Insights

Your AI pilot worked. So why hasn’t anything changed?

29 Oct 2026
Five outlined squares on a cream ground, each with a line running right. Four stop short of a dotted barrier; one blue line passes through to a blue target.

Most AI pilots I see actually work. The demo lands, the proof of concept hits its numbers, and everyone agrees it’s promising. Then nothing much happens.

That’s not just my experience. Merkle’s new playbook, The Five Patterns That Make Enterprise AI Actually Work, pulls together some uncomfortable numbers. Enterprise AI spending is projected to reach $3.3 trillion by 2027. Yet Gartner found that only 28% of AI use cases in infrastructure and operations fully succeed and meet their ROI expectations, and it predicts more than 40% of agentic AI projects will be cancelled by the end of 2027.

Leaders are bought in. Budgets are committed. The pilots succeed. It’s the step from pilot to proper deployment where things stall. Merkle calls it pilot purgatory, which is about right.

Drawing on hundreds of machine learning projects over the past three years, Merkle sets out five patterns that separate the organisations that get out from the ones that stay stuck. I agree with all five. Here’s our take on each, and what it means if you’re running a business rather than an AI lab.

1. Start where your data already lives

Every business has a place where its most useful data piles up: the CRM, the policy system, the customer service platform. Merkle calls this data gravity, and argues that’s where AI should start, not wherever the latest vendor demo points you.

Their best example is close to home for a lot of the businesses we work with. An insurance client built AI around its customer service data (call transcripts, tickets and chat logs) to prioritise and route enquiries automatically. The saving was $1.5 million a month.

The catch is that the data has to be ready. If two people pull the same number and get different answers, or the logic behind your reports lives in one analyst’s head, AI won’t fix that. It will inherit it, then give you confidently wrong answers at speed. Writing down what your data actually means is dull work. It’s also the work that makes everything else possible.

2. Start with the decision, not the technology

“Transform customer service” isn’t a use case. It’s a slogan. Merkle’s test for a real one is three questions: what decision does this enable, what delay does it remove, and how will you know it’s working in six months?

It’s the same discipline I wrote about in Define the problem before you buy the tool. If you can’t finish the sentence “this helps [this person] decide [this thing] faster”, you’re not ready to buy anything.

3. Redesign the work, not just the tools

This is the hard bit, and the one most organisations underfund. Merkle puts the split at 93% of AI investment going on technology and just 7% on people. Their line on it is worth repeating: AI applied to poor workflows “doesn’t fix them, it amplifies them.”

Merkle’s research with 100 US C-suite leaders found that tech-mature organisations are 3.5 times more likely to have a formal change management function with dedicated resources behind it. That isn’t a coincidence. If nobody has redrawn the workflow, and you can’t say which steps are AI, which are human and which are a human checking the AI, you haven’t redesigned anything. You’ve bolted a tool onto the old process.

4. Earn trust by measuring it

AI works in probabilities. Some answers will be wrong. Without measurement, teams end up checking everything by hand, the efficiency disappears, and people quietly drift back to the old way.

Merkle suggests tracking accuracy, consistency and adoption for the people running the system, and ROI and revenue contribution for the people paying for it. Their restaurant example is a good warning. When a chatbot started giving inaccurate answers across more than 6,000 locations, field teams stopped trusting it. Fixing the accuracy, and making that accuracy visible, took adoption to over 100,000 users.

If your answer to “how would we know if this got worse?” is “someone would notice”, you don’t have a measurement plan.

5. Get the right insight to the right level

A board, a sales director and a team leader need different things from AI, and at different frequencies. Some of it should be there when someone asks a question. Some should arrive without anyone asking: a weekly briefing, or an alert when something drifts off target.

Merkle’s practical test is a good one. Is there a report someone still builds by hand every month? That’s your first automation target.

What this means for a mid-sized business

Merkle is writing for global enterprises with large data teams. Most of the businesses we work with aren’t that. But the patterns hold at a smaller scale, and arguably matter more, because there’s less budget to waste on a pilot that goes nowhere.

If I had to pick one place to start, it would be pattern two. Pick one decision that’s slow, costly or inconsistent today. Find out where the data behind it already lives. Redraw the workflow around it, and agree how you’ll measure it before you spend anything. Then do it again.

None of that needs a trillion-dollar budget. It needs someone to be clear about what’s supposed to change.

If you’ve got AI stuck somewhere between pilot and proper rollout, ping me a message, or let’s grab a coffee.

Which AI pilot in your business worked well but never went any further?


Sources: Merkle (a dentsu company), The Five Patterns That Make Enterprise AI Actually Work, 2026. Gartner figures as cited in the Merkle playbook. Case studies and survey figures are Merkle’s.

Share this insight

Related insights
Four pairs of bars on a cream ground: tall outlined bars for AI's potential beside much shorter blue bars for actual use.

Almost every B2B marketer uses AI now. Most are still dabbling.

New ON24 research shows 93% of B2B marketers are using or planning to use AI. The real story is the widening gap between teams using it properly and everyone else.
30 Oct 2026
A faint grid of binary zeros and ones on cream, with a single blue magnifying glass to the right.
AI, Digital

Scotland’s biggest law firms have cracked SEO. Almost none are ready for AI search.

Scotland’s biggest law firms have cracked SEO, but almost none are ready for AI search. What we found across the ten largest firms, and the five fixes that move a firm from invisible to quotable.
2 Oct 2026
Abstract illustration: a field of identical outlined squares with one replaced by a blue target, representing a clearly defined problem.

Define the problem before you buy the tool

If you’re sitting in the C-suite right now, the pressure to buy AI is real, and it’s coming from every direction. Customers are starting to expect it. Competitors are being loud about having it. And not being able to point to your own AI story can start to feel like a disadvantage, whether or not it actually is one.
1 Oct 2026

Let's grab a coffee.

If something's stuck, or you'd like a second opinion on a plan, tell me about it. No hard sell.