Process first, automation second, AI last

Multiply a well-designed process by AI and you get acceleration. Multiply a broken one and you get faster, more expensive failure at a scale that wasn’t previously available to you.

I say this as someone who’s thoroughly pro-technology. I’ve spent twenty years building software for laboratories. The argument isn’t that AI is overhyped. It’s that the sequence matters, and that most organizations are attempting step three without having done steps one and two.

This argument is older than you think

In 1990, Michael Hammer wrote a piece in Harvard Business Review called Reengineering Work: Don’t Automate, Obliterate. His line has stayed with me for years.

His observation was that firms were using computers to speed up processes they had never examined, and that speeding up a process can’t address a fundamental deficiency in its design. Bill Gates put the same idea more bluntly a few years later: automation applied to an inefficient operation will magnify the inefficiency.

Thirty-six years on, we’re having the identical conversation with a different technology.

What the current evidence actually shows

McKinsey runs an annual survey on AI adoption. The 2026 round covered more than 1,700 organizations across 97 countries, and the headline is sobering: only 37% could point to any impact on profit at all.

A much smaller group is genuinely getting value, the 6% that McKinsey classes as high performers. So what is that 6% doing differently?

They changed how the work happens. Around three quarters of the high performers had fundamentally redesigned their workflows, against roughly a quarter of everyone else. That’s the clearest single difference between the organizations getting a return on AI and the organizations simply buying it.

RAND published a careful piece of research in 2024 on why AI projects fail, based on 65 interviews across industry and academia. Two of their five root causes are squarely about foundations: insufficient data quality and quantity, and inadequate infrastructure for managing data. The others are about leadership misunderstanding the problem and about technology-led rather than problem-led selection, which is solutioneering by another name.

Deloitte’s intelligent automation survey has found, across four consecutive years, that the top barrier to scaling automation is immature and fragmented processes that can’t be managed as a unified flow.

S&P Global Market Intelligence reported in 2025 that 42% of organizations had abandoned most of their AI initiatives, up from 17% the year before, with the average organization scrapping close to half of its proofs of concept before production.

The sequence

In PURE terms, all of this sits in the Engineer phase, and the order within it isn’t negotiable.

First, fix the process. Remove steps that shouldn’t exist. Simplify the flow. Clarify who owns what. Resolve ambiguity. Get the human process right while it’s still cheap to change, because a process is a conversation and software is a contract.

Second, automate. Once the design is clean, identify what can run without a person. Automation of a good process compounds value. Automation of a bad one compounds waste, and it does so while making the underlying problem considerably harder to see, because now it’s inside a system that somebody paid for and has to defend.

Third, apply AI. As an amplifier at the end of a well-designed chain: better decisions, prediction, anomaly detection, continuous optimization. All genuinely transformational, all dependent on the two steps before them.

What this looks like in a laboratory

Take a procurement process where stakeholder needs are understood, supplier data is normalized, inventory is visible and compliance controls sit inside the workflow rather than bolted alongside it. Apply AI and you get intelligent sourcing, predictive replenishment, automated documentation, anomaly detection, spend optimization. Real acceleration.

Now take the more typical one: spend fragmented across spreadsheets and email, inconsistent supplier data, no inventory visibility, manual compliance checks, and three people re-keying the same order into different systems. Apply AI to that and you get faster generation of incorrect orders, reports built on unreliable data, and recommendations nobody trusts, correctly, because the inputs are wrong.

Same technology. Entirely different outcome, determined by what’s underneath it.