How Danone’s AI took on 32,000 patient calls without waiting for a perfect business case

Gartner predicts that 40% of agentic AI projects will be cancelled by 2027, because they never deliver a clear outcome. A separate industry figure puts it more starkly: 95% of enterprise generative AI projects show no measurable financial return within six months of launch. Both numbers get quoted a lot. What usually gets missed is that neither is really about the technology.

Danone’s specialised nutrition business is one of the 5%. Its Homeward service delivers tube-feeding nutrition and equipment to around 37,000 patients across the UK, many recovering from surgery, some of them children, most dependent on a delivery arriving on time because there’s no other way for them to eat. Two months after going live, Danone’s AI voice agent, Nia, short for Nutricia’s Intelligent Agent, had handled more than 32,000 patient calls. It still won’t tell anyone how to manage their feeding pump.

That restriction is deliberate, and it’s arguably the more instructive part of the story. Danone got here by being precise about what AI should do first, and disciplined about what it shouldn’t do yet. Here’s how that sequencing actually worked, and what it looked like when another business got it right too.

95% of AI projects show no measurable return within six months. Most start with the tool, not the problem.

The pattern is familiar enough that Simon Woodward, Natterbox’s Director of AI Solutions, calls it AI buzzword bingo. A leader says go and find out what AI can do for us. A team spends months building something that technically works. It gets piloted, it looks good, everyone feels pleased with it. Then the same leader asks the obvious question: what’s the ROI here? Nobody has a good answer, because ROI was never the brief. Proving something worked was.

The two figures above share the same root cause:

• Gartner’s 40% cancellation figure
• The wider finding that 95% of generative AI projects show no measurable return

Both trace back to the same mistake: the project started with the tool and worked backwards to find a use for it. As Simon puts it:

“You don’t buy a hammer and then go looking for a house that needs building. It’s the same with AI.”

AI gets bought that way constantly.

The fix isn’t more caution, it’s a different starting question. Every call into a contact centre is a transaction: someone wants something done. The question is which of those transactions happen thousands of times a month, resolve the same predictable way, and can be learned as a pattern. Start there, and the tool follows. Start with the tool, and you’re guessing.

Two variables decide whether a conversation is ready for AI: impact and complexity

Natterbox’s working model plots every interaction on two axes: business impact per interaction, and interaction complexity. Complexity here means something specific. It isn’t the number of systems touched or steps involved, it’s emotional complexity: how much the outcome depends on someone feeling heard rather than simply resolved.

Where to start	What lives here	Why
Zero-risk, high-frequency	Password resets, balance checks, CSAT follow-ups	Zero emotional risk, fast to build, and it retrains customers to expect a real conversation instead of the old “say one word after the tone” IVR
24/7 and proactive	General support, out-of-hours cover, appointment reminders, order status updates, renewal notices	This is where measurable ROI starts appearing
High value, high complexity	New orders, deeper technical support, multi-step resolutions, collections, re-engaging lapsed customers	AI carries much of it, but needs a defined handover to a person as the stakes climb
Never automate	Bereavement, bespoke legal or financial advice, anything that needs to be heard	The person calling wants to be heard, not resolved quickly, that’s a different job to the one AI is built for

Freeing up that last row is the actual point of automating the other three. It’s also why the measure of success changes: from how many calls a person handled, to how well they handled the ones that mattered.

Making AI the first point of contact recovered £1.3m in revenue, plus £240k nobody was chasing

One Natterbox customer, a UK contact centre handling 4.5 million cases a year, was missing 114,000 calls annually before AI. The obvious read is a capacity problem. The real story was more specific: the team spent so much time on high-volume, low-complexity work, order status checks, delivery changes, that the same people responsible for taking new orders never got to the calls that actually made money. Buried in those missed calls were people trying to place new orders, worth roughly £1.3m a year in revenue once modelled against the conversion rate on calls that did get answered.

