AI for Invoice Processing in Energy

An invoice arrives for work at a site nobody in the finance office has ever visited. Stapled to it is a ticket, signed at shift end by whoever was at the gate, listing a crew, equipment, hours and a few words about the work. The person who has to pay it has a photograph of that signature, a rate schedule on another system, and a payment run closing this afternoon.

That is accounts payable in an energy business. Upstream, generation, networks and retail supply share one fact: the invoice describes work at a remote site or a volume measured by an instrument, so it can never be checked against itself.

The document is never the evidence

An ordinary supplier invoice is settled by asking whether you ordered the thing and whether it arrived. Neither question applies here. Much of this spend has no line by line purchase document, only a call off against a framework, and nothing arrives at a loading bay. The evidence sits elsewhere: a ticket signed in the field, a rate schedule attached to a master services agreement, a meter record, an interest in an asset shared with other companies.

Reconciling those sources is what an agent does: it extracts the invoice and its attachments, matches each line to the field evidence and the contracted rate, proposes a coding, attaches its evidence, and queues what it cannot resolve. It does not approve, classify capital spend, or send anything to a partner.

The field ticket and what was signed for

The ticket is signed on site when the work ends. The invoice turns up weeks later, built in the contractor's back office, and the gap between them is where leakage lives. A hand appears on the invoice who is not on the ticket. Equipment recorded as standing idle is billed as working.

An agent pairs every invoice line to the ticket it rests on, reads the handwritten quantities where it can and says so where it cannot, then reports lines with no supporting ticket, tickets with no authorised signature, and differences between what was signed for and what was billed. Nobody has ever had time to compare every line against every ticket, which is why that last category pays for the work.

Day rates, standby and mobilisation under a rate schedule nobody has read

Behind most field spend sits a rate schedule appended to a master services agreement, negotiated by supply chain and stored as a scanned appendix nobody in finance has opened. Rates change by shift, by day of the week, by public holiday and by crew. Standby and waiting time carry reduced rates that are easy to bill at the full one. Mobilisation is payable once for a campaign and gets claimed again when the crew returns. Escalation clauses tie rates to an index, and the version in force is not always the one the supplier used.

So an agent recalculates rather than accepts: the rate class, the schedule version governing the work date, whether standby was billed as working time. The output is expected charge, billed charge, difference and clause.

Rental that runs past the off hire date

Tanks, pumps, generators, matting and tool packages arrive on hire and stay on hire. Putting equipment on hire is a deliberate decision, while taking it off hire is paperwork, and paperwork after a job finishes belongs to nobody, so the rental bills weekly, correctly, against an agreement you signed. The agent maintains a live position per item: on hire date, rate, minimum term, site, the off hire notice if one was issued, and everything billed after it. Equipment still charging against a job that closed is not an invoice error. It is your own process billing you.

Which asset or cost object the charge belongs to

Every charge lands on a well, a site, a circuit, a project, a work order or a cost centre. Coding it to the wrong one is silently wrong rather than visibly wrong: the total is right, the supplier is paid, the run balances, and the only casualty is the ledger. It surfaces later as a unit cost that makes no sense, a project over its authorised spend, or a rebill somebody disputes.

An agent codes from evidence rather than habit: the work order reference, the ticket, the site named in the free text, the purchase document where one exists, and the precedent of earlier invoices on that job, which also carries forward whatever has been miscoded for years. Where the description names one location and the purchase document another, it reports both rather than choosing the likelier.

Capital or operating, decided at the wrong desk

A single site intervention invoice can contain both a replacement that extends the life of an asset and a repair that merely restores it. The split drives depreciation, tax, and in a regulated business what enters the asset base your charges are set against. In practice it is decided by whoever is clearing the queue, from a description written by a field engineer with no interest in accounting. Wrong desk, wrong moment. An agent should propose a split, show the words that led to it, and stop. If the policy is not written anywhere an agent can read, writing it down comes first.

