AI for Procurement in Energy: Buying That Keeps Assets Running

A buying desk that supports generating plant, a network, a platform or a wind fleet is organised around assets and the dates those assets are available, not around categories. The same desk buys bearings and gaskets, rope access crews, a transformer that arrives in two winters, and the paperwork behind all of it.

Volume is lumpy, technical content is high, and lateness is measured in lost availability. This page is about where AI agents help inside that work, and where they do not.

The calendar bends around the outage window

Everything on the desk is shaped by a window that does not move. Scope firms up late, the materials list later still, and requisitions arrive in a cluster from engineers who all reached the same decision in the same week.

An agent is useful here in a dull and valuable way. It can hold the outage scope against what has been requisitioned, ordered and acknowledged, work backwards to the date each line must be placed, and keep the uncovered list current rather than rebuilding it in a meeting. Expediting is the same work under pressure: chase on a schedule, record each answer, and surface the suppliers who have gone quiet.

Spares that were classified once and never since

Criticality ratings, min and max levels and reorder points are usually set at commissioning by someone who has since left. The asset has been modified, the duty has changed, the failure history says something different, and the settings still say what they said.

An agent can put the evidence for a review in front of a planner: issue history by part, the time since each line last moved, holdings across sites, and where a rating disagrees with the maintenance record. The judgement stays human. Building the case for a thousand lines does not.

Obsolete parts behind assets still in service

Assets outlive their supply chains, and notice that a component is no longer made arrives years before the asset retires. An agent can build the case for a final purchase: where the part sits across the fleet, what holdings exist, the remaining life of the parent asset, and whether an approved alternative already exists in your own records.

The equivalent part already approved elsewhere

Interchangeability knowledge lives in a few heads. One site has approved an equivalent seal; another is still buying the original at a premium from a single stockist. An agent can surface candidates from your own approvals and specifications, as something for engineering to confirm, never as a substitution it has made.

What a technician writes, and what the catalogue already holds

A requisition that starts at a maintenance work order arrives in a technician's words. "seal kit for pump 3" is clear to whoever wrote it and matches nothing. The part usually exists in the catalogue under a description written by somebody else, so the buyer either finds it or creates a duplicate.

Agents do this well, because it is matching under ambiguity rather than judgement. An agent reads free text against your catalogue, including manufacturer references and descriptions nobody maintains, proposes the existing item and shows what it matched on. The non catalogue pile shrinks as a side effect.

Services bought against a framework

Much of the spend is services called off against rate schedules: inspection, overhaul, civils, specialist labour. The administration sits in the gap between an engineer who wrote a scope and a rate schedule the engineer has not read. Given a scope of work, an agent can identify the framework it should be called off under, map the activities onto priced items, and flag the parts of the scope that no rate covers, which is where disputes come from later.

Mobilisation, standby, travel, waiting on weather and out of hours uplift are agreed before the crew arrives or argued about afterwards, and there is no third option. An agent can list which of those a draft call off prices and which it leaves silent. Competency and site access prerequisites belong in the same check, as conditions of award rather than a scramble the week work starts.

Enquiries out, and responses that answer different scopes

An enquiry package is heavier than people outside the desk believe: the specification at the right revision, the inspection and documentation requirements, the quantities and dates you want priced, the site conditions. An agent can assemble it, check which approved suppliers are genuinely approved for the scope and asset class, and say plainly when none is.

What comes back is in each supplier's own shape. One includes testing and certification, another prices it separately. One quotes delivered to site, another to their works. One has priced the witness points, another has assumed none.

Restating that on a common basis is rule bound, high volume work an agent should carry. What matters more than the comparison is the list of assumptions the agent made to build it, because that list is usually the real set of questions for the suppliers.

A view of what the work ought to cost, built from your own comparable jobs, is worth assembling for the same reason. It is a negotiating instrument and not a fact: take it into the room as a question, because asserting it as a price is how a desk loses credibility with a supplier it needs next winter.

Long lead equipment, slots and milestones

Transformers, turbines, switchgear, cable. The commercial reality is a manufacturing slot, and the slot is as much the thing being bought as the equipment. Around it sit escalation and indexation mechanisms, staged payments against milestones, inspection and witness points.

An agent is good at the custody of that: holding every milestone and witness point against dates, assembling the evidence that a payment stage has been met, recalculating an escalation claim against the mechanism actually agreed, and raising the lead time creep that shows up in acknowledgements rather than in letters.

Supplier documents that expire between jobs

Insurances lapse between jobs, accreditations expire, and approvals are often granted for a scope narrower than the one now being awarded. An agent can keep that picture live, and chase a renewal in time: what each supplier holds, what expires when, and what is missing for the work they are about to be given.

The breakdown at night and the paperwork after

Unplanned work does not wait for a purchase order. Somebody authorises a crew or a part verbally, the plant comes back, and the paperwork is written days later from memory. An agent cannot prevent that and should not try. It can reconstruct it: tie each callout and retrospective order back to an agreement and a rate schedule where one exists, and surface the recurring emergencies that deserve a standing arrangement.

The same part at three prices across the asset base

Spend analysis across an asset base is a different question from spend analysis across a ledger, and the useful findings are not which category grew. They are that the same bearing is bought at three prices by three sites, that one contractor is engaged under three separate arrangements at different rates, that a part held as critical at one station is not held at all at its twin. An agent can produce that continuously, in the vocabulary of assets rather than of ledger codes, which is the only form in which anyone will act on it.

Buying for an asset you do not wholly own

Where the asset is shared, a buying decision has to be defensible to a partner who was not in the room, and the test is less whether it was a good decision than whether the record shows it was a reasonable one. Agents are strong at that record: who was invited, who responded, what was compared, what was assumed, why the award went where it did, built as the event runs rather than retrieved under challenge.

Where AI fails here

Some of this will not yield to a better model.

An agent cannot fix a catalogue and an asset register nobody trusts. If the same part exists four times over, or equipment numbers disagree between the maintenance system and the finance system, every output will be confidently wrong and believed for a while. That needs an owner first.

Technical equivalence is an engineering decision. An agent can find a candidate and show what it matched on, and that is where its authority ends. A substitution made on description similarity, in a safety or pressure related application, is the kind of mistake that ends a programme.

Criticality is not derivable from usage. A part never issued may be the one you cannot be without, and a system reasoning from movement history will recommend releasing the insurance spares you hold on purpose.

In a tight market the answer is a phone call. When a manufacturer is allocating slots, they go to the customers whose people are known and present. An agent can prepare for that conversation. It cannot have it.

An agent also cannot carry risk. Buying a long lead unit before the scope is confirmed, or running an asset without a spare, has consequences that belong to a person and a signature.

One category, through one outage window

Do not start with the whole desk. Take one category that matters inside one outage, such as rotating equipment spares or access and scaffolding, and pick one where you already know how the last cycle went.

Run it backwards over that cycle first. Give it the requisitions as they arrived and the responses as they came in, then compare what it produces with what your buyers did. Where it disagrees, that is the output worth having: either a gap in the agent, or reasoning a buyer never wrote down.

Then run it forward through the next window, with a buyer reviewing every proposal before it leaves. Inside one cycle you will know whether the matching holds and whether the call off checks caught something that would otherwise have been argued about after the crew left.

Adopt AI Partners builds these agents inside your own systems, against your catalogue, your frameworks and your outage history, rather than asking you to move your buying somewhere else. We stay until your buyers can see why the agent said what it said and change it themselves. To work out which part of your desk would move first, describe it to us at hello@adoptaipartners.com.