The Decision System Your AI Strategy Is About to Expose
Most AI strategy failures are not technology failures — they are AI decision making failures hiding in plain sight.
I was in this meeting.
Global FMCG. Regional HQ. ExCom quarterly review. Five in the room, three on screen. Agenda item six: AI-enabled commercial execution.
The item had been pushed twice before it landed on the agenda that morning. What follows is as close to verbatim as my notes allow. Names and markets have been changed.
CDTO: “So we’re now at month seven of the commercial AI rollout. I want to give you the honest picture before we go to the deck.”
CEO: “Please.”
CDTO: “Technically we’re in good shape. The next-best-action engine is live across nine of the twelve markets. Recommendation accuracy is sitting at 81%, which is ahead of where we modelled it at this stage. The infrastructure is stable. The data pipeline is clean.”
CCO: “What’s the sell-out impact?”
CDTO: “We’re still working through the attribution methodology with the analytics team. There are a lot of variables moving simultaneously — promotional calendar, pricing adjustments in three markets, the distribution changes in —”
CCO: “Seven months in and we’re still working through attribution.”
CDTO: “The attribution problem is genuinely complex. We didn’t want to give you a number we couldn’t defend.”
CFO: “We’re carrying €4.2 million of investment on this programme. At some point we need a number we can defend.”
CEO: “Let’s hold on attribution. What’s the adoption picture?”
CDTO: “Usage is up. 68% of field sales managers are logging into the platform at least weekly. In our lead markets — Turkey and Nigeria — it’s closer to 80%.”
CCO: “Logging in isn’t the same as acting on recommendations.”
CDTO: “No, it’s not. And that’s — that’s the more complicated conversation.”
CCO: “Go on.”
CDTO: “When we look at the recommendation-to-action rate, it varies significantly by market. In Turkey it’s around 60%. In some of the Gulf markets it’s closer to 20%.”
CFO: “20%.”
CDTO: “Yes.”
CFO: “So four out of five recommendations are being ignored.”
CDTO: “Not ignored. Unacted on. There’s a difference. The field managers are seeing the recommendations. The question is whether acting on them sits clearly within their mandate or whether it requires a conversation upstream first.”
CPCO: “This came up in the last market visit. The field teams don’t know if the recommendation supersedes the call plan or sits alongside it. Nobody’s told them which takes precedence.”
CCO: “That was supposed to be in the deployment guidelines.”
CDTO: “It was defined. It’s in the governance framework.”
CCO: “Then why isn’t it consistent?”
CEO: “Let’s get the governance framework reviewed and come back next quarter with cleaner numbers. Can we move to item 5?”
Three things just happened in that room.
A recommendation engine is producing outputs across nine markets that nobody can trace to commercial outcomes.
An ownership question — who acts on this? — has been answered on paper and not in practice.
And the decision to move to item 5 just ensured the same conversation will happen next quarter.
This is not a technology problem. This is not a change management problem.
This is what an AI decision making failure looks like when it’s dressed in the language of digital transformation.
What your AI decision making structure is about to walk into
The gap is almost never technical. It is an AI decision making architecture problem. The questions in the room are about accuracy, adoption, and attribution. The problem underneath them is structural — and it predates the AI investment by years.
The AI doesn’t create the dysfunction. It timestamps it. And the timestamp is what makes it, finally, impossible to ignore.
Four things need to be redesigned before any of those attribution numbers will make sense.
1. AI decision making rights need to be explicit. Not assumed, not implied, not distributed by seniority. Proximity over hierarchy: the person closest to the work needs to know, without ambiguity, that a particular class of decision is theirs to make. This is not empowerment in the motivational sense. It is architecture in the structural sense.
2. Funding needs to flow to the point of execution. Organisations approve AI budgets centrally. What they don’t redesign is how funding flows to the unit expected to act on what the AI produces. A very expensive signal that nobody has been resourced to receive. This requires an operating model answer — not a deployment answer.
3. Handoffs need to be designed, not managed. The point where work moves between functions — from insight to decision, from recommendation to action — cannot rely on informal relationships to function at AI speed. The process works when people who know each other manage it informally. It fails when they leave — or when pace exceeds what informal coordination can absorb.
4. Accountability needs to be as clear as authority. When nobody is sure whose call it is, nobody acts. When everybody is accountable, nobody is accountable. The two need to be designed together — not assumed to align. This is a design question. Not a leadership message.
The governance framework review will not fix this. Coming back next quarter with cleaner numbers will not fix this. The gap between recommendation and action is not a data problem — it is a decision system problem, and it will keep timestamping itself, quarter after quarter, until the architecture changes.
The technology is ready. Your AI decision making architecture may not be. The question is whether you are willing to look at what it reveals — and redesign accordingly.
The case for making the invisible visible before you fix anything.
You already know what’s broken. So does everyone on your team.
