and they do not release material.

What makes pharma procurement a different job

The distinguishing feature is GxP relevance. If a material, a service or a system can affect product quality, identity, strength, purity or safety, the supplier of it falls inside the quality system. That means a qualification pack, a quality agreement or technical agreement setting out who does what, an obligation on the supplier to notify you of changes, an audit history, and a requalification cycle that never ends.

Direct and indirect spend are governed differently

Direct spend is where the quality entanglement is heaviest. Active pharmaceutical ingredients and intermediates, excipients, primary packaging in contact with the product, secondary packaging carrying printed regulatory content, single use consumables such as filters, tubing and bags, reference standards and critical reagents. Each of these sits in a specification, and the specification is often referenced by a marketing authorisation. Changing the supplier, or in some cases only the grade or the manufacturing site of the same supplier, can be a regulatory filing change rather than a sourcing decision.

Indirect spend is not free of this either, which is where general procurement instinct goes wrong. Contract research and contract manufacturing organisations, analytical testing laboratories, calibration and sterilisation services, pest control and cleaning in classified areas, computerised systems holding GxP records: all of these are qualified suppliers with audit expectations and agreements, even though the finance system treats them as services. Only the genuinely unregulated tail, office supply, travel, general marketing, behaves like ordinary indirect spend, and that tail is usually a small share of the money.

Add serialisation and traceability expectations, which push requirements down into packaging suppliers and contract packers, and the picture is clear. Category strategy in pharma is constrained sourcing. The cheapest compliant bid wins, and plenty of bids are not selectable at any price.

The real workflow, and where the drag is

Written out, the process looks orderly: a request, a specification, a sourcing event, qualification and quality review, a contract, a purchase order, goods receipt and release, performance review, then periodic requalification. In practice the drag concentrates in a few predictable places.

None of this is work that rewards heroics. It rewards completeness, and completeness is where human attention is least reliable.

industrial robot arm on a factory line
Photo via Unsplash

Assembler, comparer, chaser

Set the expectation correctly: an agent here is an assembler, a comparer and a chaser. Within those verbs it can take a very large amount of work off the team.

Assemble and chase qualification evidence. Given a supplier and the category of material or service, an agent can work the required evidence list, see what is already held and current, request the rest, and chase on a schedule until it arrives. It reads what comes back, checks the document is the right document at the right version covering the right site, and files it. What reaches quality assurance is a complete pack or an explicit statement of what is missing and who was asked.

Compare supplier quality responses against the specification. When a supplier returns a questionnaire, a certificate of analysis template or a technical data sheet, an agent can read it line by line against the approved specification and return a difference list: this limit is wider than ours, this test method is not the compendial one, this grade statement does not cover the animal origin requirement, this site is not the site named in the licence. That is mechanical comparison work, and doing it well is mostly a matter of not getting tired.

Keep the requalification calendar and the expiring document register honest. Audit due dates, certificate and licence expiry, agreement review dates, performance review dates. An agent can hold that register against the actual documents rather than against a spreadsheet someone last touched a year ago, and start the chase far enough ahead that a lapse is avoidable instead of discovered.

Triage supplier change notifications. This is the highest value item on the list. An agent can take an incoming notification, classify what is changing, and trace it through: which specifications reference the affected material or grade, which products use it, which filings name it, which quality agreements oblige a response within a stated period, which open batches or campaigns are exposed. It hands quality and regulatory a scoped impact list with citations instead of an email forwarded with a question mark.

Prepare the tender comparison. Not the award. The comparison: bids normalised to the same basis, total cost built up including qualification cost and lead time, compliance gaps flagged per bidder, and a clear separation between bidders who are selectable and bidders who are not yet qualifiable.

Draft the first pass of the quality agreement redline. Against your template, marking where the supplier's version shifts responsibility, weakens notification obligations, limits audit rights or contradicts what the technical agreement says. A quality reviewer then argues with a marked-up document rather than building the markup.

Second sourcing is a programme, not a purchase

The single most useful reframing for anyone new to this category. Qualifying a second source for a critical material is a project with technical, analytical, regulatory and commercial workstreams, and it runs for months or longer depending on the change classification and the markets involved. Treating it as a sourcing event produces a supplier you cannot buy from.

Where agents help is in making that project legible: holding the evidence and change control dependency list, tracking what is outstanding against whom, and keeping the picture current so the business can judge whether the dual source will exist before the risk it was meant to cover materialises. The sequencing judgement, and the regulatory strategy, stay human.

Where AI fails here

Be blunt about this, because the failure modes are the expensive ones.

A qualification decision stays with a named accountable human. An agent can show that every required element is present and consistent. It cannot approve a supplier. Approval is a quality decision carrying personal accountability, and the signature has to belong to someone who can defend the judgement to an inspector.

A release decision stays with a named accountable human. Nothing an agent produces releases material or product. Goods receipt checks can be prepared; release is signed.

A validated system has validation consequences. If the agent writes to or influences a GxP record, it falls inside your computerised system compliance obligations: intended use documented, risk assessed, qualified, controlled, audit trailed, with change control over its own behaviour. An agent that reads and drafts outside the validated record is a far lighter proposition than one that writes into it, and most of the value sits in the lighter version. Deciding which side of that line you are on is the first design decision, not an afterthought.

Regulatory impact classification is advisory only. An agent can identify which filings reference a material. Whether a change is a variation, a notification or nothing is a regulatory affairs call informed by market-specific rules and precedent, and should never be auto-asserted.

Judgement about supplier capability does not come from documents. Culture, data integrity maturity and how a site behaves when something goes wrong are learned from audits and from the relationship. A complete pack and a weak supplier are compatible states.

Garbage specifications produce confident nonsense. If the approved specification is ambiguous or out of date, comparison output will be precise and wrong. Fix the specification first.

interior of a large industrial factory
Photo via Unsplash

Where to start

Pick change notification triage. It is painful, it is unowned, the inputs are already in email, and nothing in it requires writing to a validated record. Run it alongside the current process for a cycle and compare the scoped impact lists against what humans produced. Then take the requalification register, because the value there is simply that it stops being a lie.

If you want to talk through where the line between drafting and deciding should sit in your own quality system, write to hello@adoptaipartners.com.

Tell us what your teams keep redoing

One conversation is enough to say whether there is a system worth building, and what would make it fail.

Describe your process