The Art of Efficiency

Talk to almost any leadership team right now and you'll hear a version of the same two sentences. "We know we can't afford to ignore AI." Followed, a beat later, by "we're not entirely sure what we're doing about it yet." That isn't a criticism. It's genuinely where most companies are. Call it the discovery phase: past denial, short of an actual strategy, spending money to find out what even works.

We've watched companies move through exactly this phase before, and it is worth being precise about what happened to the ones who treated it casually.

We have done this before

When the internet went from novelty to infrastructure through the 1990s and 2000s, the companies that hesitated didn't get a quiet pass. Borders spent years treating the internet as a passing trend, outsourced its entire online business to Amazon in 2001, and didn't build a web strategy of its own until 2008. It filed for bankruptcy in 2011. Blockbuster dismissed mail order and streaming as fringe ideas long enough for Netflix to build an unassailable lead, and it filed for bankruptcy in 2010. Neither company failed to notice the internet. They failed to treat adopting it as urgent, and the market did not wait for them to catch up.

That pattern shows up in the labour data too, not just the corporate-collapse headlines. The economists David Autor, Lawrence Katz and Alan Krueger tracked what computerisation did to wages through the 1980s and 90s and found it explained something like 30 to 50 percent of the growth in relative demand for skilled workers since 1970. Being fluent with the new tool wasn't a nice line on a CV. It measurably changed what people were paid.

PwC's 2025 Global AI Jobs Barometer found that workers with specialised AI skills earned a 56 percent wage premium in 2025, more than double the 25 percent recorded the year before. By its 2026 update that premium had climbed again. Productivity growth in the industries most exposed to AI has nearly quadrupled since 2022. The people and companies moving early are not just feeling good about it. They are being paid for it, in a way that is already measurable.

Adoption without a plan

None of this means most companies have actually worked it out. McKinsey's 2025 State of AI survey found that 88 percent of organisations now use AI in at least one business function, up from 78 percent the year before. Encouraging, until you look at what "use" means in practice: only around a third report scaling it across the enterprise, and just 7 percent say it's fully scaled. Only 39 percent of respondents could attribute any enterprise-level profit impact to AI at all, and most of those put the number under 5 percent.

MIT researchers, in a widely reported 2025 study of enterprise generative AI, found something starker still: roughly 95 percent of enterprise GenAI pilots fail to deliver measurable business impact. The interesting part isn't the failure rate. It's the reason. The researchers didn't pin it on model quality. They pinned it on a "learning gap": generic tools are flexible enough for individuals to get value from immediately, but they stall inside a company because they don't learn the company's actual workflows, exceptions, and edge cases. The pilots that did work shared a pattern: deep customisation to a specific workflow, rather than a general-purpose tool bolted on and hoped for.

Put those two findings together and the shape of the discovery phase gets clearer. Companies are not failing to try AI. They are trying it broadly and shallowly, and shallow adoption doesn't compound into anything.

"How good are you with AI?" is not a usable question

Here's where most companies get stuck next. They know adoption needs to be deliberate rather than scattershot, which means understanding what their own people can already do. So someone sends round a survey, or asks in a one-to-one: how comfortable are you with AI? It's a well-meaning question and it produces almost nothing you can act on. Everyone rates themselves somewhere in the middle, "comfortable" means something different to every respondent, and nobody involved learns whether the finance team can actually use AI to close the books faster or whether they've just used a chatbot to draft an email.

The scale of the variance underneath that vague question is real, and it is wide. BCG's 2025 AI at Work survey describes a "silicon ceiling": 75 percent of leaders and managers use generative AI several times a week, against 51 percent of frontline employees, and only 5 percent of all employees use it in ways that meaningfully change how they work rather than as a faster search box. The same research found 77 percent of employers planning to reskill their workforce for AI between now and 2030, against just 13 percent of workers who have actually received any AI training so far. And the gap isn't evenly spread: adoption is highest in technology and finance, and consistently lower in frontline and service roles, which is precisely where "ask people how they're doing with it" is least likely to surface anything useful.

That gap is also what creates the second trap. A company that can't measure its own workforce's capability is a company that can be sold almost anything. It's easy to end up hiring an outside "AI expert" whose entire qualification is comfort with a chat interface, no different from what plenty of existing staff already have, minus the years of domain knowledge, company history, and industry context those staff already carry for free. Go back to the MIT finding above: the pilots that actually worked succeeded because of deep customisation to a specific workflow. Nobody understands your workflow, your compliance quirks, or your customers' actual complaints better than the people already doing the job. The honest starting point isn't hiring a stranger who is fluent in prompts. It's working out, properly, what your existing team can already do, and what specifically they need to close the gap.

Efficiency isn't what you do after the discovery phase

Most of what companies spend on AI right now will get chalked up to R&D while they work out what actually earns its keep, and that's a reasonable way to categorise it. Spend during a discovery phase is expected to be somewhat exploratory. But "exploratory" is not the same as "undisciplined," and this is where a lot of that spend quietly goes to waste.

Look again at the numbers above. Only a third of companies in McKinsey's survey have reached the scaling phase. Only 5 percent of MIT's studied pilots delivered real impact. Those aren't random outcomes. They correlate directly with whether the company treated the pilot with the same discipline as any other engineering or operational work: clear scope, a real definition of done, and a habit of writing down what was learned so the next attempt doesn't start from zero. Efficiency isn't the reward for figuring AI out. It's the mechanism by which some companies figure it out at all, faster and cheaper than competitors who are also groping in the dark but keeping no notes on what they've tried.

What that discipline actually looks like

This is less abstract than it sounds. On FlowForge, our open-source workflow automation platform, every AI-assisted development session runs against a short, plain-text file called efficiency.md. It isn't a productivity slogan. It's a literal set of constraints the work has to fit inside, and it is worth looking at what's actually in it, because the pattern generalises well beyond code.

Before any task starts, the file requires a short, explicit scope: the goal, the exact files expected to change, the behaviour being added, the tests that need to pass, and, just as importantly, the non-goals: what must not be touched. That one habit kills the single most common failure mode in AI-assisted work, which is a small request quietly turning into an unreviewable rewrite.

It also keeps a running list of what the team calls "cost hotspots": specific, previously expensive mistakes, written down so they're never paid for twice. One entry notes that a particular filter pattern in the codebase silently fails to match certain records, and that this exact bug cost two rounds of debugging before someone wrote it down. It will never cost a third round, because now the fix is a line in a file instead of a lesson trapped in one engineer's memory. That's the entire value proposition of institutional knowledge, captured at essentially zero ongoing cost.

And it ends every task against a plain checklist for what "efficiently done" actually means: acceptance criteria met, only the required files changed, focused tests passing, nothing speculative built for a future phase that hasn't arrived yet. Efficient, in other words, isn't a synonym for fast. It's a synonym for scoped.

None of that is specific to software. A support team, a finance function, or a marketing department running AI-assisted work can keep the same kind of one-page, living document: what's in and out of scope for a given task, the specific mistakes that have already cost time once, and a clear definition of when something is actually finished. That document is cheap to write and gets more valuable every time it's used. It is also exactly the kind of asset a hired outsider cannot hand you on day one, because most of its value is your own company's memory of what has already gone wrong.