AI Energy Services Need a Data Permission Design Before They Need an Agent

AI Trust & Governance

16 October 2026 | By Ashley Marshall

A smarter energy recommendation is useless if nobody can show who authorised the data, where it came from or whether it saved money. This week’s UK Smart Data developments put those questions first.

The new opportunity is permissioned energy data, not another chatbot

The Department for Energy Security and Net Zero opened a consultation on an energy smart data legislative framework on 14 October 2026. It proposes that relevant data holders make specified datasets available securely to authorised services when consumers choose to share them. The consultation covers England, Scotland and Wales and closes at 11:59pm on 8 January 2027. These are proposals, not a live access entitlement. That distinction matters if you are planning an AI energy product or buying one. A pitch that assumes every supplier must already provide a neat stream of customer data is selling tomorrow as though it exists today. Read the consultation.

There is a real business opportunity nonetheless. A service with reliable, permissioned usage data could spot tariff mismatch, forecast demand, explain unusual consumption or suggest when equipment should run. For a small manufacturer, hotel group or multi-site retailer, those recommendations could be more valuable than a generic AI assistant that can only read uploaded bills. The value comes from fresh, structured data and an action the customer can take. The model is the last mile, not the foundation.

The government is asking about datasets, participants, consumer protections, accreditation, governance and funding. That is a buyer checklist in disguise. Before evaluating a vendor demo, ask which dataset it needs, who holds it, how frequently it updates and whether the provider has a legitimate route to receive it. Ask whether a recommendation can be reproduced from the underlying readings and tariff assumptions. If the answer depends on a manually exported spreadsheet today, treat the service as an assisted analysis pilot, not an always-on optimiser.

This is not just for energy technology firms. Any UK business buying AI-led energy advice should distinguish three things: access to the data, permission to use it for a stated purpose and authority to act on the recommendation. Collapsing those into one broad tick box invites disputes when an agent starts making choices that the customer only expected it to explain.

Design the permission before you design the agent

The ICO published guidance for Smart Data scheme designers on 9 October 2026. It describes Smart Data as secure, consented sharing of customer or business data with authorised third parties. It also says scheme design can support compliance but does not transfer a participating organisation’s own UK GDPR responsibilities. A controller remains responsible for its processing decisions; a processor has direct obligations and must follow documented instructions. The ICO guidance is addressed to scheme designers, so buyers should use it as a design signal, not pretend it is a product approval stamp.

For an AI energy service, the permission record should say who is granting access, which premises or account it covers, which data fields can be retrieved, the permitted purpose, the duration and how to withdraw. A customer might accept half-hourly usage analysis to compare tariffs while rejecting use of the same data to train a general-purpose model. Those are different purposes. A vendor should not quietly convert a recommendation service into a data acquisition pipeline. Put the boundary in the contract and the technical access layer.

In practice, make an access screen understandable without legal translation. Instead of “allow data access”, say “share electricity usage for Site A for six months to identify avoidable peak charges”. Show the named recipient, whether any onward sharing occurs and the effect of revocation. Store the permission event and the version of the notice shown. Then test withdrawal: does the API token stop working, are downstream jobs stopped, and does the customer receive a clear explanation of what historic records are retained under the lawful basis?

The misconception is that “consented sharing” means every downstream processing activity automatically uses UK GDPR consent as its lawful basis. It does not. The customer authorisation mechanism within a scheme and the lawful basis for personal-data processing require separate analysis. In business accounts, consumption data may still reveal individual behaviour, especially at small premises or home-working sites. A privacy assessment should identify that possibility before the first pilot, not after a customer complains.

Make the recommendation auditable at the level of a single decision

An attractive AI demo will produce an answer such as “switch this site to a different tariff” or “move this load to 2am”. An operational product must show the evidence underneath. Record the source dataset, retrieval time, account and site identifier, data gaps, tariff version, forecast assumptions, model or rules version, and any human override. If the supplier later corrects a reading, you need to know which recommendation was affected. Otherwise the service looks confident precisely when it is least explainable.

The ICO says its risk-based approach to Smart Data focuses on lawfulness, fairness and transparency, meaningful control and rights, security and accountability, and protection from harm. That is not a demand for a verbose explanation from a language model. It is a demand that an organisation can demonstrate how a decision was reached and respond when the customer challenges the inputs. A simple calculation trail may be more useful than a polished narrative. See the ICO’s stated regulatory approach.

What this means in practice: require every recommendation to carry an evidence bundle and a confidence boundary. “Estimated annual saving: £4,800” needs the comparison period, tariff assumptions, excluded charges and expected implementation cost. The figure here is an illustration, not a market average. If the service cannot calculate a credible range because data are missing, it should say so and ask for the missing input. A system that fills gaps with plausible-sounding guesses is not doing financial analysis.

Keep action rights narrower than insight rights. Reading a meter stream, producing a report, sending a supplier enquiry and changing a contract are progressively different capabilities. A sensible initial deployment allows the first two and requires an authorised person to approve the latter two. Capture that approval against the same evidence bundle. This also gives finance and operations teams a shared record when claimed savings are later measured against real bills.

Use the consultation to ask better procurement questions now

The energy consultation is 78 pages and explicitly seeks evidence on datasets, accreditation, governance and consumer protection. A buyer need not wait for final regulations to improve procurement. Create a two-column test: what can the service do with legitimately available data today, and what would become possible if a future scheme grants standardised access? Vendors should mark assumptions rather than blend the two. The consultation’s January 2027 closing date is a planning milestone, not a go-live promise. The consultation page sets out the scope and response route.

