AI for Procurement in Retail

A retail procurement function gets described as one department and behaves like two. On one side, buyers decide what a customer will want next autumn and which supplier will hold a margin through a bad season. On the other, somebody is working out why a store paid a local firm to clear a blocked drain when a national contract covers exactly that, and why a packaging supplier's insurance certificate lapsed in the spring unnoticed. Both sets of people work in procurement. Only one of those jobs is a sensible candidate for an agent, and it is not the glamorous one.

Resale buying and everything else are not the same job

Buying goods for resale is a commercial judgement job. Range, price architecture, what a customer will pay attention to in nine months, when to walk away from a negotiation. Much of what decides those calls is unwritten: a buyer's read of a market, a supplier's behaviour over three years, a trading director's appetite for risk this quarter.

An agent has very little honest claim on that work.

Everything else is where the volume is. Goods bought for the business rather than the shelf, and the administrative tail of several thousand supplier relationships. That half is reading, checking, matching and chasing against agreements that already exist, which is the shape of work an agent does well.

The half that stays with the buyers

This is deliberate rather than cautious. The buy, the range decision and the negotiation belong to buyers, and anyone claiming otherwise is selling something.

What can change is what a buyer walks into a meeting holding. Supplier history, current terms, open claims, performance against the agreement and funding already committed can be assembled in advance rather than remembered. Useful, and not the decision.

The categories nobody puts their name to

Goods not for resale is a label covering an odd collection: store fit out and fixtures, cleaning and waste, uniforms, maintenance, energy, packaging, print and the material that sits by the till, agency labour, logistics services. Individually most feel too small to manage. Together they are a serious number, and they rarely have a buyer in the sense the resale side understands the word.

They have a budget holder instead, in property or operations or marketing, who buys as a side effect of a different job. The agreements get negotiated once, land in a folder, then compete for attention against the actual work. Nobody is careless. No one's week is spent making sure a cleaning contract gets used the way it was written.

Why a store manager rings someone local

A freezer door will not shut, the store trades in two hours, and the manager knows a company that answered last time. A reasonable decision, taken with no sight of a national agreement.

A quarterly spend review finds it months later, by which point the conversation is about discipline rather than information. Naming it at the requisition changes the tone: this sits under an agreement, here is the contracted supplier and the response time you already pay for, do you still want to proceed. Some will proceed anyway, and the reason they give is the most valuable thing the business throws away.

Hundreds of small urgent jobs, raised by whoever noticed

Maintenance and facilities requests arrive from every store in the estate. Each is small, each is urgent to the person raising it, and each is written in whatever words that person uses. A light is out in the back corridor. The shutter is sticking again.

An agent can read that request, recognise what trade it needs, route it to the contract that already covers it, and tell the raiser what to expect. Where no contract covers it, it can say so plainly rather than leave the store to improvise. It can also spot the fourth request about the same shutter in two months, which is a capital conversation disguised as a repair queue.

A supplier sends a new cost and a deadline

Cost price increase requests arrive by email and spreadsheet ahead of a deadline, in volume, each with a reason attached and a date it takes effect. The honest failure mode is not that they get refused badly. It is that a tired team accepts most of them because a supplier sent them and there is no time to check each against what was agreed.

Every agreement has a mechanism. Some lines are fixed for a term, some move with a stated index, some require notice of a particular length, some are not open for review until a date. An agent can read each request against the trading terms it belongs to and produce the exceptions rather than the list: these sit outside the agreed mechanism, these arrived without the required notice, these contradict a commitment from the last review. A buyer then argues the cases that matter instead of skim reading the pile.

Promotional funding, agreed in a room and claimed in a system

Promotional support is agreed in a conversation, recorded in a document, executed in a trading calendar and recovered through a deduction. Four places for the same number to live, with no guarantee they agree.

The gap is rarely dishonest. A promotion moves, a volume assumption does not hold, somebody agrees a contribution verbally and the paperwork never catches up. The result is funding claimed that cannot be evidenced, or funding agreed that nobody claimed. An agent can hold the agreement next to the activity and the deduction, and flag where the three fail to reconcile while the period is recent enough for someone to remember it.

