How to choose AI solutions for utilities

Where does AI pay off first in a utility, and why?

AI pays off first where skilled people spend their day reading and cross-checking: a complaint against the bill, the meter reads and the call notes; a held bill against the tariff; a meter alarm against the network; a condition report against the asset register. The rules are written down, the evidence already sits in your systems, and the backlog is visible. That makes the gain easy to measure and the risk easy to contain.

The pressure is rising on both sides of the business. The IEA's World Energy Investment 2025 report puts grid spending at about USD 400 billion a year worldwide in 2025, against around USD 1 trillion on generation, and says grid spending has to move towards parity. More assets, more connections and more data arrive faster than teams can be hired, so the work that grows fastest is the work of reading, checking and deciding.

• Customer care and billing: complaints, customer email, held bills, high-bill enquiries, disputes
• Metering: VEE exceptions, meter alarms, theft and tamper scoring, leak alerts on water meters
• Assets and field work: inspection findings, drone images, condition reports, work-order close-out
• Water operations: water balance and non-revenue water, quality results against limits, service line records
• Regulatory, security and HSE: data requests, licence reports, audit evidence, incident investigations

Which processes should we start with?

Start with one process that has a regulatory clock, a measurable backlog and a person who already signs the outcome. In Great Britain, Ofgem says a domestic energy customer can take a complaint to the Energy Ombudsman once it has been open eight weeks, or sooner if the supplier issues a deadlock letter, and the Gas and Electricity (Consumer Complaints Handling Standards) Regulations 2008 require suppliers to tell customers of that right. In September 2024 OVO agreed to pay £2.37 million after an Ofgem investigation, including £378,512 in compensation to 1,395 customers whose complaints were in some cases delayed by up to 18 months, which shows what a missed clock costs.

Good first candidates share three traits: the inputs are documents or records you already hold, the correct answer can be checked by a person in minutes, and the volume is high enough that a few weeks of running produces a real number. Complaint investigation, billing exceptions, VEE exceptions and inspection finding triage usually qualify. Leave anything that switches the network or changes a protection setting for later, and keep a person in charge of it.

What should we demand on accuracy, citations and human sign-off?

Ask for every extracted value and every drafted sentence to point to its source: the meter read, the bill line, the clause of the tariff, the page of the inspection report. If a reviewer cannot click from a claim to the passage it came from, the reviewer has to redo the work, and the time saving disappears. On Botminds, outputs are cited to their source, so every number is traceable to the page it came from.

Decide in writing which decisions a person must sign, and make sure the system enforces it rather than relying on habit. Customer letters, redress amounts, disconnection decisions, switching orders and regulator submissions should never leave unsigned. Ask how accuracy is measured before go-live and after it, on your own documents, and what happens when a result falls below the bar.

• Field-level citations to the source document or record
• Named approval steps for every consequential decision, recorded with who signed and when
• Accuracy measured on your own sample before go-live and monitored after
• A clear route for low-confidence cases to a person, never a silent guess

What about security, deployment and our data?

Utilities run critical infrastructure, so the questions are stricter than in most industries. Ask where the data lives, who can see it, whether it trains anyone's model and whether every action is logged. Botminds runs as multi-tenant SaaS, in a dedicated cloud environment, or fully on-premises, with the same controls in each; your inputs and outputs are not used to train models; data is encrypted with TLS in transit and AES-256 at rest; access is controlled by role through your own sign-in; and every action, by a person or an agent, lands in an audit log that cannot be altered. Botminds holds SOC 2 Type II and ISO 27001, and its controls are designed for GDPR.

Regulation is moving too. Annex III of the EU AI Act lists as high-risk the AI systems used as safety components in the management and operation of the supply of water, gas, heating or electricity. An amendment to the Act, published in July 2026 and in force since 27 July 2026, moves the date those high-risk obligations apply from 2 August 2026 to 2 December 2027. Keep work that touches operational safety under human control, and ask any vendor how its records would support a conformity assessment. North American utilities should also ask how a solution fits NERC CIP: CIP-007-6, for example, requires security patches to be evaluated at least once every 35 calendar days.

How do we run a pilot and measure it?

Run the pilot on your own documents and your own backlog, not on a vendor's demo set. Write down the baseline first: cases a day per person, time to close, share answered inside the regulatory clock, rework rate. Then run the new solution alongside the current process for a few weeks, with your own staff reviewing every output.

Measure two things separately: how often the drafted result is right without correction, and how much time a reviewer spends per case. A pilot that cannot fail is not a test, so agree in advance what number means go, and what means stop.

• Pick one process, one team and one clear owner
• Record the baseline before the pilot starts
• Review every output during the pilot, and log every correction
• Agree the go and stop thresholds before the first case runs

Should we build, buy a point product, or use a platform?

