truck8.ai

Leading AI Adoption: Why Pilots Don't Scale

Almost every company has run an AI pilot. Very few work differently because of it. That is not down to the technology or your people. It is down to where you put the pilot.

8 min readTom Mekenkamp

Why AI pilots don't scale

You may know the pattern. A team gets a few licences and a few weeks. The demo is convincing and everyone is enthusiastic. Six months later only the two people who set it up still use it, and an order takes just as long as it did last year.

That is not an exception. In BCG research among 152 CEOs of large companies (July 2026), 64% say their company pursues AI pilots. Only 26% embed AI as part of a broader business transformation. The companies that do get value from it are, according to the same research, roughly seven times more likely to redesign their workflows end to end.

That last point is the heart of it. A pilot rarely fails technically. It fails because it sits next to the real work and never touches the work itself. There are three reasons this happens so often.

1. The pilot improves one spot, not the flow

Copilot for sales, a chatbot for HR, a writing assistant for marketing. Every department looks more productive. And an order still takes weeks. Goldratt put it plainly: an hour saved at a point that is not the bottleneck saves nothing. The work simply waits an hour longer at the next step.

2. The waste sits between departments

Your customer experiences one company. You run it as departments. Between those departments, work waits, gets retyped, gets chased and gets translated from one system to another. That is where most of the waste sits, and that is where AI helps most.

A process inside one department can be improved by the head of that department. A flow across departments can only be changed by the board. That is why AI adoption is a board topic, not an IT project.

3. Every improvement needs a change in your systems

Making a step disappear almost always means the CRM has to pass something to the ERP, a form has to change, an integration has to be added. And that is exactly where it stalls. The team that can change systems is busy with maintenance, migrations and incidents. The pilot waits in the queue, like every other request.

The real bottleneck: your capacity to change your systems

Ask your leadership team this question: what share of your department's capacity goes to keeping things running, and what share goes to improving them? Have everyone write down a number first and only then say it. It is a standard question in my boardroom workshop, and in most boards the average is shockingly low.

That is no criticism. It is the logical result of a time when custom software was expensive. You bought standard packages, adapted your work to them and had IT keep them running. IT is stuck in maintenance as a result. Trapped, not lazy.

What has changed: an application for your own process now takes days, not months. 'The ERP can't do that' was a fact for years. Now it is a choice. Most constraints in an organisation are rules and habits, not laws of nature. Once you see that, you also see that the change team is already in house: the low-code builders, the integration developers, the people who know your data. They just never had the time.

The method is lean. The speed is new.

There is nothing mysterious about AI transformation. The method is plain lean: find the bottleneck, remove the waste, move on to the next bottleneck.

  • Find the bottleneck: where does work wait longest?
  • Apply AI at that point, not where it happens to be easy.
  • Remove the waste: make a step disappear rather than making it faster.
  • Next bottleneck: the constraint moves. Start again.

Don't automate waste

One round of that loop used to take months, because of IT lead times. Now it can take days. But only if you really make the step disappear. Bolt AI onto the old process and you get the same process, slightly faster, with one more tool to maintain. The Lead to Order case shows what that looks like in practice.

A model line instead of a pilot

Lean has a better idea than the pilot: the model line. You pick one real value stream and redesign it completely, from start to finish. Not as a trial next to the work, but as the work itself. The difference comes down to six things:

  • A real value stream that counts in the results, such as quote to order or complaint to resolution.
  • Fully redesigned, not one step automated.
  • One owner from the business with a mandate across departments. Not IT.
  • Live early, with real customers and real orders. Not a demo.
  • Measured on the flow: lead time, first-time-right and cost per order. Those measures outrank departmental KPIs.
  • The scaling decision is made up front: if the stream hits its targets, the money and the next stream are ready.

Who is on the team

A model line runs on a small, mixed team: people from every department the stream touches, two or three builders, someone who knows the data and a fixed contact in IT. That team has a mandate to change processes, roles and systems. Without that mandate it is a pilot after all.

The model line is also your training ground. Whoever takes part carries the way of working to the next stream. Your capacity to change grows with every round, instead of starting from scratch with every new application. Builders learn on the job, for example in the 4-week course Building with AI.

