Why fluency, not tools, is the bottleneck

Here is an uncomfortable observation from three years of working inside enterprise AI programs: the tools have stopped being the differentiator, because everyone has them. Your organisation and your fiercest competitor can, this afternoon, buy access to essentially the same frontier models at essentially the same price. Whatever advantage exists cannot be coming from the technology, because the technology is a commodity you both rent.

And yet the performance gap between organisations using AI is enormous, and widening. Same models, wildly different results. Something else is the bottleneck, and I want to name it precisely, because the imprecise version — "it's a people thing" — leads to exactly the wrong interventions.

The bottleneck is fluency. Not familiarity, not enthusiasm, not having done the e-learning module. Fluency: the practised, judgement-laden ability to work with these systems on consequential tasks — to know what to delegate and what to hold, to specify work so it comes back right, to smell a confident wrong answer, to read a vendor's claims and know which are real. Fluency is to AI what literacy was to the printed word. The presses are everywhere. The readers are scarce.

What fluency actually consists of

Part of the problem is that organisations picture the skill as prompting — a knack for phrasing requests. Prompting is the visible sliver. The capability underneath, watching genuinely fluent operators across government, banking and my own engineering work, decomposes into something like five layers.

Calibration. Knowing, from accumulated contact, what these systems are reliably good at, where they degrade, and — critically — where they fail confidently. A fluent operator holds an internal error model: the model will be excellent at this, shaky at that, plausible-but-wrong at this other thing, so the verification effort goes there. Non-fluent users hold a binary — trust it or don't — and both poles are expensive. Blanket trust ships errors; blanket distrust re-does all the work and captures no value.

Decomposition. The mid-skill that most separates fluent from casual users: taking a real, messy piece of knowledge work and cutting it into machine-shaped and human-shaped pieces. The fluent analyst does not ask the model to "do the report". They hand it the synthesis of source documents, keep the framing of the recommendation, hand it the first-pass structure, keep the political judgement about what the audience can hear. This carving skill is invisible in demos and decisive in practice.

Specification. Fluent operators are good at saying what they want — which turns out to be a rigorous act of thinking, not a communications nicety. Vague intent produces plausible mush. The discipline of specifying context, constraints, format and what good looks like is the same discipline as briefing a capable new hire, and the organisations whose managers were already good at delegation are visibly faster at this. It is not an accident that fluency and management skill correlate.

Verification. Every serious use of AI has a checking regime, and fluency includes designing one proportionate to the stakes: what gets read line-by-line, what gets sampled, what gets tested against ground truth, what requires a second human. The non-fluent either skip this (and eventually ship the memorable failure that sets their program back a year) or apply maximal checking to everything (and quietly destroy the productivity gain they were promised).

Renewal. The layer nobody budgets for: the capability moves. Models improve, tools change, last quarter's limitation silently disappears. Fluency includes the habit of re-testing one's own calibration — noticing that the thing that failed in March works in September. Individuals and organisations alike accumulate stale fluency, confidently working around limitations that no longer exist. In a fast-moving capability landscape, the shelf life of "what these tools can't do" is about six months.

Look down that list and notice what it is made of. Judgement, decomposition, specification, verification, updating. These are thinking skills, not software skills. Which is why the standard enterprise response to the AI skills gap — buy licences, run a tools webinar — so reliably fails. It answers a judgement deficit with an access grant.

Why the bottleneck is worst at the top

Fluency is unevenly distributed in most organisations, and the distribution is upside down. Graduates and juniors, with the least to unlearn and the most time-per-dollar to experiment, are often the most fluent. The senior leaders making the consequential AI decisions — what to buy, what to automate, what the policy permits, what the strategy claims — are often the least.

This inversion has a specific cost. When decision-makers lack calibration, every AI decision in the organisation is made by proxy: by the vendor's claims, by the consultant's deck, by the most confident voice in the room, by analogy to the last technology cycle. I have watched procurement committees approve seven-figure AI platforms on the strength of demos that any fluent operator would have dismantled in ten minutes — and block low-risk, high-value uses because nobody in the room could size the actual risk. Both failures have the same root. The organisation's judgement about AI is concentrated exactly where its fluency is not.

You cannot fix this by surrounding senior leaders with fluent advisers, for the same reason you cannot outsource literacy to a reader who follows you around. The calibration has to live in the head that makes the call. A leader does not need to build systems. They do need enough hands-on contact — real hours, on their own work — to know what these systems feel like from the inside: where they astonish, where they slip, what supervising one demands. That threshold is lower than most executives fear. It is perhaps twenty deliberate hours. Almost none have spent them.

The economics of closing the gap

Framed as economics, fluency has three properties that make it the best AI investment most organisations can currently make.

It is the complement to everything else. Every dollar already spent on tools, platforms and data infrastructure is throttled by the fluency of the people operating them. Raising fluency raises the return on the entire existing stack — it is the multiplier input, not another additive one.

It compounds. Tools depreciate; judgement appreciates. A fluent workforce gets more from each successive model generation, because calibration transfers. And fluency spreads socially — one genuinely fluent operator on a team lifts the team, through example, shared patterns and borrowed confidence — which means well-placed investment propagates beyond its recipients.

And it is scarce in exactly the market where you compete for it. Fluent-and-experienced is the rare combination: plenty of AI-fluent juniors lack domain judgement, plenty of judgement-rich seniors lack fluency. The organisations that build the combination internally — by making their experienced people fluent, which is faster than making fluent people experienced — will not be able to hire the gap closed later at any reasonable price. This window is the cheap moment.

The practical implication is unglamorous: treat fluency like the serious capability it is. Assess it honestly, including at the top. Build it against real work, not modules. Protect the time. Expect the dip before the gain. And measure judgement, not usage — the question is never "how many people used the tool this week" but "who can now be trusted with a harder delegation than last quarter".

Fluency is Pillar I of our practice for a reason: it is the gate everything else stands behind. If your organisation's AI results do not match its AI spend, this is almost certainly why — and it is fixable. Book a workshop.