The bottleneck after the pilot
Why working AI solutions don't take hold in daily operations, and three levers that change that

At a glance
- Many AI solutions work technically and still never become part of processes, roles, metrics and operations.
- Impact comes when the process is redesigned around the solution, a person in the business answers for the benefit and targets apply across unit boundaries.
- These questions belong before the pilot, because they determine what the pilot is supposed to prove.
The capabilities of AI models are growing fast, while the value in companies grows more slowly. We call this gap the AI delta. It closes when companies catch up in five areas, and this article deals with one of them, embedding. It covers solutions that work technically yet never become part of processes, roles, metrics and operations.
How to recognise missing adoption
- Pilots are presented and praised, then little happens.
- Once the project ends, nobody feels responsible for running the solution.
- One step in the process has become faster, but the end-to-end lead time is unchanged.
- Business units wait for IT; IT waits for requirements.
- The benefit appears in presentations, but nobody has measured it.
None of these signs can be fixed with a better model. A better model can even widen the gap, because it promises more than the unchanged process can absorb.
A typical pattern
The following pattern is a composite of typical situations and doesn't describe a client project. A business unit tests an AI solution that reads incoming invoices and suggests postings. The pilot goes well, and processing time per invoice drops noticeably. Three months later the solution is still in pilot mode.
The reasons lie outside the technology. IT hasn't budgeted for running it, because the pilot was paid for from an innovation budget. Nobody checks samples regularly, because that task isn't in anyone's job description. Payment approval still follows the old rhythm, so faster capture never reaches the supplier. And what was measured was time per invoice; nobody knows the lead time to payment. Each of these obstacles can be solved, just not by the project team alone.
What research says
The idea that technology alone is rarely enough isn't new. In 2022 Boston Consulting Group described the "10-20-70" rule of thumb: about 10 per cent of the effort goes into the model, 20 per cent into data and technology and 70 per cent into new processes and the way units work.
The economists Brynjolfsson, Rock and Syverson offer an explanation. General-purpose technologies such as AI need complementary investment in new processes and skills, which is hard to measure and takes time. In the early years, therefore, mostly cost shows up, and the benefit follows later. They call this the productivity J-curve.
Studies of work with AI also show that the value depends on who uses it and how. In the customer service operation of a software company, an AI assistant raised the number of issues resolved per hour by 14 per cent on average. For novices it was 34 per cent; for the most experienced, hardly anything. For adoption, this means roles, onboarding and quality control have to change along with the solution, or part of the possible value stays on the table.
The common view and ours
On the weight of processes and organisation, we agree with the common view. Where we differ is timing. The usual approach is to build the pilot first and talk about roll-out and change afterwards. The pilot then proves that the technology works, not that the new process pays off.
| Common view | Our view |
|---|---|
| Build the pilot first, then talk about roll-out and change. | The future process, the owners and the metric are settled before the pilot begins. |
| The pilot succeeds when the technology works. | The pilot succeeds when it shows that the new process pays off. |
| Change management accompanies the end of the project. | Change is part of the solution and has its own budget in the business case. |
| The benefit is assessed after roll-out. | The baseline is measured before the pilot, so the benefit can be proven later. |
Illustration: Paice
Three levers
Redesign the process
An AI solution dropped into an unchanged process speeds up one step. Whether the whole process gets faster depends on the rest. A simplified example makes this clear. AI drafts a quote in minutes, approval still goes through the weekly meeting, and the lead time stays at a week. Only when approval is reorganised too, for example with fixed limits up to which the team decides for itself, does the gain reach the customer.
Redesigning the process means working out who does what, and when, once the solution is in place. That's more effort than introducing a tool. But this is where it's decided whether the solution has an impact.
Assign ownership before the pilot starts
Every AI solution needs two owners: a person in the business who answers for the benefit, and a team that runs, monitors and develops the solution. Both are named before the pilot starts, together with a budget for operation. If either is missing, the solution remains a project that ends on the last project day.
On top of that come tasks that didn't exist before. Someone checks samples, someone handles the cases the solution can't decide with confidence, and someone flags when quality drops. These tasks need names, time and a place in the job description, or nobody does them.
Co-determination is part of this. Under Section 90 of the German Works Constitution Act (BetrVG), the works council must be informed in good time about the planning of working methods, explicitly including the use of AI. In good time means while the process can still be shaped. A works council that is involved early can help design the new process so that people accept it.
Align targets and metrics
If an AI solution works across unit boundaries, target-setting becomes a problem. An agent in customer service changes the work in sales; a better forecast in purchasing changes warehousing. If targets and budgets are measured by unit, one unit carries the costs while the benefit lands in another. Shared targets untie this knot.
The metric itself should be one the business understands, such as lead time, error rate, cost per case or customer satisfaction. Usage figures for a tool show whether it gets opened. They say little about its impact.
The baseline matters. It is measured before the pilot, over a period that covers the usual fluctuations, such as month-end or holiday periods. Without it, you can only claim later that the solution improved something. With it, the claim becomes evidence that even the units carrying the costs will accept.
What follows
An AI strategy that only describes technology and use cases leaves adoption to chance. It needs a chapter on processes, ownership and metrics, and that chapter belongs at the start of every initiative. In practice, that means settling five things before a pilot begins.
- 01Sketch the processWhat the work looks like with the solution and who does what in future.
- 02Name the ownersA person in the business for the benefit, a team for operation.
- 03Set the metricA business measure and its current baseline.
- 04Plan operationBudget, support and monitoring for the time after the project.
- 05Involve the works councilInform it under Section 90 BetrVG while the process can still be shaped.
Illustration: Paice
Three questions for your next meeting:
- Which of our current pilots have a named person in the business and a budget for operation?
- For which solution have we measured the baseline we'll use to prove the benefit later?
- Where does one unit carry the costs of a solution whose benefit lands in another?
Sources
- Five Rules for Fixing AI in Business, Galley, de Bellefonds, Duranton, Boston Consulting Group, 25 Feb 2022
- The Productivity J-Curve, Brynjolfsson, Rock, Syverson, American Economic Journal: Macroeconomics, 2021
- Generative AI at Work, Brynjolfsson, Li, Raymond, NBER Working Paper 31161, 2023
- Section 90 BetrVG, Federal Ministry of Justice, gesetze-im-internet.de



