The Last Mile: Smallpox, Refrigeration, and where AI value is actually decided

The vaccine had existed, in one form or another, for 170 years. The science was settled, the strategy was written, the money was there.

And still the disease killed two million people a year, because a vaccine only becomes medicine when it arrives, cold and intact, in a village at the end of a dirt road.

What finally won was something wonderfully unglamorous called the cold chain: the relay of fridges, cool boxes and careful hand-offs that carries a vaccine from the factory to the very last mile without letting it spoil on the way. Warm vaccine is just liquid. The cold chain is what keeps the value alive all the way to the person who needs it; where the smallpox vaccine can either flourish into something miraculous, or be poured down the drain while the wasted bucks are counted.

As you know, I spend my weeks knee-deep inside AI programmes across sectors, industries, enterprise large and small, and it seems to me that the same last-mile problem is deciding all of them.

I'm writing this for the people who own those programmes: the transformation and operations leaders, the chief people officers, the CEOs of growing firms who signed off the licences and are now wondering, privately, what actually changed. The leaders thinking, are we using AI properly? Are we behind? Are we ahead? If that's you, you're in good company; read on.

The strategies are usually pretty good. I read thoughtful, board-approved AI strategies most months. The tools are certainly bought, for the most part now. Mostly the naysayers who couldn't see the value in AI are now seeing it's something that needs attention.

However, by most counts around two thirds of Copilot seats sit unused, which suggests the procurement worked perfectly and something after it didn't.

MIT's researchers found that around 95% of enterprise AI initiatives show no measurable return, and when I sit with the leaders behind numbers like that, almost none of them have a technology problem.

The value is real; it just never made the last mile. It spoiled somewhere between the boardroom and somebody's Monday morning.

And there is another number underneath that one. At last count, over 90% of people already use AI for their work, very often on free tools nobody sanctioned. So the value is not waiting to be introduced to your organisation; it is already inside the building, unrefrigerated. The question is not, then, whether AI arrives. Or even whether it was adopted. It is whether it arrives intact: within your rules, at your standard, visible to you.

Here is a story of an Eight Person Team

An investor relations team, eight of them, spent most of every week writing reports. We set things up so the reports drafted themselves, to a checked template, in about ten minutes.

Now, that moment can go two very different ways, and I've seen both. In one version the team is frightened; reports were the job, and the machine just did it. So people use it quietly and a bit badly, paste things into free tools nobody approved, and hope. Quality drifts, risk climbs, and the programme joins the 95% while the dashboard still shows green.

What actually happened was the other version. The team was delighted. They spent the reclaimed week talking to investors, which is the human work they joined the firm to do. Quality went up, the numbers went up, and, the part that never makes it onto a slide, their wellbeing went up too; we measure it, and it really moves. Up.

The difference between those two versions had almost nothing to do with the tool. It was everything that happened before the tool. Leadership in a room agreeing what this was for. A handful of plain documents: a strategy, a tool list, a usage policy, a plan for the first ninety days. Use cases sorted into tiers, so everyone knew where they could play freely and where a human must always sign: draft an email and lose thirty seconds if it's wrong; client work gets a human check; sensitive data and filings get real guardrails and a full review, every time. Rules people could actually see. Another word for all of that is trust, and trust turns out to be the temperature control on the whole chain.

The research backs this up from every direction, which is worrying but also in some ways comforting - because at least we know where the blockage on value is.

BCG puts only about 10% of AI's value in the algorithms and 20% in the technology; the remaining 70% sits in people and ways of working, which is precisely the part most budgets skip. The organisations that support their managers, the warm hands in the middle of the relay, see adoption above 90%, against the mid-thirties for everyone else. And the thoughtful voices in this space centre on the same idea from different angles: governance done well is not the brake on innovation, it's the traction; and guidance has to live inside the flow of the work itself, because that's the only place behaviour really changes.

There is one more discipline the successful programmes share, and it is the one I would press on gently: they measure the right thing. Licences bought is spend. Logins is attendance. Tokens are simply expenditure. What tells you whether the value survived the journey is capability: can your people use AI well, on a benchmark that stays current, and is that showing up as hours recovered and quality raised, team by team? If your dashboard tracks logins, you are simply measuring how many tickets were given out, not what they used them for.

Which brings me to the thing I most want to say. I don't think AI is overhyped. If anything I think it's underhyped, but hyped in the wrong place.

The noise is all at the factory end: the models, the launches, the demos.

The value is at the other end, in the last mile, in whether a real person doing a real job on a Tuesday works differently and better. A single well-built use case in the right hands can return hundreds of percent.

Most organisations are running a handful of shallow ones, and nobody can see the gap, because you cannot manage a last mile you cannot see.

This, honestly, is why we built Atheni.ai : a cold chain for capability. Your strategy, your policy and your approved tools go in once; every person's AI coach carries them, intact and translated into their own role, into their daily work the same day; and one screen finally shows you what arrived, measured as capability rather than logins.

If you take one thing from this, let it be the question I'd gently offer any leadership team right now: we know what our AI strategy says; do we know what arrived?

Walk the last mile of your own organisation this month, ask ten people at random what they did differently last week, and see what the answer tells you. My experience is that it's the most useful, and occasionally the most sobering, hour a leadership team can spend.

My co-founder, Louise Ballard (Moody) wrote thoughtfully on this yesterday here, considering the new recommendations from the AI and Jobs Taskforce, chaired by Martha Lane Fox, which were accepted by Sadiq Khan. Until the flow of work changes, we are simply scratching the surface with AI; and we don't know what we don't know.

And a piece of news, for the first time in one of these: after a year of building and a spring of beta cohorts, we open Atheni's doors this month. If you'd like to walk your last mile with someone who has done it a few dozen times, or simply see your own organisation's AI roll out on that one screen, reply to this or find us at atheni.ai; we are staggering access cohort by cohort on as we open, and I would love yours to be one of them.