From 40% to 67% self-service: what Eneco, Rabobank and KPN teach us about AI agents

For many organisations, AI agents still sound like a distant prospect. Yet a look at the customer cases Microsoft publishes shows that Dutch, and international, organisations are already achieving measurable results with them today. We have pulled together the most interesting publicly reported figures, and, more importantly, the lessons you can take from them for your own organisation.

The figures organisations report themselves

Eneco reports that its customer-service agent, built with Copilot Studio, handles around 24,000 chats a month, more than twice as many as the old chatbot, and resolves 67% of conversations on its own, up from 40% previously. Rabobank reports that its AI assistants process some 20,000 phone calls and 7,000 chats every day, with 62% self-service on chat. KPN says that salespeople using Copilot for Sales save almost two hours a week, and that 69% of account managers saw the quality of their quotes improve. And chemicals group Dow has an autonomous agent check around 100,000 freight invoices a year; errors that previously took days or weeks to track down now surface in seconds.

An important caveat: these are results as the companies themselves publicly report them in Microsoft customer cases. They are no guarantee for your own situation, but they do show what is realistically possible, and in which processes agents prove their worth fastest.

Lesson 1: start with one well-defined, high-volume process

What all the successful cases have in common is this: they did not begin with "let's do AI", but with one concrete, high-volume process governed by clear rules. Customer questions about energy contracts. Freight invoices. Quotes. That is no coincidence: an agent proves itself fastest where there is plenty of repetition, where quality can be measured and where a human can step in to handle exceptions. The eye-catching applications come later; the business case starts out dull.

Lesson 2: measure before you start

The reason Eneco can say "from 40% to 67% self-service" is that it knew the old situation. Anyone who only starts measuring after going live can never demonstrate what the agent has delivered. Every agent business case therefore begins with a baseline measurement and a chain of metrics: from usage figures (how many conversations does the agent handle?) through operational KPIs (resolution time, hand-offs) to business outcomes (cost, customer satisfaction).

Lesson 3: build iteratively, with humans in the loop

Holland America Line built the first version of its virtual agent in three months; biotech firm Amgen built its first agent in six weeks. The pattern is the same: a limited but working first version, which is then improved on the basis of real conversations. Just as important: in almost every case the agent hands complex or sensitive matters over to a member of staff with a warm transfer. The agent does not replace people; it absorbs the volume so that people are free to do the work where they make the difference.

Lesson 4: governance from day one

The more agents you have, the more the ground rules matter: who is allowed to build agents, what data may they use, who owns them and how is quality assured? In regulated sectors in particular, with Rabobank's financial world leading the way, this is not a side issue but a precondition. In our experience: governance that is in place before the first agent accelerates every one that follows. Governance that has to be bolted on afterwards holds everything back.

What is your first agent?

The common thread running through all these cases: they began with one well-chosen process, a baseline measurement and a small team. Any organisation can do that, even without the budget of a bank or an energy group. In a half-day working session we map out your processes and prioritise agent opportunities by impact and feasibility. Schedule an introductory meeting and discover where your organisation will see results first.