AI can’t keep its promises on broken platforms

Recently, I have spent a lot of time with energy and utilities leaders discussing AI, agentic systems and the move from experimentation to operational value. The ambition is clear. Intelligent grids, predictive operations and AI-driven customer journeys are no longer theoretical.

Most organisations know what they want AI to do. The harder question is whether their platforms can support it.

MIT’s recent State of AI in Business research puts a number on something many CIOs already recognise. Despite significant enterprise investment in generative AI, most organisations are still seeing little or no measurable return. The report did not point primarily to model quality, talent or regulation. It pointed to implementation: systems that do not adapt, workflows that do not fit the business and solutions that never properly integrate with day-to-day operations. A lot of these processes were designed with an early 2000s view towards applications and aren’t ready for 2026-2035.

That matches what I see in the field. The problem is not that AI cannot generate useful output. It can. The problem is that value only appears when that output is applied in context, governed responsibly and scaled into the operational workflows where decisions are actually made. Without that, organisations get impressive demonstrations that never reach production.

The real problem is not AI

Energy and utilities do not have an AI ambition gap. They have an enablement gap.

That is the gap between what organisations want AI to do and what their delivery system can actually support.

This distinction matters because it changes the work. If the problem is ambition, the answer is strategy. If the problem is model quality, the answer is better tooling. But if the problem is enablement, the answer is harder and more structural: data, platforms, governance, operating models and adoption.

Most stalled AI programmes are not short of ideas. They are short of foundations. CGI’s own work across energy and utilities points to the same recurring barriers: fragmented OT and IT data, complex legacy estates, heightened cyber and regulatory risk, and pilot fatigue. None of those are AI problems in isolation. They are delivery problems with AI on top. That is why so many programmes look promising in a lab and then slow down the moment they touch production reality.

Why this is harder in energy and utilities

Generic AI advice does not land for a utility CIO, and there is a reason. You are not a high street retailer. You are not a bank. You operate mission-critical infrastructure where a bad recommendation can have physical consequences: transformers, grids, pumps, field crews, market obligations and customers without power on a hot day.

Your OT and IT systems were often procured decades apart by different teams, for different reasons, from different vendors. Your data is fragmented across asset, network, customer, market and field systems that do not always share the same definition of the truth. Your regulatory, safety and cyber requirements are non-negotiable. And the assets you are trying to instrument have lifecycles measured in decades, not quarters.

In this sector, intelligence is only valuable if it is operationally trustworthy. That is the bar. And it is a bar most AI programmes are not designed to clear, because they are still optimising for model performance when the real constraint is the system the model has to live inside.

Five design choices that determine whether AI scales

This is the part where I would usually expect a consultant to give you a maturity model. I am not going to. What is more useful is a small set of design choices to test your AI work against.

When you are evaluating a platform, a delivery approach or a new AI initiative, hold it up against these five things and look for the gaps.

1. Connected data across IT and OT

Data has to work as one system.

Asset, network, customer, market and field data cannot remain five overlapping silos with three different keys. Without connected data, AI is blind to half the operating environment. It can see the meter anomaly but not the planned outage. It can see the customer contact but not the field constraint. It can generate a confident answer from a partial truth.

That is not intelligence. It is risk with a better interface.

The practical work here is not glamorous, but it is essential: common data products, trusted lineage, consistent asset identifiers, integration patterns and clear ownership. Until those foundations exist, AI will keep exaggerating the quality of the data beneath it.

2. Operational context by design

Raw data is not enough.

AI needs engineering meaning: asset hierarchies, grid topology, outage history, maintenance records, weather signals, control room procedures and the relationships that make a reading interpretable.

This is where concepts such as digital twins and digital triplets become more than architecture language. A digital twin can mirror the asset or network. A digital triplet goes further by adding an AI-driven decision layer that can interrogate operational data, simulate scenarios and explain recommendations within defined guardrails.

