AI for Procurement in Higher Education

A requisition lands on a Tuesday morning. The description field reads "reagents for project". The value is 46,000 pounds, the cost centre is a research grant, and the supplier is a manufacturer the institution has never paid. Attached is a six week old quotation and an email from a laboratory manager saying the instrument has already been delivered and is standing in a corridor because nobody has found a bench for it.

Before that becomes an order, someone must establish what is being bought, whether the value aggregates, whether a framework covers it, whether the grant permits the cost, and whether the supplier has been checked. None of it is in the requisition. All of it is somewhere.

A control function with no line management

A university procurement team is small, perhaps six people. Spend is devolved to hundreds of budget holders who do not report to it: heads of school, principal investigators, laboratory managers ordering consumables weekly. Procurement owns the order number, the supplier record and the tender process. It does not own the budget, the specification or the supplier relationship, and it cannot instruct anyone. Its leverage exists only between a budget holder deciding and an order being raised, and much of the job is discovering that the window has closed.

The purchase that has already happened

The fait accompli has a predictable shape. An academic meets a supplier at a conference, agrees a price, takes delivery, and the invoice reaches accounts payable with no order against it. The institution has a no order no payment policy and an exception route, because the alternative is not paying for goods it is already using.

The artefacts are familiar: a retrospective requisition raised to cover an invoice, a three word description that cannot be checked against anything, a quotation that became an order confirmation without anyone agreeing terms. An agent earns its place before the escalation, reading the quotation and the email thread and turning three words into a structured record of item, quantity and terms, so the person handling the breach argues about the right thing.

Most of this is reading one request against four rulebooks

Every request is tested against several sets of rules at once. Financial regulations set delegated authority and quotation requirements. Public contract regulations apply if the institution is a contracting authority, which most are. Funder terms apply if the money is a grant. Internal policies engage depending on what is bought: security review for anything touching student data, chemical approvals, estates standards for anything fixed to a building.

Nobody holds all of that in their head. The daily work is reading an unstructured request, working out which rulebooks engage, finding the clause and routing it to the right approver with the evidence. That is language work bounded by written rules, and it is the most agent shaped task in the function. What sits outside it is narrow: deciding whether a sole supplier case is acceptable, deciding how much risk the institution carries, and telling a senior academic that what they already bought cannot be paid for yet.

Thresholds set the route before anyone argues about the supplier

Value determines procedure. The question is never what this order costs but what the requirement costs: the purchase, the service contract and the consumables that only fit this instrument, across the full term including extensions and aggregated with similar purchases expected. Something that looks like three quotations becomes a tender once aggregated, and the aggregation is the step that gets skipped. Thresholds vary by jurisdiction and change, so the agent applies the institution's current written rules and shows its arithmetic.

The justification written to match the decision

Sole supplier justifications are usually written after the supplier was chosen, which is why so many read like the supplier's product literature with the brand removed. An agent can compare the justification against the specification, flag requirements that are proprietary rather than functional, and search the institution's history for comparable items bought elsewhere. Accepting or rejecting the case belongs to a named person.

The framework question nobody can answer quickly

Universities buy through consortium and framework agreements, and the recurring question is whether a supplier is actually on one this institution can use. Answering it takes four checks: the agreement is live, the institution is a named eligible body, the lot covers this scope rather than something adjacent, and the call off rules permit direct award rather than further competition. Then the price has to be read against the framework schedule of rates, because a supplier quoting list price to a framework customer is common and rarely caught. An agent holding those documents answers in seconds with the clause cited, including the answer nobody wants: the convenient framework does not quite cover the requirement.

Eligible cost clauses, read four years later

Grant funded purchases carry a second rulebook written by the funder. The award letter defines eligible and ineligible costs, equipment treatment, whether the funder requires competitive quotations below the institution's own threshold, and the date after which no cost may be incurred. Those clauses are read once, at award, by someone in research services. The buying happens in a department that has never seen the letter.

