Oddbox found a better use for AI than making ads
When Oddbox's board asks why the company isn't making more ads with AI, Gastón Tourn says the next question...
Read more
A pilot is a controlled test: one team, one process, clean-ish data, someone senior watching. Production is the opposite. It means the AI runs every day, on real messy data, for people who did not design it and will not baby it, inside the workflows the business already depends on. The gap between the two is where most AI transformation actually happens, or quietly doesn’t.
And most of it doesn’t. In aibl’s State of UK AI Adoption Survey 2026 (755 UK mid-market leaders), only 14 per cent have scaled AI across three or more functions with a measurable return. Roughly half report a measurable return somewhere. Getting one pilot to pay back is common. Turning that into something the whole business runs on is rare. That distance is the central challenge of AI transformation, and it is worth being honest that a promising demo tells you almost nothing about whether you will cross it.
A pilot is built to succeed. You pick a friendly team, a tidy slice of data, a metric you already know moves. Production strips all of that away. The stall usually comes from four things the pilot never had to survive.
Data is the first. In a pilot you use a clean sample. In production the model meets the real thing: duplicates, gaps, records entered five different ways, the exception nobody documented. Accuracy that looked fine on the sample drops the moment it hits the full estate.
Then governance. A pilot needs almost none. Production needs a lot, because now the AI touches customer records, regulated decisions and money, and someone has to be able to answer for it. In the survey, measurable ROI climbs from 22.2 per cent in companies with no governance to 85.3 per cent in those where governance is mature and embedded. Same tools underneath, a 63-point gap in return. Governance is not the brake on scaling. It is the thing that lets you scale at all.
Third, capacity. A pilot runs on one enthusiast and a lot of goodwill. Production needs people who can keep the thing running, retrain it, spot when it drifts, and pick up the phone when it breaks. If that skill lives in one person’s head, the pilot cannot become a system.
Fourth, the work itself. A pilot sits beside the day job. Production has to sit inside it. If using the AI means opening another tab, copying data across and remembering a new step, people won’t, and the return you measured in the pilot evaporates the moment it is optional.
Put plainly: the pilot proved the AI can work. Production asks whether your organisation can run it. Those are different questions, and the second is harder.
The data has to hold up at full volume, not on a sample. Governance has to be real before you widen access, not bolted on after something goes wrong. The skills have to sit in the team rather than in a consultant who leaves. Change management has to start before go-live, because people need to know what is coming and why. And the AI has to be built into the systems people already use, so the right action is the default one and not an extra chore. None of these showed up in the pilot. All of them decide production.
Move in a deliberate order, because each step sets up the next.
Fix the data before you widen access, since bad data breaks everything downstream and does it silently. Stand up governance next, at least enough to answer who owns this, what it is allowed to touch and how you would catch a bad output. Then integrate: put the AI inside the workflow rather than alongside it. Build the team’s capability in parallel so you are not dependent on one person. Roll out to one group, watch it properly for a few weeks, and only then widen.
Speed helps, as long as it is speed with a spine. Companies that can approve a new AI tool in days report 62 per cent measurable ROI, against 38 per cent for those where approval drags on for months. Quick sign-off is not recklessness here. It usually signals an organisation that has already decided who owns AI decisions and how they get made, which is exactly what production needs.
The most advanced version of production is agentic AI: systems that take actions across steps and systems, not just answer a prompt. This is where the pilot-to-production gap bites hardest, and where it pays off most. In the survey, agents that work across departments report 84.7 per cent measurable ROI, while narrow single-task or retrieval agents sit at 24 to 29 per cent. The value is in the connectivity, not the cleverness, and connectivity is precisely the thing a sandboxed pilot cannot test.
Assume something will go wrong, because for almost everyone it already has. In the survey, 79 per cent of leaders have had at least one AI failure and only one in five reports a clean record. The most common failures are a data or security breach, a tool failure that broke a process, and unchecked output causing errors downstream.
So the goal is not a spotless record. A spotless record usually means you never left the pilot. The goal is to govern failure: catch it early, contain the blast radius, and know who is accountable before it happens rather than after. That is what separates the companies that scale from the ones that keep running safe little pilots that never touch the business. Treat the first failure as information, fix the specific thing that broke, and keep going.
Because pilots are built to succeed and production is not forgiving. The pilot used clean data, needed almost no governance, ran on one person’s enthusiasm and sat beside the real workflow. Production hits messy data at full scale, demands real governance, needs skills in the team rather than one head, and has to live inside the systems people already use. It is common to get a pilot to pay back and rare to scale it: only 14 per cent of UK mid-market firms have done so across three or more functions.
Plan for months, not weeks, and expect the return to build over a year rather than arriving on day one. The organisations that move fastest are not cutting corners. They have already settled who owns AI decisions, which is why they can approve a tool in days, and those companies report far higher returns than the ones where every choice takes months.
Treat it as normal and fix the specific cause. Nearly four in five leaders have had at least one failure, so a clean record is the exception, and it often just means the AI never really went live. Work out what broke, whether it was data, the tool, or an unchecked output, contain it, name who owns it, and carry on. Governing failure is what lets you scale. Avoiding it usually means you never left the pilot.
When Oddbox's board asks why the company isn't making more ads with AI, Gastón Tourn says the next question...
Read more
A quick one on research this week. I was interested to read a recent article from The Strategic Management...
Read more
Account-based marketing (ABM) for an AI vendor means choosing a short list of UK organisations that fit. You then...
Read moreGet ahead with the most actionable insights, playbooks and real-world AI use cases you can adopt right now, in your inbox every week