That context is what allows AI to move from answering questions to supporting decisions. Without it, the model is not reasoning about your network. It is pattern-matching around it.

3. Governance shifted left, not added at the end

Security review, model risk, data lineage and compliance all need to exist. That is not the debate.

The question is where they sit.

If a security review takes eleven weeks because it happens after the build, you do not have governance. You have a queue.

The mature pattern is to embed the constraint in the template the engineer starts from. The standard is met before the work begins. The review becomes lighter because the default path is already compliant.

This is shift-left as a delivery practice, not as a slogan.

It means approved patterns, secure templates, built-in logging, default data classification, policy-as-code where it makes sense, clear model assessment routes and guardrails that help teams move faster because they are not waiting to discover the rules at the end.

Governance should make safe delivery easier than unsafe delivery.

4. AI embedded in operations

AI has to live where the decision happens.

That means the dispatch system, the maintenance scheduler, the planning tool, the control room workflow, the field mobility application and the customer service queue.

If AI still requires someone to leave the workflow, open a sandbox, copy data into a prompt and manually interpret the answer, it is still a pilot.

The hardest part of getting AI to production is rarely the model. It is the integration into the workflow the human actually uses, with the right context, auditability and escalation path.

This is where operational value appears: not when AI produces an answer, but when the organisation can trust that answer enough to act on it.

5. A living platform, not a launch

Platforms are not projects.

They do not have a finish line.

If the platform team disbands when the project closes, you do not have a platform. You have a pile of well-intentioned code that nobody is looking after.

A real platform needs a product mindset: permanent ownership, funding, onboarding, documentation, contribution routes, inner sourcing, regular communication and a deliberate adoption plan.

Those are not soft things. They are how platform value compounds.

If three teams use your platform and forty do not know it exists, the platform is not a platform. It is a clubhouse.

Back to top

What to focus on over the next 24 months

For many energy and utilities organisations, the next 24 months should not be about launching another disconnected proof of concept. The work is more practical than that.

Make what you already have visible

Most large organisations have already built useful patterns, data products, reference implementations, architecture decisions and reusable components.

The problem is that nobody can find them. An internal catalogue that surfaces the work that already exists is one of the highest-return investments available. It does not need to be perfect. It needs to be maintained, searchable and supported by a curation function and a communication plan. Do not underestimate the value of making good work easy to discover.

Move one painful governance gate earlier

Pick the approval process that causes the most delay and embed it into tooling.

Do not remove governance. Move it earlier. Put the standard in the template. Make the compliant route the easiest route. Turn review from a late-stage inspection into confirmation that the established pattern has been followed. This is how organisations reduce friction without reducing control.

Build one initiative to a repeatable standard

Choose one initiative on the roadmap and build it to a standard others can copy.

Start with something that every team will need: APIs, data products, integration patterns, deployment pipelines or agentic workflow templates.

For example, a production-grade API template can enforce authentication, security, observability, data classification and documentation by default. Every team that uses it inherits those standards automatically.

That is how the gating steps become less visible. Not because they disappeared, but because they were already met.

Prioritise enterprise foundations over isolated pilots

This is the strategic commitment the others depend on. Reusable platform services. Trusted data products. Permanent platform teams. Funding that survives beyond the first launch. Adoption treated as a product outcome, not an afterthought.

The reason many utilities are not getting AI to production is not that they are bad at AI. It is that they have not funded the foundation AI needs to stand on.

The bet

Here is what I think happens over the next decade.

The utilities that win will not simply be the ones with the best AI strategy. Plenty of organisations already have a good AI strategy. That is not the constraint. The winners will be the ones whose platforms can keep the promises their AI strategies make.

They will be the ones that decided, around now, that the gap was not ambition. It was enablement. And then went and fixed it.

The failure rate of AI pilots is not a model problem. It never was.

Learn more about how we’re embedding AI in energy and utilities or contact me to discuss your AI-readiness