I'm being facetious, but the point stands.

Yes, some companies are getting real value from AI. Very few are....yet.  

Only 11% of  enterprise AI Agents are making it to production. McKinsey's State of AI puts the same gap another way: 77% of organisations have not scaled AI beyond experimentation.

AI is no more a silver bullet than it is one thing

Start here, because almost everything else goes wrong downstream of this.

Like nearly any technology, AI is not a single thing. Across the discipline you have machine learning, deep learning, symbolic and neurosymbolic approaches, classical optimisation, generative models. That's just a start.

If you walk into a board meeting believing that large language models are AI, you are starting in the wrong place. You will pick the wrong tool, apply it to the wrong problem, and then conclude that "AI doesn't work" when what actually happened is that you brought a hammer to a plumbing job.

Stop asking "what's our AI strategy?"

The question isn't how you implement AI. It's how you run your business so that it's still alive in three years.

If I were on the board, I wouldn't open with the AI strategy, any more than I would demand an internet or an electricity strategy. I'd open with: who is going to kill us the way Netflix ate Blockbuster?

What does that company look like? What can they do without thirty years of legacy tech drag and data mush? Where does their advantage actually come from? And what would we have to become to seize the advantage? The one thing you don't have is the cold start problem, so if you get his right, you are better off than the newbies. It’s yours to lose.

That's all business strategy. AI may or may not be part of the answer, but you cannot solve the job by starting with the tool.

Don't use a spaceship to do the school run

The opportunity isn't optimising one step in an existing workflow. Shaving 10% off a task that shouldn't exist is an expensive way to stand still. If you use a spaceship to do the school run, you've wasted a spaceship.

The opportunity is rewriting the process entirely, so that something which took days takes seconds. Those are the step changes worth chasing, and ones that are already being delivered and genuinely proving ROI.

Don't retrofit AI onto how you work today. Design for a world in which AI exists.

This matters more than it sounds, because it's also a maths problem. If your board has set a target of 3x return, or a material reduction in operating cost, you cannot get there through incremental gains on today's process. The arithmetic simply doesn't work. Transformational outcomes require transformational design.

AI is the system. It is not the transformation.

Any transformation has four components: people, process, systems and data. In this case, AI is the system. One of four.

What organisations keep trying to do is shoehorn the system in and hope the other three sort themselves out behind it (and this isn't new, we have been doing it for at least the 25 years I have been working on transformation).

But we seem to be even worse with AI. My hypothesis is that it's because you can log into ChatGPT and get value in about thirty seconds. Organisations then, unintentionally, map the immediacy of that consumer experience onto an expectation that implementing AI in the corporate world should be similarly frictionless. It never is. Implementing change inside a business, be it technology change, process change, any change, has never been that simple, and AI hasn't repealed that.

This is borne out by where the failures actually cluster. The root causes are structural: pilots built on infrastructure that was never designed for production, fragmented data estates, and success measured on model accuracy rather than business outcome. Model capability ranks last on the list of reasons AI projects fail. It is almost never the model.

So, taking the other three in turn.

Data: rubbish in, rubbish out

This is the single biggest reason AI cannot deliver.

If your data estate isn't in good nick, the likelihood of most AI solutions producing genuine return is close to negligible. No model, however capable, fixes data it cannot read.

Be brutally clear about where you actually are with your data. Then be equally clear about what has to happen for AI to read it, use it, and produce something you'd be willing to put in front of a customer or a regulator. That's not a pre-project inconvenience. It is the project (phase 1).

People: nobody does this on the side of a desk

Anyone who thinks AI change can be delivered by an existing team, in their spare time, with no additional budget, is smoking the wacky-backy.

It is not possible to upskill sufficiently, while doing your day job, to understand these systems well enough to know what good looks like, and then implement something that delivers real return.

And "knowing what good looks like" is something you cannot afford to underestimate.

Take AI coding assistants. They have dramatically increased the speed at which code gets written and shipped. But we are now seeing a spate of platform outages, including incidents tied to the use of AI coding tools.  

When you generate far more code, far faster, and the people reviewing it don't have the depth to judge whether it's right, you may get a productivity gain. But you also get the ramifications of failure.

The same pattern shows up everywhere AI gets deployed into a team that hasn't been trained to evaluate its output.

So: people need training, and they need protected time and budget to do this properly. If the board's target is a 3x return, the reasonable follow-up question is how much time and money are we prepared to put behind that? Because the honest answer cannot be zero.