What the board does in AI adoption

An AI strategy for the board is not about which tool to buy. It is about three decisions nobody else can make.

1. Give the bottleneck an owner

Name the main bottleneck in your own words and put a name next to it. Usually the bottleneck is not a department but the handoff between two departments. That is exactly why it has no owner today.

2. Choose the value stream

Choose the stream for the model line. Not the easiest one, but one the customer notices and that shows up in the numbers. Give the owner a mandate across departments and make room in the team's diaries. Boris Cherny, head of Claude Code at Anthropic, put it this way in a 2026 interview: "It's up to leaders to create space for this. If you don't create space, you keep doing it the old way and have to force people to do it any new way."

3. Decide what happens to freed-up capacity

When a stream goes from three weeks to three days, capacity is freed up. Do you cut it, or reinvest it in the next stream, in customer contact, in growth? That is a board decision, and it helps to make it explicitly and in advance. Capacity to change is scarce: whoever reinvests speeds up the next round.

If roles change at scale, involve employee representatives early. In the Netherlands, the works council has an advisory right (Dutch Works Councils Act, WOR art. 25). Say so early and out loud, so nobody reads the model line as a reorganisation in disguise.

Governance that speeds things up instead of slowing them down

Whoever leads AI adoption also gets the questions about rules: what is allowed, who decides, what does the AI Act require? Those questions are fair. But governance that sends every application past five committees turns every model line back into a pilot waiting in the queue.

It works better the other way round. Agree up front who decides on which AI applications, which guardrails always apply and which data an agent may use. Then the model line team can build fast within those limits. How to set that up as a board is covered in AI governance for SMEs.

One guardrail deserves extra attention: an owner for every application. Cheap building also means lots of small applications appear quickly. Without an owner who maintains it and decides when it can go, you will have hundreds of unmanaged apps holding customer data a year from now.

Resistance: room and context, not reassurance

People rarely resist AI. They resist a change nobody explained and that they have no say in. Vague reassurances such as 'AI won't take jobs' make it worse. What works is being specific.

  • "It will take my job." Say what you will do with freed-up capacity, and say it before anyone asks. That decision sits with the board, not with the team.
  • "It isn't reliable enough." Often true. Let the team decide where human judgement stays. That is exactly the work AI does not take over.
  • "We already tried it." Ask what exactly did not work. Usually it was a tool on top of the old process, without the process itself changing.
  • "Legal won't allow it." Make clear what is allowed, what needs approval and what is never allowed. Clarity makes action possible. Training belongs with it: the AI Act already asks for it, see AI literacy.

How to start: follow one order through your company

You don't have to wait for a programme. Pick one stream this month, for example quote to order, order to payment, complaint to resolution or new hire to productive.

Put the steps on a wall, with a lane per department. Mark four things: where the work waits, where it gets retyped, where it is handed off to another department and where someone has to chase it. Keep it high level: eight to fifteen steps is enough.

Then ask one question: who owns this entire stream? The answer is usually nobody. That is where your model line starts. The wider route, from inventory to ongoing management, is in how to implement AI in your business.

Key takeaways

  • AI pilots rarely fail technically. They sit next to the real work and never touch the redesign.
  • The waste sits between departments. An hour saved at a point that is not the bottleneck saves nothing.
  • The real bottleneck is your capacity to change your own systems. With AI an improvement takes days instead of months.
  • A model line is one real value stream, fully redesigned, live, with one owner, flow metrics and a scaling decision made up front.
  • The board gives the bottleneck an owner, chooses the value stream and decides what happens to freed-up capacity.
TM

Written by

Tom Mekenkamp

AI consultant & founder of truck8.ai

15+ years leading transformations at AB-InBev, Royal BAM and beyond. Now building AI products and helping SMEs implement AI.

Find your own bottleneck, with your whole board?

In the Boardroom Workshop (2 or 4 hours) you follow one order through your own company, name the bottleneck and choose a model line. Within a week you receive the strategy document with the decisions and the first 90 days.

See the Boardroom Workshop