AI for compliance evidence in utilities: assembling the proof before the auditor asks for it
Every network and supplier we speak to can tell you what they are obliged to do. The licence conditions are written down. The permit terms are written down. The reporting calendar is on a wall somewhere. What nobody can tell you quickly is where the evidence is.
That is the actual problem. Not the obligation. The proof.
The evidence exists, it is just not joined up
Consider a statutory inspection cycle on a class of assets. To prove each asset was inspected at the required frequency you need the asset register, the work orders, the completion records, the engineer's notes describing what was found, any follow-on defect jobs, and the dates on all of it.
The asset register lives in one system. The work orders live in the work management system, which uses identifiers that only partly match the register because of a migration eight years ago. Completion is a status flag plus a free text note typed on a handheld in the rain. The photographs are in a mobile app whose export nobody has run since the last audit. Three sites were done by a contractor who reports in a monthly spreadsheet emailed to a shared inbox.
Now do the same for environmental permit conditions, evidenced by laboratory results with sample reference numbers, flow and pressure records pulled from SCADA exports, and a consent document whose terms were summarised into a spreadsheet by someone who has since left. Or metering and billing accuracy, evidenced by meter test certificates, exchange records, exception reports and the billing system's audit log. Or the priority services register, where what was offered to a vulnerable customer sits partly in the CRM, partly in contact centre notes and partly in a field visit record.
None of this is disorganised out of carelessness. Each system is well run for its own purpose. The gap is that none of them was built to answer the question an auditor asks, which is always some version of: show me, for this obligation, across this period, the records that prove it, and show me the set is complete.
What a submission actually costs you
The visible cost of a periodic submission is the reporting team's quarter. The invisible costs are larger.
There is the reconciliation cost. A number goes into a submission. Six weeks later somebody asks how it was derived, and the person who derived it is reconstructing a query, a filter and three manual adjustments from memory and a working spreadsheet. If it does not come back clean, the conversation moves from a data question to a confidence question, which is expensive.
There is the assurance cost. Every internal audit and external assurance sample lands on an engineering or operations team who have to stop and go and find something. The requests are individually small and collectively enormous.
There is the gap cost, which is the worst one. An inspection that slipped outside its window, a permit condition whose monitoring lapsed for a month, an incident reporting deadline met in operational reality but not evidenced in writing. Discovered internally, these are manageable findings with a remediation plan. Discovered by an auditor or a regulator, they are a different category of problem, because now the question is not the gap but whether you knew about it.
Where an AI system genuinely helps
The useful framing is not compliance automation. It is evidence retrieval and evidence assembly, done continuously, with a citation on every claim.
Assembling the evidence for a named obligation
The system holds a structured description of each obligation: what must happen, to what population, at what frequency, and what would constitute proof. Given one of those, it finds the candidate evidence across the work management system, the asset register, SCADA exports, laboratory results, contractor spreadsheets and the email attachments never filed anywhere better. It returns an evidence set where every item points at the exact record it came from, with the system, the identifier and the timestamp. An assurance lead can click from assertion to source in one step, which is the only output worth anything in an audit.
Reconciling a reported figure back to source
Given a figure in a submission, the system reproduces the population behind it and shows the records that make it up, including the exclusions and the reason for each. When the figure does not reconcile, it says so and shows the difference rather than presenting a plausible total. Reconciliation stops being archaeology and becomes a query.
Spotting gaps in a cycle before an auditor does
This is the highest value use and the least glamorous. Run the inspection population against the obligation frequency continuously. Surface every asset approaching or past its window, every completion record with no supporting note, every work order closed with a status that does not actually evidence the work, every register entry with no maintenance history, every identifier mismatch that means an asset might be compliant in one system and invisible in another. The output is a worklist with a named owner, months before anyone external asks.
Reading field notes into structured completion evidence
A field engineer writes in shorthand, recording what was checked, what was found, what was left in service and what was referred on, in language shaped by the job rather than the reporting schema. A language model is genuinely good at reading that note and proposing structured fields: asset, activity performed, condition observed, defect raised or not, follow-up required. Proposing, with the original text kept alongside and a confidence value attached. The structured field is a searchable index into the note, never a replacement for it.
