The Hardest Part Isn’t the Math
The hardest part of forecasting is getting to the end of a planning cycle and having to go to leadership with bad news: the expense forecast is too high or the revenue forecast is going to fall short. The best case outcome of such a conversation is going back for another pass, while blame and finger pointing are not unusual responses. The forecast miss becomes more about the analyst who delivered it rather than the conditions that produced it, and the finance team shies away from surfacing bad news next time.
This is where AI-powered forecasting changes the dynamic, shifting the conversation from who is wrong to what the data is telling us. Most of what’s been written about AI in forecasting focuses on the obvious wins: speed and accuracy. The effect I want to talk about is quieter and, as far as I can tell, mostly unexamined: AI changes who delivers the bad news, and that changes everything else.
Why Bad News Gets Stuck
A forecast is never just a forecast, and never just the endpoint of an analytical process. It is as much a social artifact capturing commitments and assumptions, built upon the careers and credibility of the analysts who developed it. Is it any wonder, then, that people might be reluctant to share the result when it’s worse than expected? There are many possible reasons why: different assumptions or definitions across Sales, Operations, and Finance, data that resides with a single person (“We always ask Susan…”), and entire categories where everyone quietly ignores the official forecast. The end result is often a process that involves trying to avoid the difficult conversation until it’s impossible to do so any longer. Once a number is wrapped in that much social weight, it becomes much easier to delay the conversation than to have it, and easier still to quietly hope the next forecast cycle smooths things over.
The Real Culprit Isn’t a Person
The worst part of the finger-pointing isn’t who gets blamed – it’s that the blame is aimed at the wrong target. A forecast rarely misses because of individual human error, and the analyst delivering the number is, more often than not, just a messenger. The actual culprit is harder to point at: dozens of small factors stacking up quietly over months or years. Undocumented business logic that lives in one person’s head. A product hierarchy that got reorganized two years ago and never quite reconciled back through reporting. A metric whose name stayed the same while its definition drifted between business units, or between this year and last. Any one of these, on its own, is too small to notice. Together, they degrade the forecast in ways nobody sees until the number lands on someone’s desk and looks wrong. By then, of course, it looks like the analyst’s problem. It never was.
What Changes When the Model Delivers the Number
How, then, can AI change the conversation? The biggest shift comes when the source of the forecast becomes a model rather than an individual. Models have drivers and documented assumptions. They can also run continuously rather than once a cycle, which means bad news arrives in smaller pieces, earlier, when it’s still cheap to act on. The cumulative effect is a change in tone: from inquisition to inquiry. Instead of asking how a particular analyst missed a key factor, the discussion turns to why the model couldn’t see it.
A few things make that shift possible. The model becomes a neutral narrator. When the unfavorable number arrives, no one in the room delivered it. The drivers are visible and shared, so “you were wrong” becomes “these inputs moved.” And because everyone is working from the same model and the same inputs, the conversation starts from a common baseline rather than three competing versions of the truth. It’s a much healthier starting place.
The other meaningful shift is from a point forecast to a probabilistic range. Even the best forecast is “wrong” the moment it’s published; reality never quite cooperates. Replacing a single number with a range and a confidence interval changes what the conversation is even for: instead of arguing about whether the forecast was right, the room can talk about which scenario it’s preparing for.
What Has to Be True
AI doesn’t do this work on its own. To take full advantage, an organization has to put a few things in place – and most of them aren’t technology.
Start with ownership. Someone has to own the inputs, and someone has to have the authority to declare the output final. Without that, the finger pointing won’t disappear – it will simply relocate. Closely related is definitional alignment: you can’t align on a number that means different things to Sales, Operations, and Finance. And when humans disagree with the model (which they should, sometimes, because people often have context the model doesn’t), the overrides and the reasoning behind them need to be visible. An undocumented override is just a quieter version of the old problem.
The cultural piece is harder. Forecast post-mortems have to be treated as learning events rather than indictments, run with curiosity rather than the search for someone to blame. None of that holds if leaders still shoot the messenger when the number is unfavorable. In that case, AI hasn’t depersonalized anything – it’s just given everyone a new messenger.
Four Things That Actually Help
None of this is automatic, but a few moves consistently push organizations in the right direction. While none of these are dramatic, all of them are harder than they look.
- Invest in the “boring” layer: Documentation of business logic. Robust change management. These are not the flashiest topics, but they are what make the model work. The old adage, “an ounce of prevention is worth a pound of cure,” applies here, as keeping on top of the inputs is far easier than trying to fix the outputs later.
- Keep a close eye on overrides: Areas with frequent or consistent overrides are where conversations need to occur, and are often where there is the widest trust gap. Tracking overrides as a first-class metric is what makes the rest of this work.
- Normalize imperfection early: As stated earlier, even the best forecast is wrong the moment it’s done. The key question is whether forecasts are wrong in useful ways or in surprising ones. The failure mode here is subtle: organizations don’t usually punish forecast misses overtly. They just quietly stop trusting the analyst who delivered the last one. That’s the dynamic to interrupt.
- Run forecast retrospectives, not autopsies: After major outages, tech organizations often run no-fault incident reviews, where the goal is not to place blame but to learn where the system or process failed. Finance retrospectives need the same posture. Conversation should center on what drivers the model missed, not which person did. The goal is to improve cycle over cycle and learn for the next forecast rather than assign blame for the last one.
The Bottom Line
The promise of AI-powered forecasting isn’t a perfect number, it’s a less personal one. Even the best forecast will still miss; bad news will still happen. Expenses may run hot, revenue may fall short. What changes is what happens in the room when the news arrives: who delivers it, who gets blamed, and whether the team is set up to learn or to defend. When the model owns the miss, the team can own the fix.
Where in your forecasting process does bad news get stuck? What have you tried to unstick it?
If these challenges sound familiar, you’re not alone. At BDA, we work with organizations to improve forecasting through AI, stronger data foundations, and better planning processes that turn difficult conversations into earlier, more actionable decisions. Contact us to learn how we can help.