“Let Them Fail” – AI Insights from the Strategic Innovation Sprint

Following last week’s Strategic Innovation Sprint in London, we’ll be sharing insights focusing on various conversations from the event.  This piece, from participant Yaakov Chaliotis, discusses our conversation on AI implementation:

“Let them fail.”

At Strategic’s Innovation Sprint in London on 17 September, our working group landed on an idea that deserves a fairer hearing than it usually gets. Let organisations pursue AI in whatever way they believe will work. When that approach falls short, professionals should be ready to assess what went wrong, show where human expertise makes the difference and help build something better. The group broadly agreed: failure creates an opening to show where human experience must sit alongside AI.

I support that thinking, however not for the reason some might assume. The opportunity is not to prove that organisations should have stayed away from AI. It is to help them discover how much more valuable AI becomes when its use is shaped by people who understand both the technology and the work itself and the specific expertise. 

Some leaders may need that encounter with failure before they are ready to listen to a different approach. Our responsibility is to make the encounter useful.

Failure challenges both overconfidence and scepticism

The group’s diagnosis was that the central obstacle is human rather than technical. Limited literacy, fear and leadership attitudes are pulling organisations towards two extremes: trusting AI far too much or dismissing it altogether.

Naivety takes either form. One leader assumes that access to a tool removes the need for professional judgement. Another sees a disappointing output and concludes that AI has little serious value to offer. Both miss the relationship between what a tool can do and how it is applied. AI does not produce intelligence, skilled people do, and the tool is only as good as the judgement behind it.

This is where failure can become productive. It exposes assumptions nobody had bothered to test: about what the technology could do, what instructions it needed, or how much expertise the task actually required.

Interpretation matters, however. Without expert assessment, an unsuccessful attempt simply hardens scepticism. The conclusion becomes “AI does not work”, when the more useful question is “why did this approach not work?”

The session left that distinction open. How do we separate AI error from human error, and what does failure actually look like in practice? For me, those questions are the starting point of the intervention, not reasons to abandon the proposition.

Diagnose first, then rebuild

The process I would build from this discussion starts with the organisation’s own experience. It has tried something, hit a problem and now has a concrete reason to reconsider.

The expert’s first task is to understand what happened. What was the intended outcome? What was delegated to AI? What information and instructions were given and their quality? How was the output assessed? Where was human judgement missing, and where could AI have contributed more?

That assessment should distinguish a limitation of the technology from a weakness in how it was used. It should also identify what worked. A disappointing outcome does not make every part of the attempt worthless.

The next step is to turn diagnosis into an adoption process: keep the useful capabilities, close the gaps and place human expertise where it adds value. This reflects the group’s principle of coexistence – human experience and intelligence working alongside AI, with people positioned exactly where the gaps are.

The expert therefore does more than repair an output. They help the organisation understand how to produce better work again and again, not just once. The intervention becomes the beginning of a different relationship with AI.

Build adoption around real work

Once the problem is understood, the conversation can move from broad enthusiasm or resistance to specific decisions. Which tasks should AI support? What should remain a human decision? Is the organisation deploying existing tools or developing its own capability? The group raised that last distinction explicitly, because it shapes what is needed in skills, data, privacy and governance.

I would use those questions to build a practical workflow around the work that actually needs doing: define the outcome, allocate responsibilities and agree what evidence is needed before the result can be trusted.

Literacy belongs inside that process. The session repeatedly called for training that is frictionless, role-specific and tied to real use cases, alongside a better understanding among leaders of AI’s limitations. After a failed attempt, that learning has an immediate purpose. People are no longer sitting through an abstract presentation about AI. They are examining their own experience and working out what to do differently.

The subsequent working drafts also propose named human review and clearer disclosure of what AI contributed and what people verified. These are proposals rather than standards agreed in the room, however they offer practical building blocks for a redesigned approach.

Demonstrate the value leadership could not see

The most important stage comes after diagnosis and redesign, and that is showing the improvement.

A sceptical leader needs more than reassurance that AI is powerful. They need to see how a better approach changes the work. An overconfident leader needs to see that human involvement is a source of value, not merely an additional cost.

I would go back to a task the organisation understands and assess the revised approach against its original objective. Is the analysis stronger? Is the communication more useful? Can the evidence withstand scrutiny? Has time been saved without weakening the result? That comparison gives leadership something concrete to judge rather than a slide about the future, and the conversation shifts from whether to believe in the technology to how to use it well.

This requires a constructive attitude. There is little value in arriving with “we told you so”. The opening should be: we can understand what happened, learn from it and build a more capable approach together.

Letting organisations fail does not mean engineering mistakes or ignoring foreseeable harm. It means recognising that organisations will make their own choices, and that their experience can create the conditions for meaningful change.

For communications professionals, the opportunity is larger than crisis repair. It is to help organisations move from unsupported experimentation to expert-led adoption.

Let them fail. Then turn that failure into evidence, learning and a better way of combining human intelligence with AI.

+++

Yaakov Chaliotis is the Founder and Managing Director of Group of Verified Intelligence, an AI-assisted strategic intelligence and research consultancy. He advises organisations on decision intelligence, verified research, reputation risk and the responsible use of AI in analysis. He previously held senior communications and analytics roles at Shell International and the UK Department for Education, where he worked across digital communications, insight and strategic decision support.