Three months after handing over the AI mandate, the updates look fine. Meetings happen. The delegate is across it. The deck has grown. But something has shifted in the tone of the conversation, and you cannot quite put your finger on it. You have not seen a number that changed. You cannot point to a workflow that runs differently. You are not in the weeds, so you are not sure whether that is normal progress or a slow drift toward nowhere.
This is a familiar position for a founder who has handed off an AI initiative. The stall, when it comes, rarely arrives as a crisis. The signals are subtler than that, and they are readable from the top if you know what to look for.
Why do AI stalls rarely announce themselves?
Research from MIT NANDA, cited widely in AI implementation analysis, found that 95% of AI pilots show no measurable P&L impact. The failure mode for a delegated AI initiative is not a sudden crash but a sustained absence of result. Business as usual continues around it. The work proceeds, the expenditure accumulates, and the outcome never arrives.
BCG’s research on AI adoption tracks a consistent pattern across sectors, showing rising usage alongside flat business impact. Adoption metrics look healthy while outcome metrics do not move. For a founder reading updates from a distance, this creates a particular blind spot. The signals of a stall are second-order, things you will pick up in the language of the updates, the staging of the presentations, and the direction in which decisions flow, rather than in any single moment that announces itself as a problem.
The pattern is not unique to AI. A project that shows spend tracking to plan while the deliverable slips shows up clearly in a standard management review. A delegated AI initiative that shows activity tracking to plan while the outcome stays flat is harder to see because much of the activity is invisible to you from where you sit. You are not in the meetings, not on the tool, not talking to the team. You are reading signals at one remove.
What does a shift in update language tell you?
When a delegate first takes on an AI mandate, the language tends to be specific. They are testing this tool with this team for this outcome. Three to four months in, if the update language shifts from outcomes to activities, “we are building AI readiness”, “we are still in the exploration phase”, “we are aligning the team around the approach”, that is a signal worth noticing.
The distinction matters because both sets of language sound professional. The question is not the tone but whether a named outcome sits in the sentence. “We have tested document processing with the operations team and cut review time by 30 minutes per case” has a result attached to it. “We are making good progress on our AI readiness programme” has warmth and a clear activity but no result.
Update language that names what changed is a signal of a working initiative. Language that names what was done is, at best, neutral. A steady drift toward activity language over four to six weeks is one of the more reliable early indicators that the mandate has run out of momentum and the delegate is managing the narrative rather than the programme.
What does a more polished presentation actually mean?
There is an established pattern in AI implementation called AI theatre, where demonstrations get polished and prototypes grow impressive while no measurable outcome appears. The question to hold when reviewing any progress update is whether presentation quality has grown at the same rate as result quality. When the answer is no, the polish is the signal.
A founder who is not inside the workflow has limited ways to evaluate AI progress directly. The quality of the communication tends to substitute for the quality of the content. On the delegate’s part, this is a predictable response to the pressure of an initiative that is not landing as expected. The delegate works harder on explaining what is happening because explaining what happened is harder.
Look for the specific business outcome in every update, not the tool name or use case or test result, but the actual change in a real workflow with a number attached to it. If you cannot find that sentence in the last three updates, it may not exist in the programme either.
What does ownership silence look like from above?
When an AI mandate is working, the named owner owns it visibly, surfaces the blockers, and returns with conclusions rather than questions. When decisions that should sit with the delegate start coming back to you instead, or the named owner grows harder to pin down, the mandate has stalled somewhere in the chain.
Two ownership signals are worth watching. The first is a delegate who goes quiet on the initiative itself while remaining present and productive on everything else. That combination, effective on the day job but absent on the AI mandate, often means the AI work has become a source of discomfort, and the delegate is managing it by deprioritising it. Korn Ferry’s research on AI readiness in organisations documents this pattern; AI leadership gets handed to capable operators who lack the AI-specific competency for the role, which creates a gap that compounds as the initiative stalls.
The second is work that flows back upward. The founding principle of delegation is that the person who holds the mandate should be able to carry it without decisions queuing back to you. When vendor proposals land on your desk for sign-off, when team concerns about the initiative reach you before the delegate has addressed them, when you are being asked to approve something you expected simply to hear about, the mandate has not been properly held. That points less to the delegate and more to how the handoff was structured, and it is fixable, but only once it is named.
How do you raise it without pulling the mandate back?
The instinct when something feels off is often to step back in. That usually makes things worse. The delegate reads re-entry as withdrawn confidence, and the dynamic shifts in a way that is hard to recover. The more useful move is a single question that creates space for the delegate to name the problem themselves, before you have to name it for them.
Try asking what the single biggest improvement to the programme would be in the next thirty days, and what they would need from you to achieve it.
That question does three things at once. It surfaces the blocker the delegate may not have raised explicitly. It returns ownership of the solution to them rather than inserting you into the day-to-day. And it signals that you are paying attention without implying you have lost confidence in the person.
If the answer is coherent and specific, the initiative is probably fine and running slower than expected for legitimate reasons. If the answer is vague, or the question is met with surprise, you have found your signal. From that point you can address it directly, whether that means restructuring the mandate, adding a resource, or having a clearer conversation about what success actually looks like in measurable terms.
Spencer Stuart’s research on AI and CEO delegation notes that the founder who disengages entirely from the initiative and then re-enters with unrealistic demands once the gap becomes impossible to ignore is a more common failure pattern than a team that simply underdelivers. Staying in the conversation without re-claiming the mandate is the tighter line to walk.
