AI for Contract Review in Energy: Offtake, Construction and the Obligations That Follow

A reviewer opens the fourth draft of an offtake agreement late in the evening. The pricing schedule has moved again. The floor is now expressed by reference to a different index, the indexation date has shifted by a quarter, and somewhere in the definitions the meaning of available capacity has been narrowed in a way that nobody mentioned in the covering email. The commercial team wants a view by the morning because the investment paper is due. The reviewer also knows, without being able to prove it before morning, that the same counterparty group signed something with a different business unit two years ago, on fallback wording that was argued over hard and would answer half of tonight's questions.

That is the shape of the problem. Not a shortage of skill, but a shortage of hours against documents that are long, repetitive in form and unforgiving in the detail. A well scoped agent reads every page, every schedule and every side letter, and puts the differences in front of a human who can then spend their time on the two clauses that matter.

Why energy agreements do not behave like ordinary commercial paper

Most commercial contracts are short lived and cheap to exit. Energy agreements are neither. Terms run for fifteen, twenty or thirty years. Capital is committed before the ink dries. The counterparty will still be there in two decades, which changes how hard you push.

The other difference is that value hides in clauses that do nothing in a normal year. An indexation floor, a curtailment compensation mechanism, a change in law definition: each sits quiet until the year when it is the only thing standing between a project and a write down. Review effort should follow that asymmetry and usually does not, because the reviewer is working to a deadline and reads front to back.

Offtake and power purchase agreements

Pricing and indexation

The mechanics are where the money is. Whether escalation is capped, collared or uncapped. Which index applies and what happens when that index is discontinued or rebased. Whether a difference payment structure pays out symmetrically in negative price periods or leaves the generator carrying them. Whether imbalance and shaping costs are the seller's, the buyer's, or shared above a threshold. An agent can pull the pricing stack out of every agreement into one comparable view, including the parts buried in schedules that never reach the summary memo.

Volume, availability and curtailment

Minimum delivery obligations, annual quantity bands, make up and carry forward rights, and the deemed generation payment that decides whether a curtailed hour is compensated. The distinction that matters is who instructed the curtailment. Network instructed, buyer instructed and seller elected curtailment are often treated differently in the same agreement, and the ranking between them is frequently the least clearly drafted part of the document.

Change in law and who carries a bad outcome

Change in law clauses vary enormously in scope: whether tax is included, whether a change foreseeable at signature is excluded from relief, whether the remedy is a price adjustment, a sharing of cost, or a right to terminate, and whether only discriminatory changes count. Two agreements on the same site can allocate the same regulatory event in opposite directions.

Construction and engineering contracts for plant and network assets

Liquidated damages and performance guarantees

Delay damages, performance damages, the caps on each, whether they are a sole remedy, and how they interact with extension of time entitlement. Performance guarantee testing is its own trap: the guarantee means nothing without the reference conditions, the measurement method, the tolerance and the degradation curve that applies in later years. An agent that extracts those into one table across a build programme gives the engineers something to argue about.

Taking over, reliability testing and interface risk

Taking over usually starts the defects liability period, releases part of the security and shifts operating risk. Its conditions, and the reliability run that precedes it, deserve reading rather than assuming. Where the works are split across main equipment supply, balance of plant and connection works, the interface gaps are where claims are born. Ask which contract defines the interface points, and whether any obligation appears in no contract at all.

Long term service agreements on major equipment

These often outlive the people who signed them. The availability guarantee is the centre: what counts as available, which hours are excluded, how scheduled outages are treated, and whether bonus and damages are symmetric. Behind it sit the operating obligations, such as overhaul intervals expressed in operating hours or starts, and what happens when the plant is run differently from the agreement's assumption. Then the parts and spares commitments: whether the supplier guarantees lead times, holds a spares position, and what follows when a part does not arrive. The cross check that matters is against the offtake, because an availability guarantee weaker than the availability obligation you owe downstream is a gap somebody is carrying.