Ask five concrete questions in demonstrations. Which organisation is the data holder? Who is the authorised recipient? What is the smallest dataset that supports the use case? What happens when a customer revokes access? How will the vendor prove a recommendation’s financial benefit without cherry-picking the comparison period? These questions work whether the supplier uses rules, machine learning or a language model. They expose weaknesses in the operating model rather than rewarding a polished interface.

For a pilot, choose one site and one measurable decision, such as identifying avoidable peak usage. Agree the baseline before the service sees the evaluation period. Include energy price changes, weather or production volume where they materially affect the result. Keep the pilot in advice mode until the false-positive cost is understood. If an incorrect suggestion could interrupt production or breach a supply contract, include that loss in the decision threshold. The best pilot may produce fewer recommendations but better accepted ones.

There is a sensible counterargument: why create all this structure when a human energy consultant can review bills and deliver advice now? For a handful of accounts, that may be the right answer. Structured, permissioned data becomes more valuable when the same decision recurs across many sites or when timely changes matter. Do not automate a low-frequency decision merely because an API may eventually exist. Prove that frequency, value and data quality justify the operating burden.

Separate data portability from continuous data access

The government’s August 2026 publication on data intermediaries explicitly discusses personalised AI services that combine information from multiple sources. It also describes legal ambiguity around third-party delegation of certain UK GDPR rights, including data portability. That makes an important distinction visible: asking for a one-off portable export is not the same as operating a continuous, standardised feed for an AI service. Read the government consultation.

The ICO guidance explains that Smart Data schemes may require continuous provision in real time, a specified electronic format, wider categories of information or business data. By contrast, the UK GDPR portability right is handled case by case and the ICO notes that organisations can take up to a calendar month to comply. A product promising live optimisation cannot quietly assume a portability request supplies the same capability. Nor should it build a workaround that asks users to share passwords or scrape a portal without evaluating security and contractual implications. The ICO’s comparison is worth putting in the product requirements.

What this means in practice is a phased architecture. In phase one, accept a documented export or an existing authorised integration and disclose its refresh rate. In phase two, add standard APIs if and when an applicable scheme and participant authorisation make them available. Keep the data model independent of any single supplier format, but do not claim interoperability that has not been tested. Include provenance and timestamps in every record so analysts can distinguish missing readings from genuine zero usage.

The AI layer should degrade gracefully. A monthly export may support bill analysis but not real-time load shifting. Make that product limitation explicit in the proposal and user interface. Businesses buying the service then know whether they are purchasing a monthly advisory report, daily monitoring or operational control. Those are different contracts, different risk profiles and different prices. Precision about the data channel protects both the buyer and the supplier from inflated expectations.

The decision for UK leaders: buy the data operating model

The interesting question is not whether AI can find patterns in energy readings. It can. The hard question is whether your organisation can receive the right readings for the right purpose, explain a recommendation, respect withdrawal and measure a real saving. The October developments give leaders a practical moment to review those fundamentals: a new energy framework consultation from DESNZ and fresh ICO guidance on Smart Data scheme design. Neither automatically certifies a vendor or removes the buyer’s due-diligence work.

If you are an energy data holder, map what information a customer could reasonably authorise you to share, what granularity is safe and how you would authenticate an authorised third party. If you are a service provider, build permission, revocation and decision evidence into the first release. If you are an end-user business, ask for a demonstration using a realistic site, imperfect data and a disputed recommendation. The response to the disputed case will tell you more than the dashboard.

Assign one owner for data access, one for operational action and one for measuring benefit. Sometimes one person fills two roles, but the responsibilities should still be explicit. Agree when the service is allowed to recommend, when a person must approve, and when it must stop because the data are stale. Review the first month’s recommendations against actual bills or meter outcomes and record the cost of false positives alongside the value of correct suggestions. That is how a pilot becomes a defensible investment decision.

The opportunity is substantial for businesses with repeated energy decisions across many sites, but there is no need to pretend the regulatory framework is finished. Use the consultation period to build a product or procurement standard that works with today’s authorised data and can adapt to tomorrow’s scheme. The competitive advantage will belong to services that make trust and evidence operational, not to those that attach “AI” to an uncertain data pipe.

Frequently Asked Questions

Is the energy Smart Data framework already law?

No. DESNZ opened a consultation on 14 October 2026, closing on 8 January 2027. The proposed framework would use powers under the Data (Use and Access) Act 2025, but the proposed details are not a live access entitlement.

Does customer permission to share data let an AI vendor train on it?

Not automatically. The sharing authorisation and every downstream purpose need to be clearly described and assessed under applicable data protection law. Model training should not be quietly bundled into permission for energy advice.

Can a UK GDPR portability request support real-time optimisation?

Do not assume so. Portability requests are case-by-case; Smart Data schemes may establish continuous, standardised access. An export may support periodic analysis without supporting live control.

What should an AI energy pilot measure?

Agree a baseline and compare recommendations with actual outcomes, accounting for tariff changes, weather and production where material. Track false positives and implementation cost as well as estimated savings.

Who is responsible for privacy compliance in a Smart Data scheme?

Participating organisations must determine whether they are controllers, joint controllers or processors and meet their respective duties. Scheme design can support compliance but does not transfer those responsibilities.

Should an AI service be allowed to change an energy contract?

Start with read and recommendation rights. Require a named human approver for supplier contact, tariff changes or contractual commitments, with the supporting evidence recorded.

What if the meter data are incomplete or stale?

The service should flag the gap, limit its claim and avoid operational action. Define a freshness threshold and a fallback to human review before the pilot starts.

Who can respond to the energy consultation?

DESNZ invites energy and technology organisations, consumer groups, academics and any interested organisation or individual to respond by 8 January 2027.