Process: don't automate what you do today

Which brings us back round. Don't automate the process as it currently exists. Ask how you'd achieve the objective if you were designing from scratch, today, knowing what these tools can do.

Otherwise you get marginal gains, dressed up as transformation, measured against targets that were only ever achievable through genuine redesign.

And then there's the bit we've never had before

Everything I've described so far is recognisable. Hard, but recognisable; the same shape as every technology change of the last thirty years.

This is the part that isn't.

Suppose you do all of it right. You get your data into genuinely good order. You hire the right people and manage the change properly. You redesign your processes for the future state rather than paving over the present one. And you pick the right kind of AI for the job.

You can do all four and still fail, because the tool itself cannot be governed.

If a system can't tell you why it reached the answer it reached, and can't be relied on to reach the same answer twice, it will never get past compliance for tasks where precision and defensibility matter. In those cases, nor should it.

An MLRO put it to us just this week: it may not be a popular position, but if I can't control it, can't govern it and can't audit it, how do I know what it's actually doing? How do I know what risk we're accepting on the firm's behalf? How do we stand by any decisions we make? Decisions that, if wrong, could kill the company.

That's the gate. Everything upstream — data, people, process, tooling — is necessary, but none of it is sufficient. For a subset of tasks, where the outcomes matter, either because of GDPR, SMCR, the EU AI Act, Consumer Duty, HIPAA or other regulation, or just because the result matters, the black box will never be enough.

In a regulated business, explainability isn't something you bolt on at the end of the programme. It determines whether the programme ever goes live at all.

Fortunately, not all AI is black box. So, we now need to add control to people, process, systems and data. This makes "which AI?" a governance question, not merely a technical one.

Knowing what AI can actually do today

With any planning exercise you need to separate what these systems can genuinely do now from what they might do later. While that's partly crystal-ball gazing, we are starting to see real trends, and to know where progress is being made and where it isn't.

And the answer depends entirely on which AI you mean.

Generative, agentic AI is currently more likely to get a multi-step task wrong than right. On τ³-Banking, a benchmark for agentic performance in banking contexts, even the strongest models sit at just 55% a single attempt, and reliability degrades sharply when the same task is repeated (35%). That is not an acceptable error rate for regulated decisioning, be it in finance, healthcare or elsewhere.

And this is not a gap that scale alone closes. It is architectural. A large language model produces its answer by predicting likely output, not by following a chain of logic you can inspect. The explanation it gives you afterwards is generated the same way the answer was, it is not a record of how the answer was reached. That is why the same question can produce different answers on different days, and why a correction applied today carries no guarantee of holding tomorrow. You can log what an LLM did. You cannot reconstruct why it did it.

The FCA's concern specifically is auditability and reproducibility of outcomes at scale. It's central to the Mills Review published in July: the efficiency potential of agentic AI is real, but currently constrained. "A useful answer is not enough if the basis for it cannot be reconstructed." (Mills, 2026)

There are tasks large language models are excellent at, and tasks they are not. But generative AI isn't the only kind of AI, and the Mills Review points to alternative approaches that may be better suited to jobs where outputs need to be precise and defensible, including neurosymbolic systems, as a route to AI whose decisions firms can actually evidence and govern.

So, if I were on your board

I'd be asking:

• What problem are we solving, and what are the success criteria for solving it?

• What does good look like, how is it measured, who owns that number, and when do we check it?

• To get there, what do we need to deliver against data, control, people, process and systems?

• Are we redesigning the work, or just doing today's work faster?

• If our answer is wrong, do we need to be able to reconstruct how we reached it?

• Given that answer, which kind of AI is right for this task, and who in this room can tell the difference?

• Who's doing this, what are they stopping in order to do it, and what's the training budget? Or do we need to bring specific skills in-house?

Used the right way, the opportunity for AI is the most significant we have seen in a generation. This is well understood.

But it's not plug and play, and your job as a board member is to make the company the best version of itself, not beat your exec into doing the right thing in the wrong way.

About Onteric

Where today's agents guess and hope, Onteric's prove and commit.

Our proprietary AI is built from the ground up for precision, transparency and defensible outcomes, not retrofitted onto generative AI. Because logic, not probability, is the primary reasoning component, our AI and AI agents are controllable, auditable and governable.

We have supported over 95,000 end users through client platforms including Confused.com, First Central and Mojo Mortgages, delivering 3.5x ROI, 23%+ conversion uplift and over an hour saved per case in mortgage processing.

If you have a workflow that would be transformed by AI, but you need to be able to trust the output, get in touch.

‍