Fuel, feedstock and commodity supply

Take or pay obligations, with the carry forward and make up rights that soften them, and the nomination windows that decide how much operational flexibility you really have. Delivery flexibility is worth reading alongside the quality specification, because a narrow spec with a tight nomination window leaves little room in a bad month. On off specification delivery: whether you may reject or must accept at an adjusted price, whether the supplier's liability covers consequential damage to plant, and whether the cap sits below the cost of the repair such a delivery would actually cause.

Connection and network use agreements

These are short, standard in form, and mostly ignored after signature, which is why they cause trouble. They impose obligations on operations rather than on legal: capacity limits, notification of planned outages, compliance with operator instructions, testing and data submission by fixed dates, and charge pass through mechanics. Nobody in the legal team will be the person who misses a submission deadline, so extracting those obligations into something an operations manager can hold beats any elegant summary of the clause.

Force majeure and regulatory risk across the portfolio

Read one agreement at a time, these clauses look acceptable. Read as a portfolio, they show the real exposure. A single change in law or extended outage event will trigger relief under some agreements, no relief under others, and an obligation to keep paying under a third. The net position is what the business cares about and is almost never written down. Answering it means reading every agreement, which suits an agent and does not suit a person.

Credit support and what a downgrade sets off

Parent company guarantees with net worth tests, bonds with expiry dates and evergreen renewal wording, letters of credit with rating triggers. The obligations that follow a downgrade are time limited and specific: post additional support within a set number of business days, or be in default. When a group rating moves, somebody has to know within hours which agreements just created an obligation and how much aggregate support is required.

Departures from the form, listed before they are accepted

A great deal of energy contracting starts from a familiar industry form and then departs from it. The departures are where the risk moved: a deleted sub clause, a definition quietly broadened, the word reasonable removed from an approval right, a carve out added to the liability cap, a notice period cut by two thirds. Comparing the draft against the unamended baseline and listing every substantive departure is mechanical, tedious, and the highest value thing an agent can do on this paper. Which departures to accept stays entirely with the reviewer.

Obligations and notices after signature

Signature is the halfway point. Each agreement creates things somebody must do, by a date, with evidence: reports, insurance renewal confirmations, consent requests, notices in a prescribed form to a prescribed address, condition subsequent deadlines. An agent can turn an executed agreement into a draft obligation register with owner, trigger, deadline and required evidence against each line, for a human to confirm and hand to the operating team. The failures that end up in dispute are rarely failures of drafting. They are notices sent late, in the wrong form, or to the wrong person.

One regulatory change, read across every agreement at once

Which agreements contain a most favoured nation clause. Which need counterparty consent on a change of control, and so sit in the critical path of a disposal or refinancing. Which are exposed if one specific regulatory regime changes. Which renew or reach a price review in the next eighteen months. Each is a week of reading for a person and an afternoon for an agent across a full document set, which changes what the legal team can offer the business rather than just how fast it closes a file.

Where AI fails here

It does not give advice. It cannot decide risk appetite, it has no view on the relationship, and it does not know what was traded away in the negotiation that produced the clause.

It struggles with the parts of these documents that are not prose: rate tables, performance curves, appendices that arrived as scans, and defined terms whose meaning lives three cross references away. Schedules and side letters in separate files will be missed unless someone assembles the complete instrument first. Date arithmetic is a known weak point, so treat every computed date as a draft.

Most importantly, a wrong answer arrives in the same confident register as a right one. That is why output belongs as extracted text with a pointer to where it came from, so the reviewer verifies rather than trusts. The advice, the sign off and the responsibility stay with a qualified lawyer.

One agreement type, one desk, first

Pick the agreement type you review most and one desk that handles it. Long term service agreements and amended construction forms are good first choices because the review is repetitive and the departures are findable.

Then prove it backwards. Take reviews the team has already completed, including the ones that caused an internal argument, and replay the agent against them without showing it the answer. Count three things: what it found that the human missed, what the human found that it missed, and what it asserted that was not in the document. The third number decides whether this is usable. Keep a record of every override, because the pattern in them tells you where the scope should stop. When the false assertion rate is close to zero on paper you know well, widen by one agreement type, and not before.