Ask anyone individually and they’ll tell you: too many priorities running in parallel, work arriving mid-cycle without warning, prioritisation criteria that shift depending on who’s asking. They know. They’ve known for months. Some of them have known for years.
So why hasn’t it been fixed?
Not because people lack awareness. Because each person holds a different version of what’s wrong. Your version is shaped by what you see from your position. Theirs is shaped by what they see from theirs. And without a shared picture, every conversation about priorities starts from a different set of assumptions, a different experience of the problem, and a different theory about what needs to change.
The step most organisations skip is the simplest one: making the current system visible to the people who operate inside it. Not visible to leadership in a deck. Not visible in a retrospective summary. Visible to the room, together, at the same time.
Shared evidence is not a diagnostic step that precedes the real work. It is itself a piece of scaffolding. And without it, everything you build next is weaker, because it rests on twenty different private readings of the problem rather than one shared reading the room agreed was real.
The survey that nobody presented
In a recent cross-functional engagement, spanning brand strategy leads through to operational delivery, the teams completed a short pre-survey. Six to eight calibration questions. Nothing exotic: how manageable is your workload, how do you currently prioritise, what drives decisions in practice, what gets in the way. The findings were clear, consistent, and unsurprising to anyone who had been paying attention individually.
But the results were never presented via decks. Instead, each finding went up on the wall as a poster. The room was given a sheet of dots and one instruction: walk the gallery in silence and place a dot next to anything that matches your experience.
No slides. No narration. No expert standing at the front interpreting the data.
The room went quiet for a few minutes. When people started talking again, the conversation was completely different from the one they had walked in with.
Here is what changed. Before the gallery walk, each person held a private version of the problem. “I feel overwhelmed.” “My team has too much on.” “We get pulled in too many directions.” These are individual experiences, and they stay individual until the room sees the pattern. The moment you discover that the majority of your colleagues also rated urgency as the dominant driver of prioritisation, the conversation shifts. It stops being “I feel overwhelmed” and becomes “we have a system that produces overwhelm.” That is a different sentence. It locates the problem in the system rather than in the person, and it gives the room permission to fix the system rather than cope harder.
The debrief asked four questions: What feels most familiar? What feels most damaging? Where are we compensating for system weakness with personal effort? Which issue, if improved, would unlock the most?
Every one of those questions moves the room from recognition to diagnosis. And because the evidence came from their own survey, the diagnosis belongs to them. Not to an external facilitator. Not to a leadership narrative. To the room.
The prioritisation map nobody expected
The second piece of shared evidence was more physical.
Each table took their real current initiatives, written onto cards from their own planning tools and backlogs, and placed them into a simple flow on the wall: coming in, committed, in progress, blocked or waiting, unclear or floating. For each card they marked who brought it in, whether it was planned or arrived through urgency, whether there was a named owner, and whether it was truly active.
What appeared on the wall surprised the people who built it.
Initiatives that had been deprioritised months ago were still consuming effort, because nobody had formally stopped them. Work that entered through urgency sat alongside work that entered through planning, with no visible distinction between the two. Some initiatives had no named owner at all. Others had owners who didn’t know they were owners.
The system’s actual operating rules became visible for the first time. And those rules were different from the ones the organisation believed it was following. The intended flow was strategic priorities into planning cycles into coordinated delivery. The actual flow had a second, parallel intake channel driven entirely by urgency and seniority, and that channel was often stronger than the planned one.
No one needed to be told this. They could see it. The wall told the story.
Why this is the step you cannot skip
It would be easy to treat shared evidence as a warm-up. Context-setting before the “real” work begins. That would be a mistake.
Every prioritisation decision your team makes after this moment either rests on visible facts or invisible assumptions. That is not a small difference. A team that has just seen its own data showing that urgency drives most decisions is ready to build criteria that explicitly address urgency. A room that has just watched its own initiative cards reveal three priorities with no named owner is ready to have the ownership conversation. Without the shared picture, those same conversations produce polite agreement that collapses within a planning cycle, because each person is solving a different version of the problem.
Once the room has seen the same picture, every subsequent conversation references it. Trade-off exercises are more consequential because the initiative cards already showed what is genuinely in flight versus what is drifting without an owner. Conflict conversations are more honest because the survey already named the patterns. Follow-up check-ins have teeth because there is a baseline to measure against.
This is the difference between scaffolding and decoration. Decoration makes a session feel grounded. Scaffolding makes every subsequent decision structurally stronger.
Before you build the protocol, build the picture. Before you ask the room to make trade-offs, show the room what it is trading off against. Before you introduce a new system, make the current system visible.
The room already knows. It just hasn’t seen it together yet.
This is the second in a series on the structures underneath good prioritisation. Previously: Prioritisation is a scaffolding problem. Next: the wall where the real rules live, and what happens when you give a room three minutes of silent writing and a set of category headings.