Starbucks, a cup and a million AI brains

By Dan Ilett

Starbucks has a cup problem. The Frappuccino arrives in a plastic cold cup and single-use plastic is on its way out.

Whatever replaces it has to hold a blended iced drink, pass the food-standards test, satisfy the environmental rules and survive the walk from the counter to the car.

No material on the market does all of that yet. Somebody has to invent it.

Simon Hill, CEO and founder of innovation firm Wazoku, is orchestrating a search for the answer.

"We have a network of a million super smart people in our crowd," he says. "I'm doing a project for Starbucks at the moment, trying to see if we can come up with a new environmentally compliant substance packaging for the Frappuccino cold cup. They want it to meet all of the standards, both environmental and food standards, to be able to do that, and no solution exists."

Same-same different

His answer is to crowdsource the response, using a blend of real people and AI agents tuned to think differently. The non-conformist agent is designed to deliver non-standard answers.

This is fascinating when you consider that High Street AI platforms often push people to the same, homogeneous answers.

Hill, however, is looking for heterogeneity. As he puts it, the expert is the last person you want to solve the problem.

"It is empirically proven that experts are the least likely people to solve an expert problem," he says. "If it's a chemistry problem and you find the ultimate expert in that chemistry field, if they don't solve it in the first five minutes, then they're the least likely person to solve it. And so you need variety. You need perspective to come from different areas.

"In innovation land there's been this perennial debate around are more ideas good or bad," he says. "The answer is if you can't get enough, you can't use it to draw the map. You can't figure out what the solution space might look like."

With a synthetic crowd he says you can.

"You can draw a really rich map. I can point you to the most likely areas now where humans should go dig, with all of the power of them plus the machine, much deeper into the solution set than you would ever be able to do before."

The industry he works in is built the other way round.

"One of the big flaws in the innovation design work that we do is we typically define a chemistry problem as a chemistry problem and then bring chemists in to solve that chemistry problem, and then don't necessarily solve it," he says. "It's complicated to know where to go look, but we believe people have been doing it wrong for a very long time."

But random variety was simply too expensive to manufacture, he says.

"It was quite hard to create heterogeneity. Our approach of doing it was to ask a million people who are super smart and they'll give you that heterogeneity.

But with AI, now I can ask even more of those, in parallel, really quickly, and draw the map for you of where the likely answer is."

He is careful to separate that from what most people are doing with a chatbot, and impatient with the way the question is usually framed.

"This generic use of AI, do you use AI for things, is a bit like saying 'do you eat food?'.

"There's a bigger, deeper question of what type and for what purpose."

The second brain effect

"We've built a whole company brain that means work that used to have to pass through lots of different meetings and people is now institutionalised and accessible by anybody," Hill says.

"It means anybody can be a builder. It requires humans to come in and check and validate, but more people can do more things quicker, faster, cheaper and better than they could do internally."

On the product side the numbers are harder to picture. "I try and track how many of those different agentic workers sit alongside my team-based workers at any moment in time. It could be thousands or tens of thousands of them up and running."

The models underneath are chosen to fit the job, says Hill.

"A lot of them are not the big flagship expensive models that maybe you would see elsewhere, because they're quite general purpose models and you want to get more specific depending on the task that you're doing."

Internally there is a company brain and increasingly a personal brain for each member of staff - synthetic sidekicks built on Claude Code.

"That's the front-end sort of UX," he says, "but it's all fed into a quite complex Snowflake based sort of cortex at the heart of all of it."

The hard yards were all in data.

"Most of the work we've been doing to get to this point was about data, and how do we get the data into a set of structures that can be used that doesn't overwhelm. We started from zero," he says.

"It had to earn its place, and it's still pretty thin what's in there, because you're trying to figure out what the minimum viable knowledge is to produce the output you want, not just give it everything and then train it and train it and train it."

The brain is also fenced in following a trend in what is now called the control plane - agent jail.

"Any of the company brain stuff operates only ring-fenced within the company brain," he says. "Although there's an AI there, it itself is operating only within those guard rails. It can't go and get extra information. It only knows what it knows, which is essential for us to know what we're feeding it."

Ripping out HubSpot

The clearest illustration he gives of what this makes possible happened last Christmas.

"We ripped out a whole load of software including our CRM. We pulled HubSpot. They tried to jack up the price. I said no. They said, you can't pull out a CRM with no notice. And we said, watch us."

Wazoku - like many other firms at the moment - rebuilt it. (I'm hearing so many stories like this.)

"We have an open-source CRM tool that we use. We build a whole set of synthetic capability on top that's fed into this company brain. So as you've got interactions happening, it can keep your CRM up to date for you. It can give you updates of things that are happening. It can tell you if something changes or not."

The cost-saving moved with it.

"From being quite isolated on a HubSpot side and quite limited in terms of what we could do, within six, nine months we've got it as part of a much broader core that's now got a whole bunch of other capabilities coming off the back of it.

"It used to cost me significant six figures. It now costs me maybe £20k or something a year."

He is quick to head off the obvious conclusion. "Is SaaS dead? No. But you've got to get the right tools to be able to do the right things. It doesn't mean that buy is gone and build is everything."

What has changed is what he will buy.

"We want our technologies to be open and connectable and ideally have MCP connectivity, because our data lives somewhere else really, and they're just systems of record for very specific things."

The Copilot company

I put to him the situation I see most often, which is a leadership team that has bought a licence, has no strategy and does not know where to start.

He gives three ideas:

The first is the shape of the data. "It isn't going to be in a bunch of PDFs or in a bunch of Excel files and be really useful.

"So there's a whole data structure layer that we spent a lot of time on. We've done this in house. We didn't bring extra people in. You often don't need that much data to start to make this really, really useful."

The second is to understand the distance between the three things people lump together.

"There's a very big difference between just writing prompts into an AI and then building quite rudimentary agents in Claude Code, and then really getting sophisticated agents up and running."

He would rather people got their hands dirty than waited.

"I wouldn't be terrified of it. You can get quite hands-on and start playing around and practising on quite interesting things, like the personal brain side."

The third is to start and keep going.

"It can feel really overwhelming, but just getting started is good. And then expect to be continuously adapting.

"You don't have to be this frontier. It might sound like, oh, we're really far behind, but my job is to push the boundaries. I live right on the edge of the innovation space.

"A lot of these things I'm talking about, they're great and they're in the daily core, but there's still lots of kinks around them, and almost all of my customers are a million miles away on this stuff as well.

"We have evidentially multiplied our productivity with AI, which we wouldn't have been able to do if we'd not pushed on this. We're trying to be an exemplar of what can be possible."