Are you being served?

"Are you free, Mrs Slocombe?" "Do I look free, Captain Peacock? Of course I'm not free; I'm busy dealing with this AI malarky which will not work."

I was talking to someone recently about their AI-powered recruitment process. They were genuinely proud of it. AI writes the job ads, another AI screens the applications, a third shortlists candidates. Righto, but what are the applicants doing?

Same thing, obviously. Their AI writes the cover letter. Their AI tailors the CV. Their AI optimises for the ATS keywords that the employer’s AI is scanning for. Average applications per listing jumped from 41 to 184 in the space of a few months (per ACS data via Skill Society), driven by a mix of economic pressure and AI making it trivially easy to apply for everything. But the bottom line is at some point, someone, some actual human, still has to read them. And the ones that get through are painful, because they’ve been written to satisfy a machine, not to communicate with a person.

We’ve built a system where AI talks to AI and a human sits in the middle, wondering what happened to their Tuesday. Nobody stopped to ask whether this process should exist as it stands… or at all?

Faster horses

This isn’t only recruitment either. It’s a pattern I’m seeing everywhere. Organisations (and departments) everywhere are bolting AI onto existing processes and calling it transformation. Not redesigning. Not rethinking. Just accelerating whatever was already there, including the bits that weren’t working. Hi 2026, it’s cloud transformation from 2018… You may have forgotten… but this didn’t work last time.

McKinsey’s 2025 State of AI report puts a number on it: only 21% of organisations have redesigned even some of their workflows around AI. The other 79% are layering it on top of what they already had, and the results reflect that. 88% of organisations now use AI in some form, but only around 6% are seeing meaningful impact on their bottom line.

That gap is not a technology problem. It’s a design problem. BCG found that organisations that actually redesign their operations, using AI as an enabler rather than a bolt-on, achieve 2x the revenue growth and 40% greater cost reductions for future-built firms vs. laggards. Meanwhile, 65% of mid-market companies have deployed at least one AI tool in the last 18 months without touching their operating model. Fewer than 1 in 4 report measurable improvement. That’s a lot of money spent making the horse faster.

AI fills time. It doesn’t save it.

This is the bit that I think leaders need to sit with. HBR published a study in February tracking 200 employees at a U.S. tech company over eight months. 83% said AI had increased their workload. Not decreased it. They took on more tasks, switched between them more often, and drifted into 12-hour days without anyone asking them to. AI made it possible to handle more, so they did, and the volume expanded until it consumed every hour available.

I see this in my own work (despite my best efforts). AI removes friction, which means you can say yes to more things, which means you do, which means you’re busier than before. In my case, I’m stoked to say that I’m doing far more fun and fulfilling work, but I do appreciate that I am very much in the minority.

PwC surveyed 4,454 CEOs across 95 countries and found 56% say they’ve gotten “nothing out of” their AI investments. Economists are drawing parallels to Solow’s 1987 computing paradox: “You can see the computer age everywhere but in the productivity statistics.” That lag took 15 to 20 years to close. We might be standing in the middle of the same fog right now, except this time we’re moving faster and understanding less.

The problem underneath the problem

Here’s what I keep coming back to, in my consulting work and in building The Write Stuff: you cannot effectively automate a process you haven’t properly understood (and I don’t mean the actual coding for automation, everyone should have a crack at vibe coding).

If the workflow is unclear, the data is messy, or nobody actually owns the outcome, automation doesn’t fix it. It just breaks it faster and makes it riskier.

MIT found in 2025 that 95% of enterprise AI initiatives fail (their words… I won’t go into the definition of failure here…!) Not necessarily because the technology is wrong, but because the foundations aren’t there. Unclear governance. Poor data quality. No change management. No process redesign. The 5% who succeed invest in understanding the end-to-end system before they go near the technology. Gartner predicts 60% of AI projects will be abandoned by the end of 2026 because the data isn’t AI-ready, not just clean, but structured and contextualised for the way AI actually needs to use it. And boy howdy have I learnt some lessons with The Write Stuff.

The formula that keeps showing up across multiple studies is roughly 10% algorithms, 20% data and technology, 70% people, processes, and cultural change. Most organisations are pouring money into the 30% and hoping the 70% sorts itself out. It doesn’t. Also, when are we going to ditch this ratio mix?!

The room is not one person

Most organisations roll out AI training as a group exercise. Everyone gets the same workshop, the same tools, the same “here’s how to use Copilot” session. Then leadership wonders why adoption is patchy. It was the same with MS Teams roll out and SharePoint before that (speaking from personal experience).

BCG’s 2025 research found that tailored, persona-based learning journeys deliver higher adoption rates than broad-based approaches. They identified five distinct personas: champions who run ahead, independent explorers who figure it out alone, organisational adopters who follow the group, passive observers who watch from the edges, and cautious sceptics who resist until they see proof. More than 85% of employees sit in the middle three. Blanket training doesn’t shift them, because it doesn’t meet them where they are.

I keep coming back to what I call “deficit first.” Before you train anyone, figure out where their individual gaps are. What does this person actually need? What’s their work style? Where are they stuck? Then build from there. Not a group mandate. Not a one-size rollout. An honest assessment of what each person needs to move forward.

Paul Gibbons used the word “tinker” in a recent interview on AI adoption, and I think this is critical for both leaders and practitioners to understand. You can’t wait for someone to hand you the standard operating procedure; knowing more about what’s possible is more likely to become confusing and overwhelming. Take a minute out of your day and have a think about a task or piece of work that brings no joy, and have a crack at using AI to help you in your context, to enhance your way of working. The organisations that create space for that will build genuine capability. The ones that hand everyone the same playbook will build dependency.

What I’d actually do

Stop automating processes you haven’t analysed. If the workflow was broken before AI, AI will break it faster and more expensively.

Start with the system, not the tool. Map the end-to-end process. Ask who this actually serves. Ask whether the output is for a human or for another machine. If it’s machine-to-machine with a human awkwardly wedged in the middle, question whether the process needs to exist in that form at all.

Invest in the 70%. People, process, culture. The technology is genuinely the easy part. The hard part is shifting how humans work, and that takes individual attention. Not a town hall. Not a lunch-and-learn. Real, specific coaching and support for real, specific people.

Give people permission to tinker. OpenAI’s own research shows a 6x productivity gap between AI power users and everyone else. That’s not a technology gap. It’s a practice gap. The people pulling ahead are the ones experimenting, failing, adjusting, and trying again. They’re not waiting for the manual because the manual doesn’t exist yet.

And be honest about what AI can’t do. It can’t fix a broken process. It can’t replace systems thinking. It can’t tell you whether the thing you’re automating should exist in the first place.

---

Sources:

- McKinsey, State of AI 2025
- BCG, From Potential to Profit: Closing the AI Impact Gap (2025)
- BCG, Strategies to Tackle the AI Skills Gap (2025)
- HBR, AI Doesn’t Reduce Work, It Intensifies It (Feb 2026)
- PwC, Global CEO Survey (2025)
- MIT, Why 95% of GenAI Pilots Fail (2025)
- Gartner, AI-Ready Data (2025)
- Skills Society / ACS, Too Many Applications (2025)
- OpenAI, Productivity Gap Report (2025)