Parag K Adarkar

Insights AI that pays off

Start with the workflow, not the tool: where AI really pays off in a utility

Parag K Adarkar · · 6 min read

Enterprise AIUtilitiesWorkflow automation

A chatbot can answer questions. An enterprise needs more than answers. It needs AI that understands context, triggers the next action, supports a decision and leaves an audit trail behind it.

That is why I believe the AI conversation in most boardrooms is still framed the wrong way. The question being asked is “Which AI tool should we use?” The question that creates value is “Which business workflow should become more intelligent first?”

The evidence is now hard to ignore

Three pieces of research, read together, tell a consistent story.

McKinsey’s global survey on the state of AI tested 25 attributes of how organisations deploy generative AI. The one with the biggest effect on bottom-line (EBIT) impact was the redesign of workflows, and yet only about one in five organisations using generative AI had fundamentally redesigned even some of their workflows.

MIT’s NANDA initiative, in its 2025 report The GenAI Divide, found that around 95% of the enterprise generative AI pilots it studied showed no measurable impact on profit and loss. It is worth reading that figure carefully: much of it reflects pilots that were never connected to a business process, or never measured against a baseline. The report also observed that budgets flowed mostly to visible front-office experiments, while the clearer returns came from less glamorous back-office automation.

Gartner, meanwhile, predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, because of rising costs, unclear business value and inadequate risk controls. It also warns of “agent washing”: by its estimate only around 130 of the thousands of vendors marketing agentic AI offer the real thing.

None of this says AI does not work. It says AI that floats outside the way work actually gets done rarely pays for itself.

Where AI creates the most value

In my view, AI earns its keep where at least one of five conditions is present. Each of them is common in a power utility.

1. Repetitive work. High-volume tasks that follow a pattern but still need human judgement at the edges: validating meter-reading exceptions before bills go out, classifying consumer complaints, checking documents for a new connection.

2. Decision delay. Places where work waits for someone to read, compare and approve. Load sanctions, connection approvals and vendor bill verification often spend more time in a queue than in actual review.

3. Data overload. Smart meters, SCADA, outage logs and call-centre records produce more data than any team can read. AI is very good at turning that volume into a short list of things a human should look at today.

4. Compliance pressure. Regulatory filings, audit responses and standard operating procedures involve assembling evidence from many systems. AI can draft, cross-check and flag gaps; accountable people still sign.

5. Poor process visibility. If nobody can say where a complaint, a fault or a field job is right now, that is a strong candidate. The International Energy Agency estimates that AI-based fault detection can cut outage durations by 30–50%, largely by locating faults faster.

If a process has none of these five traits, it is probably not where your first AI rupee should go.

What “connected to the workflow” actually means

An AI capability is genuinely inside a workflow only when it is connected to six things. I use them as a simple checklist before approving any pilot:

  • Business process: it sits at a defined step, with a clear trigger and a clear next action, rather than in a separate window people have to remember to open.
  • Data quality: the data it relies on is owned, current and trusted. AI amplifies bad master data; it does not fix it.
  • User adoption: the people doing the work helped design it, and their day gets easier, not longer.
  • Security: access follows the same roles and permissions as the underlying system, and nothing leaves the organisation that should not.
  • Governance: someone is accountable for its outputs, and there is a route to challenge or override them.
  • Measurable outcomes: there is a baseline before the pilot and a target after it: hours saved per bill run, days off connection approvals, minutes off outage restoration.

The last point is the one most often skipped, and it is the main reason so many pilots cannot prove their worth. Measure the workflow before you change it.

A practical way to start

For a utility leadership team, I would suggest a four-step sequence:

  1. List ten candidate workflows across customer service, billing, field operations, finance and IT service delivery, and score each against the five value conditions above.
  2. Pick two, one customer-facing and one back-office, where the baseline is easy to measure and the data is already in good shape.
  3. Embed, don’t bolt on. Build the capability into the systems people already use, such as the CRM, the billing platform or the ERP, rather than adding another tool.
  4. Decide in 90 days whether to scale, fix or stop, based on the measured outcome rather than on enthusiasm.

Be particularly careful with “agentic” promises. Agents that take actions on their own are powerful, but in a regulated, safety-critical business they should earn autonomy step by step, starting with drafting and recommending before they are allowed to execute.

The real shift

The future of enterprise AI will not be built from isolated experiments. It will come from embedding intelligence into the way work is actually done, one well-chosen workflow at a time.

So the question I would put to every CIO and CTO is a simple one: where is your biggest AI opportunity — customer service, operations, finance, HR, IT service delivery or field workforce management? Start there, measure it honestly, and let the results make the case for the next one.

Sources