The exposure arrives years later, when an audit samples transactions and asks why an ineligible item sits on the grant. By then the money is spent, the staff member has gone, and the finding is a clawback. An agent can convert those cost clauses into checks applied at requisition time, flagging the purchase outside the eligible period or on the wrong budget heading before the order goes out.

What has to be true before a supplier gets a number

A new supplier record is a control point with a long list. Verified bank details, because mandate fraud against universities is persistent and convincing. Insurance certificates that will expire mid contract. A modern slavery statement and, in higher risk categories, real answers about the supply chain rather than a policy attachment. Export control screening for dual use equipment and its end destination, which matters at an institution with overseas campuses. Data protection terms where student data is processed, and a conflict of interest check for the supplier owned by a member of staff.

Most of that is collection, validation against a checklist and chasing. An agent runs it end to end and keeps the expiry dates live. It should never approve a supplier, and never confirm bank details, which need a call to a number found independently.

The estates programme and the pipette tips share one team

The same team handles a capital estates project worth tens of millions and a weekly order for gloves. Capital work uses recognised contract forms and turns on stage payments, retention and whether a variation has grown large enough to have needed its own competition. The everyday work is thousands of low value lines with no strategic content. They want opposite treatment. On consumables an agent can take the whole volume: classify the request, find the catalogue line, check the framework price, route the approval. On capital it should assemble and never conclude, tracking bid commitments into the contract and flagging cumulative variation value for a human to judge.

Eleven ways to describe one chemical

Spend analysis at a university stops at the point it becomes interesting. Supplier names exist in several spellings, half the spend is free text on non catalogue lines, and chemicals are ordered by trade name in one school and catalogue number in another, so one substance appears eleven ways at four prices with no field connecting them. An agent reading line descriptions, quotations and catalogues can classify the whole ledger and show the duplicates. The output is a question for a category manager, not a savings number for a board paper: these eleven lines look like one item, here is the price spread, and somebody who knows the laboratories decides whether they really are.

Nobody was holding the date

Contracts auto extend because the notice window passed while everyone was busy. The register is a spreadsheet, the dates were typed from the executed PDF, and some were typed wrong. The review that should have run six months before expiry happens six weeks after the extension triggered. An agent can read executed agreements, extract the term, extension mechanics, notice period and price review clauses, and raise each in time. The valuable part is what it attaches: spend since signature, service issues logged, and whether the volumes that justified the original pricing ever materialised.

Telling a professor no is not an agent's job

The hardest part of this function is a conversation. A head of school has committed to a supplier, the paperwork does not support it, and someone has to say so to a colleague whose cooperation they will need for years. That requires reading the politics and judging whether this is the hill worth choosing. An agent can prepare that person: the clause that applies, the exposure if it is waived, precedents where similar cases were allowed, and the compliant alternatives. It should not send the message or be the reason given. The authority here is personal, and borrowing it from a system is how a team loses it.

Where AI fails here

The document corpus is worse than anyone admits. Framework agreements live on consortium portals, local copies go stale, and an agent citing a superseded schedule of rates with confidence is worse than one saying it does not know.

It cannot tell you whether a sole supplier case is true, only whether it is argued and whether comparable purchases exist. Whether the researcher needs that instrument to continue a method validated over three years is beyond it.

Spend classification is probabilistic: a category view that is mostly right is useful for direction and unsafe as a savings claim.

Funder terms vary by funder, by call and by variation letter. An agent reading only the headline award letter misses the amendment that changed the equipment rules, so ingestion must cover the whole award file.

It cannot create authority it does not have. Departments route around central processes that feel slow, and if the agent only adds a step they will route around it faster.

Where to start

Take non catalogue requisitions from one faculty below the tender threshold, and build triage only. The agent reads the request and its attachments, normalises the description, checks framework coverage and grant eligibility, aggregates against related spend, and routes to the right approver with the evidence. Run it in shadow for a quarter against requisitions already processed, and measure it on the contested cases.

Leave capital estates alone. Leave sole supplier approvals alone. Do not let it write to the finance system, and do not let it tell anyone no. If you want to work out which parts of your own buying could run this way without crossing that line, describe it to Adopt AI Partners at hello@adoptaipartners.com.