AI for Document Review in Energy: Revisions, As Builts and the Document Estate Nobody Has Read End to End

Every energy business of any age carries a document estate larger than anyone has read. Drawings, data sheets, specifications, method statements, operating manuals, commissioning records, as builts, survey and environmental reports, land agreements, consents, and the vendor data book that arrived on a disc with a piece of plant two decades ago. Some sits in a document management system, some on a project share nobody migrated, some as paper in a site office cabinet.

Document review is reading across that estate to answer a question. Which revision is the crew working from. Was the modification executed last spring ever drawn. Does the data room contain what the seller says it does. Does the plan on the deed still match the ground.

The question is almost never about one document

A practitioner rarely needs a single document, but a position supported by several. An engineering manager asking whether a pump change was captured is asking about the modification record, the as built, the current P and ID and the data sheet. A land team asked whether works can proceed in a given month is asking about a lease, a wayleave, a consent condition and a plan. The expense is the assembly: finding the candidates, deciding which are current, noticing that two disagree, and recording why. Assembly is mechanical. Judgement is not.

Revision currency, the problem underneath all the others

The same drawing exists at several revisions across the document management system, a contractor's handover folder that was never fed back, a site print marked up in pen, and a PDF emailed to a vendor.

An agent reads title blocks, revision tables, transmittal records and document registers, then builds a candidate set for one document number across every location it was pointed at. For each candidate it records the revision marking, the issue date, the source location and the evidence for both, cited to the page and field it read them from. The result is a comparison with the disagreements made loud: this copy is marked "rev 4, superseded" in its own title block while the register still lists it as the issue sheet.

Why the agent must not declare a revision current

Currency is a controlled status, not an inference. It depends on transmittal, on acceptance, and on whether the document was ever formally issued. An agent that announces "this is the current revision" has created a false record that will be relied on. One that reports three revisions, each with its evidence, and states where the register and the handover folder disagree, leaves the decision with the person accountable.

As built against as designed

The gap between what was designed and what stands on site is the oldest problem in asset documentation: the modification executed under pressure during an outage, and the drawing update that stays on a list through three reorganisations.

An agent narrows the search. Point it at the modification and change records, the work order history, the as built set and the design basis, and ask where a change is evidenced in one stream and absent from the other. It surfaces candidates: a change record naming an instrument that does not appear on the relevant drawing revision, a commissioning note describing a configuration the data sheet does not support. Each is a question for an engineer, not a finding.

Data rooms and acquisition diligence

Buying an operating asset means reviewing somebody else's filing habits under a deadline, and the hardest output is not a summary of what is there but a precise statement of what is missing, because absence prices risk.

An agent works against a structured expectation. Give it the asset register, the site boundary and the document categories your technical diligence requires, and it maps each file to a category, reports coverage, and flags where something exists but is thin: a manual with no commissioning record behind it, an environmental report whose referenced appendices are absent, a consent without its conditions document. Framed against the actual plant rather than a generic checklist, gaps arrive attached to a thing a diligence lead can escalate. It also finds near duplicates: a room that looks full is often one folder copied three times.

Land, lease, wayleave and access files

Land files age badly. Rights were granted by deeds with plans drawn to varying standards, assigned through corporate changes, varied by side letters and overlaid by later consents. The questions recur: does a right exist over this ground, what does it permit or constrain, and who must be notified.

An agent reads a bundle of agreements and produces a per parcel position: the instrument, the parties, the right as written, any restriction on timing or activity, any notice requirement, and the plan reference, each cited. Where a plan is referenced but not present, it says so. Where two instruments cover overlapping ground, it reports the overlap as unresolved. The one wayleave that constrains a works programme, a seasonal restriction buried in a schedule, is what a reader skims past at four in the afternoon.

Consents and planning conditions as a reading task

A consent is a set of obligations expressed as conditions. Reviewing it means extracting each condition as a discrete item with its wording, its trigger, and the document it says must be produced or approved. The agent returns that list with citations back to the condition text, plus a note on which conditions reference a submission the file does not appear to contain. Whether the operating reality satisfies a condition is a judgement for the people who hold it. The reading and structuring is where the hours go.

Handover and completion packs

At the end of a construction project a large body of documentation must be accepted, and presence, currency and internal consistency all have to hold across a volume no acceptance team can read line by line. In practice the pack gets sampled, and the gaps appear when somebody needs the manual.

An agent checks presence against the required document list, confirms each document carries a revision and a date, and flags documents whose content contradicts their status. More usefully, it checks consistency between documents that should agree: a tag number list against the drawings, an equipment schedule against the data sheets supplied, a test record set against the equipment it claims to cover. A pack with exceptions listed is a different conversation with a contractor than a sampled acceptance.

Submittals, technical queries and deviations

A submittal exists to answer a specification, so reviewing it means reading both. An agent lines up a submitted data sheet against the clauses it responds to, marking where the submitted value sits inside the requirement, where it sits outside, and where the specification asks for something the submittal does not address. Deviation requests get the same treatment. The engineer still reviews and signs, from a mapped comparison rather than two documents and a highlighter.

Cross document contradiction is the output that matters

The most valuable thing an agent produces here is a contradiction, stated plainly, with both sides cited. The specification says one thing, the data sheet supplied another, the as built a third. The reviewer's job is to surface the conflict and route it, not to pick a winner. Picking one without the engineering context is how a wrong number enters the record and stays.

So build the agent to prefer reporting disagreement over resolving it. Confidence is the wrong target. Coverage and citation are the right ones.

Finding things at all

Half of document review is retrieval. Legacy scans have no text layer until they are processed, drawings carry their only machine readable words in the title block, and filenames encode nothing reliable. An agent helps by building a usable index: text extraction with quality flagged per document, title block fields read into structured metadata, document and tag numbers recognised wherever they appear, and a note on every file it could not read. That last list tells you where the estate is still dark.

Decommissioning and the long record

End of life work is defended with records assembled long after the people who made them left. What was installed, what was removed, what was left in the ground, what was tested and when. Run the review while anyone who remembers is reachable, and capture the agent's unresolved questions as a list someone can still answer.

Where AI fails here

Drawings are the hardest honest limit. On a general arrangement the information is geometry: a clearance, an orientation, a routing. None of that is text, and an agent reading the words on a drawing is reading the smallest part of it. Treat drawing content as metadata and retrieval help, not comprehension.

Scans of scans defeat extraction. A faxed and rescanned data sheet produces plausible nonsense, which is worse than an unreadable page because it does not announce itself. Every extraction needs a quality signal, and low quality pages must route to a person rather than into the answer.

Where no revision history was maintained, no amount of reading reconstructs one. The agent can tell you what copies exist and what each says about itself. It cannot tell you what was issued. A confident answer about currency is more dangerous than no answer: a controller will not re-check something that looks settled, and the crew will work to it.

Absence is hard. An agent reports that it did not find a document in the locations it was given. That is not the same as the document not existing, and the difference matters in a dispute.

Finally, the agent decides nothing. It does not approve a submittal, accept a handover, confirm a right of access or declare a document current. Adopt AI Partners builds these as assembly and evidence tools with a named human decision at every point where the record changes.

Start with the records an audit will ask for

Pick the narrowest estate where being wrong is expensive and the documents are already in one place. One asset, one document class, one question. Revision reconciliation on a single plant area is a good first build because the comparison is checkable: run it over documents your controllers have already resolved and see whether it finds what they found, and what it misses.

Build the citation layer before the volume. If a reviewer cannot get from a finding to the page it came from, the finding will not be trusted and the tool will stop being used. If you want to talk through which part of your document estate is worth reading first, email hello@adoptaipartners.com.