Drafting the narrative sections from the underlying records
Submissions contain prose: explanations of performance movement, descriptions of remedial activity, commentary on exceptions. Those paragraphs are written by the few people who understand both the numbers and the regulatory register, usually late and at the cost of everything else on their desk. A system that drafts from the underlying records, with every factual claim cited back to a source record, gives them an argued first draft to correct rather than a blank page. They still own every word that goes out.
The evidence pack that is always current
The pattern worth aiming at is a continuously maintained evidence pack rather than a quarterly assembly project. The obligation register, the evidence mapping and the gap list all stay live. When a work order closes, the evidence set for the relevant obligation updates. An identifier mismatch surfaces the week it appears.
The organisational effect matters more than the technical one. Preparation stops being a scramble, because the pack is already there. Assurance sampling stops interrupting engineering teams, because the sample can be served from the evidence layer with citations attached. And the compliance lead can answer the question they currently cannot: as of today, which obligations do we have complete evidence for, and which do we not.
Where AI must not decide alone
Some boundaries here are not tunable.
No AI system makes a regulatory attestation. A submission is signed by a person with accountability, and that signature must rest on their own judgement.
No AI system judges whether an obligation is met. It assembles evidence and flags where evidence is thin, absent or contradictory. The determination is a human judgement informed by the register, by precedent and by regulatory relationship, and it belongs to the accountable owner.
No AI system makes a safety determination. Not asset fitness, not network risk, not whether an anomaly is a hazard. Where compliance evidence touches safety, the system's job is to put the right records in front of the right qualified engineer faster, then stop.
And nothing the system does may be unobservable. Every output carries its sources, every extraction keeps the original text, every automated step is logged with inputs, outputs and version. Automation without an audit trail is automation you will not be allowed to keep, which in a regulated utility is the difference between a system that survives its first assurance review and one that gets switched off.
Where AI fails here
It fabricates coverage. Asked to produce evidence for an obligation, a language model will produce something that looks like an evidence pack whether or not the evidence exists. This is the dangerous failure, because a confident empty pack is worse than a blank page: it creates a false assurance signal. The mitigation is architectural. The system must cite only records it can resolve, and must report absence as a finding rather than smoothing over it.
It cannot fix identifier chaos. If the same asset has three identifiers across four systems and no reliable crosswalk, the system will mis-join records. It can surface candidate mismatches for a data owner to resolve, but it does not substitute for the reconciliation work. Anyone promising that AI removes the need for a clean asset spine is selling you something.
It reads ambiguity as fact. "Checked, ok, will monitor" could mean the inspection was completed to standard or that the engineer deferred. A model will pick one. The only safe design routes low confidence extractions to a human rather than committing them.
It degrades quietly when the sources change. A new work order type, a changed status code, a contractor switching spreadsheet format. Accuracy drops and nothing announces it. Measurement therefore has to be continuous rather than a launch gate, with a held out set of your own hard cases re-run on a schedule.
It is only as good as the obligation register. If nobody has written down what each obligation requires and what would prove it, there is nothing to retrieve against. That mapping work is human, unavoidable, and usually the real project.
It does not survive a governance vacuum. A flagged gap with no named owner becomes noise, and people stop reading the output. Routing matters as much as retrieval.
How we would approach it
Adopt AI Partners builds this inside your environment, against your systems, with your data staying where it is, and vendor neutral because the right components depend on what you already run.
We would start with one obligation you already find painful to evidence, build the retrieval and citation layer for it, and measure accuracy on your own hard cases before anything goes live. Not a demo set. The awkward assets, the migrated identifiers, the contractor records, the notes nobody can parse. If it does not hold up on those, it does not launch. The system stops and routes to a named owner whenever confidence is low, and your people own and operate it afterwards.
If you already know the shape of your evidence problem, hello@adoptaipartners.com reaches us.