How OVO Energy turned AI adoption into behaviour change

A practical look at how one contact centre went past usage numbers to show its people were behaving differently, and what you can borrow for your own rollout.

Most AI rollouts reach the same point. Licences are live, usage looks healthy and feedback is positive. Then a leader asks whether anyone is working differently, and the dashboard has no answer.

OVO Energy reached that point too. Its agents rated their new AI practice tool 4.5 out of 5 and kept going back to it. But OVO's question was harder: when a bereaved customer rings about a rising bill, does the agent now handle that call better?

That question changes what you measure.

By answering the right-hand column, OVO showed that on the calls its agents had practised, Direct Debit process adherence rose from 65% to 97%. Here is how it approached the problem, step by step.

The challenge

OVO's contact-centre calls were getting harder faster than agents could be prepared for them. Rising bills and falling affordability meant more customers in distress. Direct Debit reviews, estimated bills and bereavement often landed on the same call.

Preparation hadn't kept up. New agents learned theory, systems and shadowing, then took live calls. E-learning showed who had attended, not who was ready. Facilitated role-play was unpopular, varied by facilitator and couldn't keep pace with continuous hiring.

The result was that the first time an agent handled a difficult call, the customer was real. Mistakes showed up as repeat calls, transfers, reopened cases and complaints.

Step 1: Define the behaviour before anything else

OVO didn't start with a tool or a topic. It started by writing down what an agent has to do differently on the call, and which business number each behaviour would move. The wording came from OVO's own policies and QA framework, so it matched the language already used on the floor.

1. 01Customer satisfaction

   Acknowledge and validate the concern before moving to process.

2. 02Direct Debit adherence

   Explain the Direct Debit calculation accurately, in plain language.

3. 03Transfers

   Recognise and respond to bereavement, distress or confusion.

4. 04Reopened cases

   Agree compliant next steps the customer understands.

Those four behaviours then ran through everything: what practice was scored on, what feedback referred to and what the pilot was judged against.

Step 2: Start where behaviour is most likely to break

Rather than roll out widely, OVO went narrow. It picked the call where it had the least confidence agents would get it right, and where mistakes cost the most.

That was the Direct Debit review. It's high volume and high complaint, and it often involves a vulnerable customer. Picture a widow calling about the first bill since her husband died: her account is in credit, yet her Direct Debit has gone up. The agent has to be procedurally right under Ofgem's rules and humanly right for someone having a terrible week.

The pilot covered that one call type and one team, OVO's Direct Debit Tiger Team.

Step 3: Practise the real moment, under realistic pressure

Knowing what to do and doing it under pressure are different skills. A course or a quiz builds the first. Behaviour only changes when people rehearse the second.

So OVO's agents practised the call itself, with an AI customer who reacts to how the conversation goes. She hesitates, pushes back, softens when handled well and can hang up if handled badly. There's no script, and the agent works with a realistic account on screens that match the ones they use for real.

The aim was a call that feels uncomfortable enough to matter, with no real customer at risk if it goes wrong.

Step 4: Treat a low first score as the starting point

The average first attempt scored 37 out of 100. That number was useful. It showed how far agents were from ready before they ever met a real customer on that call.

The improvement came through repetition. Agents took the call, read feedback on the specific lines that missed, chose one thing to change and went again. Feedback arrived within minutes, so the next attempt happened while the last one was fresh.

Between first attempt and best, the average score rose from 37 to 82. On the flagship Direct Debit scenario it rose from 32 to 87, and all four scenarios followed the same pattern.

Average, first attempt to best

37→82

Flagship Direct Debit scenario

32→87

Step 5: Measure behaviour, then test it on the real job

OVO tracked three things separately in practice, because each one can fail on its own.

• Competence: can they do it?

  How well the four behaviours were performed on the call.

• Confidence: will they do it?

  Agents rated their confidence before and after, so it was measured rather than assumed.

• Compliance: was it done the way it has to be?

  Whether the call followed OVO's policy, with a missed acknowledgement of distress treated as seriously as a wrong calculation.

Practice scores still weren't the proof. The proof was whether behaviour changed on live calls. So OVO compared the practised team with the rest of the floor, on the same calls, in the same period. The Tiger Team practised for 12 weeks alongside live work while everyone else carried on as usual. OVO measured and reported the results itself, from March to April.

That's evidence I can take to the board, not a completion certificate.

Step 6: Build practice into how people start

A pilot proves the point once. Behaviour holds only if practice becomes part of how people work, because new people keep arriving and old habits creep back.

OVO is now building practice into how new agents get ready, at the two points where habits form.

• The new-hire bootcamp

  Agents practise calls in classroom training before they go near a live customer.

• Grow Zone

  Their first weeks on real calls, where they practise alongside live work on the calls they feel least comfortable with. It ends with an assessment they must pass.

New scenarios are added whenever a gap appears, from data protection checks to offering the Priority Services Register.

Applying this to your own AI rollout

OVO's example is a contact centre, but the pattern applies to any AI rollout. Adoption depends on tools and access. Behaviour change depends on a handful of human moments, such as:

• a team leader talking to a long-serving colleague who is quietly avoiding the new tool
• an advisor checking an AI draft before it reaches a customer
• an employee under deadline pressure, tempted to paste customer data into a chatbot

Four questions will tell you whether your rollout can see those moments:

1. Have you named the behaviour you need in each moment, in your own policy language?
2. Do you know which moment is most likely to go wrong, and for whom?
3. Have people practised it with realistic pressure, more than once?
4. Could you show the difference against a group that carried on as usual?

If the answer to any of these is no, that's where to start. OVO began with one call and one team, and that was enough to show the board real change.

OVO's practice programme was built with Real Talk Studio. The full customer story is on the Real Talk Studio website.