Mark Beard, Founder and CEO, Vertex Agility
I sat in a meeting recently with a very large blue chip organisation. The CFO asked a question that would have been unthinkable two years ago. Not “what is our AI strategy” but “we have spent seven figures on AI, show me where it is running.” The room went quiet. Everyone knew the honest answer was a list of pilots, a couple of impressive demos, and a Copilot licence bill.
That question is now being asked in boardrooms everywhere, and it marks a genuine turning point. The experimental budget is gone. What replaces it is a production budget, and production budgets come with production expectations.
The model stopped being the interesting part
Here is the uncomfortable maths for anyone who thought buying the right AI would be a source of advantage. Your competitors can buy the same model you can, on the same day, at the same price. Within eighteen months the leading models have become close enough in capability that, for most business work, the choice barely matters. Intelligence has gone the way of cloud compute. It is a utility. Nobody wins by having electricity. The days of watching what your competitors spend and trying to outspend them, in the expectation of faster or more efficient results, are over.
So where does advantage come from? From the one thing your competitors cannot buy off the shelf: the shape of your own organisation. Your systems, your data, your approval chains, your regulatory obligations, your customers. AI creates value when it operates inside all of that, and inside all of that is precisely where it currently struggles.
Why pilots die
Having watched this play out across a number of enterprise clients, I would put the failure pattern in three buckets.
First, the data was never ready. Most large businesses have five or six systems that disagree about what a customer is. A demo can quietly ignore that. Production cannot.
Second, nobody owned the risk. A pilot needs a sponsor. A production system needs someone willing to sign their name to what it does at 2am on a Saturday, and to explain it to a regulator afterwards. Those are different levels of commitment, and most programmes never bridge the gap.
Third, and this is the one people least want to hear, the workflow itself was never designed. Getting a model to draft a credit decision takes an afternoon. Getting that decision through role-based approvals, exception handling, audit logging and a change process in a twenty-year-old core system is months of proper engineering. The intelligence is maybe ten percent of the build. The other ninety percent is plumbing, and plumbing is not glamorous, which is why so few people talk about it and so few pilots survive contact with it.
What the next three years look like
If capability has stopped being the differentiator, some things follow that most companies have not yet priced in.
The second wave of AI spend will be larger than the first and it will go to fewer suppliers. Procurement teams are already starting to ask a different question in RFPs: not “what is your vision” but “show us three of these running in production, and let us speak to the people who own them.” Vendors with decks and no deployments are about to have a difficult couple of years.
Regulated businesses will pull ahead, not fall behind. Banks, insurers and utilities get called slow adopters. In my experience they are simply adopters who are not allowed to skip steps. Every AI action in those environments needs validation, approval and an audit trail before it happens. That discipline is expensive up front and it compounds later, because a system built to survive a regulator survives everything else too. The firms doing this properly today will be very hard to catch.
And internal teams will hit a wall, through no fault of their own. The people who understand your systems best are the people already running them. Asking that team to also re-engineer core workflows around AI, while keeping the lights on, is how good programmes stall for a year. Every client we work with has capable people. What they are short of is people who have already made this class of system work inside someone else’s legacy estate, several times, and know where the bodies are buried before the project starts digging.
The honest test
There is a simple test I offer anyone who asks how their AI programme is really going. Count the workflows, not the pilots. How many processes in your business does AI now execute end to end, with proper controls, that a named person is accountable for? If the answer is zero, you do not have an AI capability yet. You have an AI interest.
Q&A: Counting Workflows, Not Pilots
We have spent a fortune on AI. Why can’t we point to much of it running in production?
Because the hard part was never the model. Getting a model to draft a decision takes an afternoon. Getting that decision through role-based approvals, exception handling, audit logging and a change process in a twenty-year-old core system takes months of proper engineering. The intelligence is maybe ten percent of the build. The other ninety percent is plumbing, and that is where most pilots quietly die.
If everyone can buy the same models, where does our advantage actually come from?
From the one thing your competitors cannot buy off the shelf: the shape of your own organisation. Your systems, your data, your approval chains, your regulatory obligations, your customers. Intelligence has become a utility, like cloud compute or electricity. Nobody wins by having it. You win by getting it to operate reliably inside all of that.
Are regulated industries like banking and insurance destined to fall behind on AI?
In my experience it is the opposite. They are not slow adopters, they are adopters who are not allowed to skip steps. Every AI action needs validation, approval and an audit trail before it happens. That discipline is expensive up front and it compounds later, because a system built to survive a regulator survives everything else too. The firms doing this properly today will be very hard to catch.
Can’t our own team just build this? They understand our systems better than anyone.
They do, and that is exactly the problem. The people who understand your systems best are the people already running them. Asking that team to re-engineer core workflows around AI while keeping the lights on is how good programmes stall for a year. What they are usually short of is people who have made this class of system work inside someone else’s legacy estate before, several times, and know where the bodies are buried.
How do I honestly tell whether our AI programme is working?
Count the workflows, not the pilots. How many processes in your business does AI now execute end to end, with proper controls, that a named person is accountable for? If the answer is zero, you do not have an AI capability yet. You have an AI interest. Moving that number off zero is the whole game.
Working Through This With Vertex Agility
Moving that number off zero is exactly the work our AI Consultancy practice was built to do. Our starting point is deliberately narrow: a short, fixed-price engagement, in your systems and on your data, that finishes with three things. Something genuinely working in your environment. A plain-language statement of the real blocker, which is almost never the one written in the business case. And a costed, honest path to scale, including the option of stopping. Two to three weeks, no strategy phase and no discovery theatre – just the first workflow, done properly.
Because the ninety percent that decides the outcome is plumbing through legacy core systems, our Software and Data Consultancy practices carry as much weight here as the AI itself. Role-based approvals, exception handling, audit logging and change processes in twenty-year-old systems are architecture problems before they are AI problems, and architecture is where we start. We have made this class of system work inside heavily regulated estates before, which is usually the difference between a workflow that reaches production and one that stalls in pilot.
If your board is starting to ask where the AI is actually running, it is worth having an answer ready. The honest question underneath it is a simpler one: what would you want that first workflow to be?
Vertex Agility is a technology consultancy delivering production AI, data, cloud and platform engineering for enterprise clients, including some of the most heavily regulated environments in the UK.
Want an honest read on where your AI programme really stands? Our free AI Readiness Mini-Audit is a straightforward place to start.