Joint interest billing, where a coding error becomes an external dispute

Here the stakes change. Operating an asset on behalf of co venturers, you incur cost and rebill a share to other companies under a joint operating agreement. An error no longer stays in your ledger. It leaves the building on a statement, is booked by companies who never saw the cost incurred, and returns as a dispute after the period closed.

The checks are specific. Is this cost chargeable to the joint account at all, or an overhead the agreement excludes? Which property does it belong to, when neighbouring properties share suppliers, crews and equipment in the same week? What was each party's working interest on the date the cost was incurred, rather than today, given interests move when assets are traded? An agent assembles that per line and presents the billing with its evidence attached, then runs the same checks in reverse on statements you receive as non operator.

The claim that arrives after the period closed

Partner audits look backwards, often across years, and arrive as a schedule of queried lines. The work is retrieval: the ticket, the rate in force, the coding rationale, the authorisation, the clause that made the cost chargeable. That is a poor use of a senior accountant and a good use of an agent, which builds a pack per queried line. Nothing rebilled to a partner leaves on an agent's authority. A person signs.

Charges that are arithmetic on data you already hold

A large slice of energy spend is not a description of work at all. It is a calculation over measured quantities and a published tariff: commodity volumes, network and transport charges, capacity and demand charges, imbalance amounts from the market settlement process. These are the easiest invoices to check and the least checked, because checking means rebuilding the sum. You hold the same measurement, or your own version of it. An agent recalculates from your records and the tariff in force, then reports the difference and its cause: an estimated reading where an actual exists, a wrong multiplier, a tariff applied from the wrong date.

Accruing for work done and not yet billed

At period end, most of the work performed in an energy business has not been invoiced. Contractor back offices run weeks behind. The accrual therefore carries much of the reported cost, and it is often built from memory or from last period plus a margin.

An agent already holds the raw material: open tickets with no invoice, rentals still on hire, milestones recorded as achieved, authorised spend with commitments outstanding. It can propose a schedule line by line with the evidence behind each entry, a better conversation with your auditors than one judgemental figure, and match late invoices back to what was accrued so the cost is not recognised twice.

Lump sum milestones and the invoice that arrives twice

Major project work bills differently: no hours and no tickets, a payment falling due when a defined event happens. An agent reads the payment schedule out of the contract, tracks each milestone and tests the claim against the deliverables named. Everything here also arrives more than once: a portal, a shared inbox, paper scanned at a site office. A resubmission after a query looks like a duplicate, while a real duplicate often carries a different number and date, so the only reliable test is the underlying work: the ticket reference, the site, the dates worked.

Back testing on invoices your own team has already settled

Take a period you have closed and audited, where the answers are settled, and run the agent across it from scratch. Some disagreements will be the agent being wrong, some things your team missed.

Measure the components separately rather than as one accuracy number. Extraction from typed invoices and from photographed handwritten tickets are different figures and should never be averaged. Coding accuracy has to be scored against where the cost finally sat after year end reclassification, not what payables first keyed, or you reward it for reproducing old errors. Then shadow run before authority widens.

Where AI fails here

Handwriting defeats it more often than demonstrations suggest. Tickets are written outdoors, folded into a pocket and photographed badly, and quantities and signatures are often the least legible fields on the page, which are precisely the ones that matter.

A rate schedule the agent cannot read properly produces confident nonsense. If the appendix is a scan of a scan with pen annotations, or the rate everyone uses was agreed in an email and never folded into the contract, the agent recalculates against the wrong basis with the tidy certainty of a correct result.

It cannot tell you whether the work happened. An agent compares documents to other documents. If a ticket was signed without being read, it confirms that a correct invoice matches a wrong ticket, and nothing at this desk will catch that. Verification of physical work sits with the field.

Interests, agreements and rates change before anyone documents the change. A rate revised verbally, or an assignment known to the commercial team for weeks, makes the agent wrong on every affected line until someone tells it.

If your payables team is checking the large invoices and trusting the rest, describe the process to Adopt AI Partners at hello@adoptaipartners.com and we will tell you plainly which parts an agent could take and which it could not.