
The pilot has worked. Someone has put a tick next to innovation, there’s a promising demonstration for the next leadership meeting, and the team would quite reasonably like to know whether they can keep using it. Then comes the question that ought to have been written into the brief at the beginning: what exactly are we funding next?
I’d be wary of any process where a successful pilot automatically becomes a rollout. A trial can prove that an output is possible while leaving most of the operating costs, awkward dependencies and customer consequences unresolved. Before signing the next proposal, I’d want to understand which uncertainty the experiment has removed and which ones we are about to pay to investigate.
Write down what the pilot actually proved
Consider a hypothetical marketing team that has built an assistant for preparing campaign briefs. It can produce a sensible draft using approved product information. That is useful evidence about a bounded task. It tells us rather less about what happens when the product information is contradictory, the campaign crosses several markets or the person who designed the assistant is on holiday.
The next decision should describe the finding precisely. “The assistant produced acceptable drafts for these tasks, using these inputs, with this amount of review” gives a sponsor something to assess. “We’ve proved AI works” gives them a conference slide.
I’d ask the team to preserve examples of success, failure and borderline output. Include the cases where they abandoned the tool. A result that only survives careful selection is still worth learning from, provided nobody mistakes it for evidence of dependable everyday performance.
Cost the version that ordinary colleagues will use
A pilot often borrows expertise from people whose time disappears from the apparent cost. Somebody chooses the inputs, corrects the instructions and quietly rescues the result before it reaches a reviewer. Those activities belong in the operating model, even if the person doing them happens to be enthusiastic enough to work through lunch.
For the next business case, I’d record preparation, production, checking, correction and maintenance. Then ask who will do each job once the trial team returns to its normal responsibilities. If the business cannot provide that capacity, it has a resourcing decision to make before it has a scaling opportunity.
A support arrangement matters too. Someone needs to own changes to source information, access, instructions and the underlying tool. An assistant that has no maintenance owner is a fairly imaginative way of creating tomorrow’s support ticket.
Choose how much freedom the task needs
It is tempting to turn a promising assistant into an agent because that sounds like progress. The additional freedom should have a job to do.
Anthropic’s engineering guidance on effective agents distinguishes predefined workflows from systems that choose their own processes and tools, and recommends starting with the simplest approach that meets the need. That is vendor engineering advice rather than independent evidence of a return, but it raises a useful design question: which decisions genuinely need to be delegated?
In our hypothetical briefing workflow, a controlled sequence might be sufficient: gather approved inputs, draft, check, send for review. Giving the system access to more tools or permission to take customer-facing action adds another set of decisions about authority and recovery. I’d want the expected benefit of that freedom written down before anyone enabled it.
Make the next funding decision smaller and sharper
There are several respectable outcomes from a pilot. Stop because the problem is not valuable enough. Continue because the evidence is inconclusive. Fix a dependency before testing again. Fund a limited operational deployment because the case is credible.
Each outcome needs a reason. “We’ve already spent money on it” is a description of the past, however often it appears in a recommendation about the future.
A useful next stage might involve one team and one recurring task, with agreed input boundaries and a review date. Decide what would justify further investment and what would cause the team to stop. Name the person who can make that decision. The larger rollout can wait until this stage answers something commercially important.
Bring an operational owner into the conversation
The sponsor who approved experimentation may have no authority to change how the work runs. Before progressing, bring in whoever owns the process receiving the output. They need a say in what acceptable looks like and the capacity to act when the revised workflow creates a new bottleneck.
That conversation can expose an uncomfortable result: the technology is promising, but the proposed change is not a priority. Finding that out before a wider purchase is a useful outcome. It is considerably cheaper than learning it through twelve months of polite non-adoption.
I’ve written about the broader foundations in why AI adoption stalls. At the post-pilot stage, I’d use the AI readiness assessment to identify questions the next decision needs to settle, rather than treating its score as permission to scale.
If several teams have pilots waiting for that decision, the AI Adoption Alignment Review can help compare the opportunities and the evidence behind them. Bring the demonstrations, but bring the maintenance work and abandoned attempts as well. They will probably tell us more about what deserves funding.