Find Out Why an AI Pilot Stalled Before You Spend More on It
Find Out Why an AI Pilot Stalled Before You Spend More on It
Say your team is piloting AI work for a weekly account-risk brief. Every Monday, before the leadership meeting, it should tell the room which customer accounts look shaky and what to do about them.
Three weeks in, there's still nothing in the meeting anyone can use. The sponsor, the person who argued for the pilot and owns its budget, has to decide what happens next. The first instinct is usually to add more money.
Sometimes that's the right call. But a stalled pilot can mean at least three different things. The work ran out of budget. It needed access to a system nobody had granted. Or it produced a brief that missed what the meeting needed. More money fixes one of those, at most.
In July 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. The reasons it gave were poor data quality, inadequate risk controls, escalating costs or unclear business value. Those are reasons to end a project, and two of them can start as the stalls above. A budget stall that nobody looks into turns into escalating costs. After a few Mondays of a brief the meeting can't use, the pilot's business value starts to look unclear. Before you fund another run, find out which of the three stalls you're looking at.
Start with where the work stopped
Before asking why, ask where.
Did the work reach the point where someone was supposed to review it? Did it pause at a limit, like a budget or a permission? Or did it fail before producing anything at all?
For the account-risk brief, those are different Mondays. On one, a brief exists and the meeting didn't find it useful. On another, the work stopped partway through pulling account data. On a third, nothing came back. Each one sends you to a different person for a different conversation.
When the budget ran out
If the work stopped because it used up its allowance, look at what it was doing when it stopped.
Maybe the brief was set up to review every account in the CRM each week, with full support history and usage data for each one. The meeting only needed the accounts renewing in the next 90 days. The run spent its budget on accounts nobody was going to discuss.
In that case, more money buys a longer run of the same mistake. A narrower assignment might cost less than the first one and give the meeting what it needed.
The other possibility is that the scope was right and the estimate was wrong. The FinOps Foundation's guidance on AI costs warns that forecasting may be much harder for new technology areas, and that funding for projects may need to be revisited more often until forecasts get more accurate. A first estimate for a pilot is a guess, and a stall can be the moment you learn what the work actually costs.
So ask what the run got through before it stopped, what was left, and whether the next run would need the same amount. From there the sponsor can fund more work, change the assignment, or decide the question isn't worth the extra spend.
When it needed access nobody granted
The brief also needs open support cases, because an account with three unresolved escalations is a different risk from one that just went quiet. Suppose the pilot can read the CRM but not the support system. The work stops there. Or it carries on and hands the meeting a brief with the support picture missing, which is harder to notice.
Money doesn't help here. A bigger budget reaches the same locked door.
Name the system and the action. Does the work need to read case status, or does it also need to change something, like updating a case or tagging an account? Then find the person who owns that system and can decide whether the access is appropriate. That's usually not the sponsor.
Don't grant broad access just to get the pilot moving again. OWASP's guidance on excessive agency in AI applications lists excessive permissions as one cause of damaging actions, and recommends limiting what an AI system is granted in other systems to the minimum necessary. It also recommends that a person approve high-impact actions before they're taken. For the account-risk brief, reading case status is probably enough. Nobody needs it closing tickets.
When the brief arrived and missed
This one is the hardest to spot, because something came back.
The brief lists every account whose product usage dropped last month, sorted by revenue. It's accurate. The meeting still can't use it. What leadership wanted was the three or four accounts that need a call this week, and the reason for each.
Compare the output with who it was for, the sources it was supposed to use, and what "good enough" meant. Often that last one was never written down. The pilot was asked for an "account-risk brief," and it produced a reasonable reading of those words.
The fix might be a revision with specific feedback: drop the full list, pick the accounts that need action this week, and say why each one made the cut. It might be a clearer brief, or better sources, like the support data from the last section. Or it might be the decision that the pilot hasn't earned another attempt, because the meeting has found a better way to get this information.
Write a short decision note after each stalled run
Whichever problem you find, write it down before deciding anything. A few lines is enough:
- What was the work meant to produce?
- What evidence do you have about where it stopped or fell short?
- Which cause is most likely, and what don't you know yet?
- What will you change next: scope, access, funding, sources, the brief, how the work gets reviewed, or nothing?
- When will you look at the next run, or decide to end the pilot?
Keep it as a note, not a score. One stalled run tells you something about that run. It doesn't tell you whether the pilot works.
The note also gives the sponsor a record of decisions that would otherwise live in someone's memory of a meeting. NIST's AI Risk Management Framework asks organizations to document who is responsible for managing AI risks, and to determine whether a system achieves its intended purpose and whether its deployment should proceed. For a pilot, a note after each stalled run is a practical way to make that call as you go, instead of once at the end when the budget is already spent.
Where Supanova fits
Supanova is an autonomous workforce, and the same diagnosis applies to work you hand its atoms. If the account-risk task stops partway, it shows why it stopped and the next actions available for that task. A retry is another attempt, not a promise the work will finish. Whether the question deserves more budget, or whether an access request is safe to approve, is still a decision for the sponsor and the owner of that system.
If the task finishes, the brief comes back for your review, so the sponsor can read it before anyone acts on it.
Don't spend more just because the pilot went quiet. Take one stalled assignment to its owner, work out which of the three conversations you're actually having, and make the next run earn its budget. You can start at supanova.team.