Part one argued that AI projects fail on the harness: the context, tools, loop and guardrails you build around a model, not the model itself. Get that right and something predictable happens. Work comes off people.
Here’s where I think most of the ROI goes missing, and it has nothing to do with technology.
Freed hours don’t convert into value on their own. Somebody has to decide what those hours are now for, then change how the team is organized to match. Skip that and you’ve bought a faster version of the same output, which is precisely the 39 percent problem: near-universal adoption, and a minority seeing any impact on earnings.
A team structure that was well designed for the old process is often actively wrong for the new one. The roles were shaped around work you just removed.
The legible version
Sometimes it’s easy to see. Take an accounting firm we worked with, where tax-season documents arrived in a jumble, all at once, and juniors sorted them by hand to identify the relevant tax year. We built a system that automated the initial processing, flagged the uncertain cases, and produced organized client files with confidence ratings. Work that could consume almost a business day for some clients came down to 30 minutes.
The clever part is smaller than it looks. Every file gets a confidence score, and anything below a chosen threshold goes to a human; setting that number took longer than building the classifier, because it’s really a policy about how much risk the firm accepts before a person looks. Call it a triage rule.
Then the return came from what happened next. Sorting and collating went away, and what was left was the work clients actually pay an accounting firm for. Easy to sell internally, because the higher-value work was already sitting there waiting.
The harder version
It isn’t always that legible. IKEA is the harder case, and its numbers make the point better than I can.
Ingka Group put a customer-service bot called Billie into production in 2021. From 2021 to 2023 it resolved roughly 47 percent of the enquiries it received, about 3.2 million interactions, for nearly €13 million in savings. That’s the harness half, and €13 million is a perfectly good result.
Then look at what they did with the capacity. Rather than bank it as a headcount saving, Ingka reskilled 8,500 call-centre co-workers into remote interior design, digital retail sales and complex problem-solving, and pushed the channel into new markets including the UK and US. Reuters reported in June 2023 that sales through it reached €1.3 billion in Ingka’s 2022 financial year, 3.3 percent of total sales, with a target of 10 percent by 2028.
The automation saved millions. The reorganization built a business.
And the insight that mattered was organizational. Somebody noticed that the people answering routine questions were product experts, and that a product expert is wasted on routine questions. The bot didn’t create that expertise; it stopped consuming it. Everything after was design work: new roles, new training, a new line on the P&L.
Not a panacea, to be fair. IKEA cut around 1,650 corporate roles in spring 2026, unattributed to AI and not in the remote-sales centres, so “IKEA chose reskilling over layoffs” describes a decision made between 2021 and 2023 rather than the company today.
The process shift is still the lesson. It was as much of the work as the harness or the agent.
The pattern at scale
Organizations doing this well follow the same logic. Morgan Stanley built an assistant over a corpus of more than 100,000 internal documents for its 16,000 advisors, plus a second tool that turns meeting recordings into structured CRM notes; on OpenAI’s case study page, adoption in wealth management is put at 98 percent. JLL built an in-house platform on its own transaction and market data, and in March 2026 its CTO put daily use at roughly a quarter of the company’s 110,000 people.
Neither stopped at buying a licence. They built around their own data and their own way of working, then changed what their people spent the day doing.
We’ve seen the same shape in print houses, mining and gas, ecommerce, liquor marketing, gaming and entertainment, and CPG (not a spread anyone would design on purpose). Something arrives, it gets transformed, a decision gets made, an action follows. The industry changes what the inputs are called. It doesn’t change the design problem.
Where to look first
The heuristic we use for low-hanging fruit: work that’s repetitive, high volume, rule-driven, and currently absorbing hours of junior or coordinator time you’d rather spend on judgment. If you’d hire a junior to do it manually, it’s a strong candidate.
And on the question everybody in the room is actually thinking about, what we’ve seen is that these people get pulled up rather than pushed out, because what’s left after you remove the transcription is the judgment. That isn’t automatic, though: it’s the second half of the design, and it only happens if somebody owns it.
Why initiatives stall
| Aspect | What it looks like |
|---|---|
| Too many experiments | Lots of pilots and demos, nothing in production. Energy without output. |
| Not embedded in the workflow | The AI sits on top of the old process, which still runs the same way, slightly faster. |
| No owner of the redesign | Split between IT, the business and an AI champion. The seams are where it dies. |
| No plan for the capacity it frees | The system works, the hours come back, and nothing is waiting to absorb them. This is the quiet one, and it turns a working build into a project nobody can point at a number for. |
A 90-day version that works
- Identify five workflows using the heuristic above.
- For each, write down five things: the trigger, the inputs, the outputs, the decisions a human has to keep, and what the people currently doing it will do instead.
- Build one of them all the way. Harness included.
- Roll it out, move the roles, and measure it against what it replaced.
That last item in step 2 is the one people drop. Write it down before you build, because it’s an organizational commitment and it takes longer to arrange than the software does.
One working system beats ten pilots. Pilots produce slide decks; systems produce hours.
The actual skill shift
The old job was to do the work. The new job is to design how the work gets done, and that applies to a coordinator as much as to a COO.
So, some homework. Pick one thing you do every week. If you were designing it from scratch today, knowing what AI can now do, would you design it the way you currently do? Then the harder question: if that work went away next month, what would you do with the time, and who would need to be part of deciding?
Most companies can answer the first one. The second is where the €13 million becomes €1.3 billion, or doesn’t.
Wells Stringham has spent 18+ years building for Nike, Disney, lululemon, TELUS, and the NFL. He's Co-Founder & Partner, Experience at better&co, a digital collective working on strategic clarity, AI-native product development, experience innovation, and operational evolution. betterand.co