Option	What it looks like	Outcome
Hire more people	Add headcount to answer the extra calls	Expensive, slow, headcount stays tied to call volume forever
Route missed calls straight to AI	Send all 114,000 missed calls to an AI agent	These are high-stakes sales conversations, AI won’t object-handle to the standard of a top rep, most of the £1.3m stays missed
Make AI the front door (chosen)	AI triages every call, handles everything that isn’t a sales conversation, frees reps to close	Full £1.3m captured, plus an unplanned £240k a year in routing efficiency

“Let’s make AI the front door. Have it figure out why someone’s calling, decide whether it’s an AI task or a sales task, and unblock the sales team to close.”

The full £1.3m got captured because the reps could finally pick up the phone, not because AI answered instead of them. The £240,000 side benefit came purely from routing correctly the first time, instead of callers being mis-directed through a switchboard and repeating their situation three times before reaching the right person.

Danone’s AI has handled 32,000 patient calls in two months. But nothing clinical.

Danone’s specialised nutrition arm is a different business to the yoghurt and water most people associate with the name, closer to healthcare than FMCG. Homeward supports patients on home enteral feeding, many recovering from surgery, some paediatric, with carers managing genuinely difficult circumstances. Getting a delivery wrong here isn’t an inconvenience.

Before anything else happens on a call, Nia runs identity verification, the same checks a human agent runs:

• Surname
• Date of birth
• Postcode

That’s a hard gate, including phonetic matching for when it mishears a name, and a defined exit if it can’t verify someone. Order management doesn’t start until ID&V is passed. Nia never gives clinical advice either, every call is pointed back to a healthcare professional for anything medical. No customer data is used to train the underlying model, a line Simon draws deliberately given how many regulated, safety-critical customers Natterbox works with.

Artem Gordiyenko, Danone’s Regional CIO for Northern Europe, is direct about the result: “the quality of what we see coming out of AI is clearly higher than we see from the people.” Not a comment on human effort, more a comment on consistency. Homeward’s contact centre work is a genuinely demanding, junior role, hiring and training someone to a patient-safety standard takes time, and a stretched human team’s quality varies across a long shift in a way a well-configured AI’s doesn’t.

He’s just as direct about the partnership itself. Asked how Danone found working with the Natterbox team: “We love working with the Natterbox team, and not just on this project. The guys really know their product, and they can bridge the needs of the customer into the solution very quickly and easily. They’ve done an amazing job helping us with Nia.”

What’s next:

• Full order management: updates, exchanges, delivery date changes, handled without a person involved at all

What’s deliberately not next:

• Pump troubleshooting, considered and ruled out for now on data and quality grounds rather than sensitivity grounds

Language is being extended with the same discipline. English only today, with German, Dutch, French and Spanish on the roadmap, and the actual work isn’t translation, it’s building the vocabulary of what patients really call things. Patients don’t ask for their Fortisip Compact Banana delivery. They ask if their milkshakes have arrived.

Meanwhile, the same AI analysis reviewing every call is being used to coach the humans still on the phones, flagging things like a missed chance to lock in a delivery date, turning a smaller remaining team into what Simon calls super agents rather than replacing them.

The business case doesn’t need to be perfect. It needs to point at the right square

Artem’s advice to anyone starting this: don’t spend months building a perfect business case in a lab, the learning comes from piloting, not planning. Getting to a usable starting point doesn’t need to be a multi-month preparation project.

That advice only works because Danone had already done the harder part: finding the right square on the grid. A high-frequency, low-emotional-risk transaction, a real efficiency case behind it as patient volumes grew faster than headcount could follow, and a clear read on what quality and complaint reduction were actually worth. The framework did the thinking about where to start. The pilot did the learning about how to get there. Getting the sequencing right is what let Danone move fast without it being reckless.

This piece draws on the full conversation between Simon Woodward and Artem Gordiyenko, including the live product demo and audience Q&A. The complete session is available to watch on demand: Watch the webinar replay.

For more results like Danone’s, see our customer stories.