AI, Interpretation, and the Widening Gap Between Idea and Value

This article is a slight deviation from the usual — not insights from a single theme, but a pattern I’ve heard repeated plenty of times this week that it felt worth writing down.

I’m not normally one for cultural commentary but when the same story shows itself from different directions, something is happening. Of course, I may be completely wrong, but the AI craze has got me wondering what line item on a spreadsheet people are looking at.

I want to be direct and honest about something before we get into it. I use AI. I explore it with clients. I think it’s one of the most interesting tools available to leaders and teams right now. This is not a warning against it. It’s a more precise concern — about a specific mistake I’m seeing organisations make, and what it’s costing them.

The spreadsheet wasn’t lying. It was just looking at the wrong thing.

This story starts with a small software development team piloting AI. A team of six to be precise; five developers, one tester. They were working well and shipping value, and were asked to dabble with AI.

Management discovered "agentic" AI and asked for it to be rolled out within this team before scaling across the organisation. The excitement was genuine, for everyone, sadly for some, that didn't last.

Within a short period, agentic AI was coding and building and creating and testing software. Three developers are let go. The team settles at two developers and one tester — and AI. On a spreadsheet, this makes complete sense – a 50% reduction in head count (sad times).

The management team rejoiced. Costs reduced, same amount of code (if not more) and fewer people on payroll. As I write this I grimace a little at the human cost of this, but a business that can "make things" without people is appealing to some.

However, a few weeks later came the big change. The AI provider switched from fixed monthly/annual pricing (which was already heavily subsidised) to token-based usage.

The team, like many using AI, had no idea how many tokens their work was actually consuming until they got the invoice.

Costs increased by roughly ten times - yes, you heard that right. Still, the belief held. The stance was in place. The economics, previously calculated on a spreadsheet, were now quietly ignored. It was now costing them more to use AI than with the three programmers in the team they let go.

What happened next is where it gets instructive.

AI made mistakes — not occasionally, but consistently enough that the remaining developers had to build additional code to check the AI code and do plenty of verification type activity, plus more model training, context management etc. The tester was busier than ever checking what was being built was indeed right.

Every new AI model release meant rebuilding parts of the working context around the system.

Data privacy and compliance checks added overhead nobody had budgeted for. The cost of each mistake from the AI was now billed directly (when the AI makes a mistake you still pay for the token usage).

Faster does not mean more value

Here is what I find most interesting.

AI could generate code (and tests) faster than any individual person in the team. But it was now taking nearly twice as long to ship features. Not code — features. Work that created actual value for someone outside the organisation.

If you know the idea to value system, you know what that means. Everything between the idea and the value is cost. When that gap widens, costs increase — not just in money, but in time, attention, and organisational energy, plus the delay in bringing about value.

The spreadsheet that originally made the case for AI was now telling a different story. Hence, it was being ignored.

I heard this same story from multiple people this last two weeks. Many are now hiring back. All are re-framing how they use AI.

This is what Rory Sutherland talks a lot about - the pursuit of cost reduction and efficiency is often measured and celebrated, without bearing the consequences of value degradation as a knock on effect.

What the ones getting it right are doing differently

There are teams making this work. They are worth paying attention to, because they don’t invalidate the pattern above — they clarify it.

The teams succeeding are not reducing headcount en-mass. They are redeploying it. Developers who were writing boilerplate are now doing architecture, review, and judgement calls — the work that requires someone who understands the context of why the thing is being built, not just what it should do. And of course, this requires human intelligence, noticing and understanding.

They model token costs as a variable upfront to be factored in, not a surprise. They treat prompt engineering and context management as a genuine discipline. And they measure success at the value end of the funnel — features shipped, outcomes delivered, financial value being returned — rather than at the generation end, or activity, or lines of code, or speed to build things.

The difference is not the tool. It is the understanding of where the tool sits in the system.

The interpretation problem

This is the thing I don’t think enough people are accounting for.

AI agents are interpreters. They are not the code, not the data — they sit above all of that, and they interpret what you’re asking for before attempting to give it back. That’s a significant and under-appreciated gap. I once heard someone describe AI as a goldfish with a good notebook. And we humans are now spending our time updating the notebook. (As an aside, is updating a notebook as meaningful as learning through doing the work itself?)

Anyhow.

Interpretation requires context, judgement, and an understanding of meaning. When a developer briefs an AI agent, they are not issuing a precise instruction to a deterministic machine. They are communicating intent to something that will attempt to translate that intent into action. And every translation costs. Every round of interpretation carries cost — in human attention, organisational energy, and token consumption — long before value is realised.

When you place that interpretation layer between your people and your value, you do not automatically save time. You add a new category of cost that didn't exist before.

But there's a deeper problem worth naming. Humans are also interpreters. They translate strategy/direction/needs into requirements, requirements into briefs, briefs into action — and at every hand-off, some meaning is lost or distorted. We have always known this.

Most organisational dysfunction is, at root, an interpretation problem: the gap between intention and understanding between people.

AI agents do not resolve that gap. They extend it. A developer briefing an agent must first have understood the requirements clearly themselves — must have interpreted the strategy, the user need, the constraint — before they can communicate it with enough precision for the agent to act on it usefully. If they haven't, the agent will interpret an already-imprecise brief, confidently, at speed, maybe in the wrong direction. This might not actually always be a bad thing, but worth noticing.

The question is whether the productivity on the other side is worth it. Sometimes it clearly is. The mistake is assuming it always is, without first asking where the interpretation gaps already exist in your system — and whether adding another interpreter into that chain makes them smaller, or larger.

What this is actually about

AI works well when it replaces cross-role workflows — the admin, the analysis, the tedious implementation that nobody would choose to do if they had an alternative – the grunt work. It works poorly when it replaces the people who hold the context and thinking of what the work is for.

The teams replacing headcount with AI agents, without redesigning how value actually flows, are not necessarily cutting costs. They are moving costs around, creating new ones, and, sometimes, widening the gap between their ideas and the value those ideas might generate.

That gap is not visible on a salary spreadsheet. It shows up in slower delivery, higher oversight overhead, eroded morale, and an increasing dependence on providers whose pricing models are, understandably, probably going to rise as they attempt to recover their own staggering investments, not to mention the lack of visibility sometimes into how tokens are used.

The leaders I know who are thinking clearly about this are asking a different question.

Not “how do I replace this person with AI?” but “where in my system does AI remove friction, create clarity, enable alignment – without creating new interpretation overhead, hidden costs and failure demand?”

That is a harder question. It requires understanding your system as a whole rather than looking at isolated cost lines. That understanding is, perhaps unsurprisingly, what I spend my time working on with clients.