Paperwork that quietly stops being true

Supplier onboarding in retail is a document problem before it is a commercial one. Insurance certificates, ethical audit reports, technical and factory audits, product certifications, bank details. Each arrives once, gets filed, and carries an expiry date somebody is theoretically watching.

An agent can read these rather than store them, which is the difference that matters. A certificate on file proves nothing if its cover limit sits below what the agreement requires, or if the site it names is not the site now making the product. Chasing the renewal before the date bites removes a recurring source of trouble.

The pack that is almost complete

Most onboarding delay comes from a pack that is nearly finished. An agent can assemble what has arrived, state what is missing and who it is waiting on, and keep asking without anybody remembering to. That alone shortens the path from agreed supplier to usable one.

Responsible sourcing, and the once a year scramble

Modern slavery and responsible sourcing reporting runs on evidence that was always meant to be on file: site declarations, audit outcomes, corrective actions closed or open, tiers below the direct supplier. Collected annually under pressure, it is thin by construction. Collected continuously through onboarding and renewal, the annual exercise becomes a reading job rather than an excavation.

When the label, the specification and the law have to agree

Private label adds a publishing problem to a buying one. The specification, the ingredient or component declaration, the artwork copy and the legal requirements for the market all have to say the same thing before a print run commits.

They are held by different people in different file formats, and the errors are small: a weight that contradicts the specification, an allergen in the recipe that never reached the label, a marketing claim the specification does not support. An agent reading a specification and an artwork pack against each other, and against the rules for the market, catches the mistake that otherwise becomes a reprint or a withdrawal.

Groundwork for a call the merchant makes

A range review is a decision a human makes, preceded by weeks of assembly nobody enjoys. Current performance, supplier submissions in whatever template they chose, costs, terms, availability and the commitments made last time.

That assembly is the part worth handing over. An agent can pull a consistent pack per line, normalise submissions that arrived in different shapes, and name where a supplier's claim is not supported by the figures held internally. The merchant still decides the range, with a shorter run up.

Performance against a service level you agreed to measure

Most supplier agreements contain commitments about availability, delivery accuracy, lead time and quality. Many businesses measure them sporadically, so the supplier conversation runs on impression rather than record.

Short deliveries, late arrivals and quality rejections already sit in your operational data. An agent can build a scorecard from what happened rather than what anyone recalls, raise the claims the agreement entitles you to, and prepare the file behind each. That turns a difficult meeting into a specific one.

The agreement that renewed itself while nobody looked

Somewhere there is a contract that rolled for another year because its notice window passed in a quiet month. There is also a tail of small agreements nobody has read since signature, where the service still runs, the invoice still clears and nothing is tracking either.

A renewal calendar that actually fires is modest work with a disproportionate return. An agent can hold every term, notice period and break date, raise each with runway for a decision rather than a scramble, and bring the performance record with it.

Where AI fails here

The buy is not on the table. Range, price architecture and negotiation stay with buyers and merchants, because those decisions carry consequence and context no model holds.

A supplier audit certificate can be read, but a supplier cannot be judged from paper. A clean report describes a day. Deciding whether a factory is acceptable needs people who have been there.

Bank detail changes are a fraud surface. An agent can notice the request and flag it. Verification belongs with a person, through a channel that was not used to make the request.

Delisting a supplier is a commercial decision with consequences nothing automated carries: continuity of supply, a gap on a shelf, a relationship worth more than the dispute.

An agent cannot settle a disputed deduction. It can assemble both positions accurately, which is most of the preparation, and then two commercial teams have to agree.

Where to start

Take one goods not for resale category with real volume and poor ownership, and run requisition routing end to end for a quarter. Maintenance is usually the honest choice: the requests are frequent, the contracts exist and the leakage is visible. Baseline first, measuring how many requisitions arrive, how many land on an existing agreement, and how long each takes to place.

Then add the document and renewal watch across the same supplier set. It costs little once the agreements are readable and closes a predictable category of problem.

Adopt AI Partners works on this part of the function, the half nobody is proud of, and leaves the buying to the buyers. To talk it through against your own categories, hello@adoptaipartners.com.