Building in-house gives control but ties up scarce engineers on document reading, permissions, audit and monitoring before any utility logic is written. Point products are quick for one task, but a utility needs dozens of them, from complaints to transformer oil analysis to licence reports, and each one brings its own security review, contract and data copy.

A platform lets one set of controls, one sign-in and one audit trail serve every process. Botminds is an agentic solution engine: it turns your requirements into a governed, working solution and deploys it, and its utilities library covers customer care, billing, metering, grid, assets, field service, water, generation and desalination, connections, regulatory affairs, HSE, supply chain, workforce and IT/OT security. Ask any vendor how the second and the tenth solution will be delivered, not just the first.

What changes for the team?

The job moves from gathering evidence to judging it. A complaints handler stops pulling bills, reads and call notes into a spreadsheet and starts checking a drafted, cited investigation and signing the letter. A planner stops chasing missing close-out codes and starts deciding which jobs go back to the field.

Plan for the change openly. Name who signs what, train reviewers on how to check a citation quickly, and keep the people who know the edge cases close to the pilot. Many utilities also face the loss of experienced staff to retirement, so capturing how experts decide, with citations to the procedures they rely on, is part of the value rather than a side effect.

What does it cost to wait?

Deadlines do not move because a team is busy. In the US, EPA's Lead and Copper Rule Improvements, finalised in October 2024, require water systems to identify and replace lead service lines within 10 years, which starts with classifying service line materials from old records such as tap cards and permits. NERC CIP evidence, licence reports and complaint clocks recur every month whether or not the backlog is cleared.

Every month without help, the backlog, the redress risk and the overtime grow, and the people who know how to clear it get closer to retirement. Starting with one measured process now builds the evidence you need to decide the next one.

Solutions to look at first

• Utility Complaint ManagementA workflow solution for complaint handlers that investigates and resolves complaints on time.
• Customer Email AutomationA workflow solution for back-office teams that reads, sorts and answers customer email.
• Contact Center Agent AssistA search solution for contact centre agents that answers tariff and policy questions in-call.
• Billing Exception ManagementA workflow solution for billing analysts that clears held bills with an evidenced correction.
• High Bill InvestigationA workflow solution for billing teams that explains each high-bill complaint from meter data.
• VEE Exception ManagementA workflow solution for MDM analysts that triages VEE exceptions and proposes the fix.
• Energy Theft DetectionA system solution for revenue protection teams that scores meters for theft and tampering.
• Drone Inspection AnalyticsA workflow solution for line inspectors that screens drone images and flags asset defects.
• Asset Health IndexA system solution for asset managers that scores every asset's condition and failure risk.
• Maintenance Knowledge SearchA search solution for maintenance engineers that answers from manuals, SOPs and past fixes.
• Non-Revenue Water ManagementA system solution for water loss managers that builds the water balance and NRW KPIs by zone.
• Lead Service Line InventoryA workflow solution for water utilities that classifies service line materials from records.
• DER Interconnection Application ReviewA workflow solution for interconnection engineers that checks solar and storage applications.
• NERC CIP Audit Evidence ManagementA workflow solution for CIP teams that assembles and checks audit evidence by requirement.

Every utilities solution →

A checklist for any vendor

1. Can every value and sentence in an output be traced to its source document or record?A good answer shows a click from a drafted claim to the exact meter read, bill line, tariff clause or report page it came from.
2. Which decisions require a named person to sign, and how is that enforced?Sign-off is a step the system will not skip, recorded with the person and the time, not a policy people are asked to follow.
3. How will you measure accuracy on our documents before go-live, and after?A test on a sample of your own cases with an agreed bar, then monitoring in production so drift is caught early.
4. Where does our data live, and can you run in our own cloud or on our premises?A choice of multi-tenant SaaS, a dedicated cloud environment or fully on-premises, with the same controls in each.
5. Are our inputs and outputs used to train any model?A plain no, written into the contract.
6. What security certifications do you hold, and can we see the reports?Current independent audits such as SOC 2 Type II and ISO 27001, with reports available on request.
7. Is every action, by a person or an agent, logged in a record that cannot be altered?An immutable audit trail you can export for a regulator, an internal auditor or an ombudsman.
8. How does the solution work with our CIS, MDM, GIS, asset and work management systems?It reads from and writes to the systems you already run, without a parallel copy that drifts.
9. How will the second and tenth solution be delivered, and on what controls?The same sign-in, permissions, audit and monitoring serve every solution, so each new one does not restart the security review.
10. What do you need from us for a pilot, and what number would make you recommend stopping?A short list of documents and people, a written baseline, and a stop threshold agreed before the pilot starts.

Questions