AI efficiency trap

Every boardroom now has an AI slide. It promises a once-in-a-generation reset: leaner operations, faster decisions, reinvented industries. Some of that may prove true. But anyone who sat through the 1990s will feel a flicker of recognition — and a quiet warning about the AI efficiency trap. We have run a version of this experiment before, under a different name: Business Process Reengineering.

Press rewind. The thesis in one image: a powerful new technology, and a pattern we have run before.

We have seen this movie before

In the 1990s, Business Process Reengineering was the dominant management idea. Enabled by enterprise software and the first wave of digitisation, organisations set out to redesign their processes from the ground up. The ambition was bold: strip out inefficiency, streamline operations, and lift performance by an order of magnitude.

I know, because I was there. The energy was real, the consultants were everywhere, and the promise — do more with radically less — sounded a lot like the promise being made about AI today.

What reengineering actually delivered

Plenty of companies booked real gains. Firms that reengineered targeted processes reported cost reductions of 20–50% and cycle-time reductions of up to 90% (Hammer & Champy, 1993; Davenport, 1995). But the record was uneven. Analyses from CSC Index and others at the time put the share of initiatives that fell short of their goals somewhere between half and two-thirds (Hammer & Stanton, 1995).

The deeper problem was not the failure rate. It was that the gains refused to stay proprietary. By the early 2000s, the productivity from IT and reengineering was visible across the economy, but firm-level differentiation had largely eroded. As Erik Brynjolfsson and Lorin Hitt showed in Beyond Computation, IT-driven productivity tended to diffuse across whole industries, steadily draining its value as a source of advantage (Brynjolfsson & Hitt, 2000). What started as an edge became the price of entry.

AI is following the same trajectory

Today’s AI rollout is tracing a strikingly similar arc. Across industries, organisations are using AI to automate tasks, accelerate workflows, and lean out their cost base. And, as in the 1990s, the early productivity data is genuinely good.

These improvements are real and worth having. But notice what they have in common: they make existing ways of working faster. They optimise the current business. They do not, on their own, change what the business is or how it creates value — and that distinction is where the trouble starts. (For the underlying field experiments, see Generative AI at Work, Noy & Zhang in Science, and the GitHub Copilot study.)

The efficiency trap

Improving how work gets done is important. It is not the same as changing what a company does. When an organisation pours its AI energy into efficiency, it is doing incremental improvement of its current operations — valuable, but bounded. As the tools become widely available and the best practices spread, rivals close the gap. What briefly looked like an advantage settles into standard practice.

This is the AI efficiency trap in one line: companies get faster and leaner without becoming meaningfully different. Productivity rises; relative advantage erodes. You can win the optimisation race and still finish exactly where you started, only with a tighter cost base and tougher competitors.

Efficiency makes you faster. It rarely makes you different.

Adoption is not transformation

There is a growing habit of treating AI adoption as if it were transformation — assuming that deploying the tools will, by itself, produce better outcomes. That is the exact misconception of the ERP era, when firms believed installing the enterprise system would reinvent the organisation. It did not. Technology amplifies the strategy, capabilities, and behaviours already in place. Point it at a clear strategy and it compounds; point it at confusion and it scales the confusion.

The cost-cutting parallel

One more rhyme is worth naming. In the 1990s, reengineering frequently translated into headcount. Between 1990 and 1995 the United States went through what was widely called a “jobless recovery,” as restructuring contributed to millions of layoffs, concentrated in middle management and administration (U.S. Bureau of Labor Statistics; Cappelli, 1999). It flattered the quarter, but it often hollowed out institutional knowledge and flexibility.

A similar pattern is taking shape with AI. The data is early, but many companies are already using it to run with leaner teams. Cutting headcount without redesigning the wider system tends to introduce new fragility — and to quietly cap the organisation’s capacity to innovate later.

Who actually won last time

The organisations that truly benefited from the reengineering era were the ones that refused to stop at process optimisation. They used the new technology to rethink their value propositions, launch new products and services, and build new business models. They did not just run the existing business better; they expanded into opportunities the old model could not reach.

The same fork is in front of every leadership team now. AI can automate tasks — or it can reshape how value is created and delivered, enabling new forms of customer interaction, new service models, and new ways of organising work. The technology supports both roads. Only one of them leads somewhere durable.

From better to different

Realising that potential takes more than a tooling budget. The instinct to ask “how can AI improve our current processes?” is fine as far as it goes; the more valuable question is “how can AI let us operate in a fundamentally different way?” Answering it well depends on a few non-technical ingredients.

1. Strategic clarity. Decide where AI should change what you sell, not only how you produce it. Start from a deliberate view of where the business is heading.
2. Leadership alignment. Reinvention crosses functions; it stalls without a leadership team that agrees on the destination.
3. New capabilities. Know honestly where you stand before you scale. An evidence-based assessment beats a confident guess.
4. Permission to experiment. Durable advantage lives just outside today’s business model — and you only find it by redesigning the wider operating system, not a single process.

The risk is that most organisations will not make the shift. They will adopt AI mainly for efficiency, bank the incremental gains, and discover they have arrived somewhere more competitive but no more advantaged — the efficiency trap, sprung on schedule. The opportunity belongs to the few who take the broader view.

Artificial intelligence does not guarantee transformation. It creates the possibility of it. If the focus stays on doing the same things better, the outcome is predictable: more efficiency, less differentiation. If AI becomes the catalyst to rethink what you do and how you create value, the outcome can be something else entirely. Reengineering left us a clear lesson — efficiency buys short-term results, but lasting advantage comes from reinvention. The only open question is whether this generation of leaders applies it, or repeats it.