AI forecasting: best, base, and worst cases in minutes

Why a single forecast is the wrong output

Ask a forecast for one number and it will give you one number, and that number will be wrong. Not because the model is bad, but because the future contains variance and a point estimate contains none. The damage isn't the wrongness itself; it's what a single number does to planning. Teams anchor on it, spend to it, hire to it, and when reality lands 15% below, every decision built on the point estimate needs unwinding at once.

A range does the opposite. "Revenue lands between $95k and $130k monthly by December, most likely around $112k" tells you what to prepare for at both ends. You size fixed commitments against the downside, keep options ready for the upside, and stop treating a forecast miss as a surprise. Precision is what you have about the past. About the future, you have a distribution, and three scenarios are the minimum honest sketch of one. The point isn't to predict better; it's to be prepared at more than one point on the curve.

Defining the three cases

Base: most likely, not most hopeful

The base case is what you genuinely expect if current trends continue: trailing growth rates, observed churn, actual collection behavior. The most common scenario-planning failure happens right here, quietly: the "base" gets built from targets instead of trends, which makes it an upside wearing the wrong label, and then the real downside never gets modeled at all. A useful test: if hitting your base case would make you want to celebrate, it's not your base case.

Best: what has to go right

The upside isn't "everything goes up 20%." It's a specific, named set of things going right: the enterprise deal closes, the new pricing holds, churn drops because the onboarding fix works. Naming the drivers keeps the case honest and, more usefully, tells you what to watch: if the named things aren't happening by mid-quarter, you're not on the upside path, whatever the revenue line says this week.

Worst: the planning case, not the panic case

The downside is a plausible bad year, not an asteroid strike: growth stalls, your biggest customer churns, collections stretch by two weeks. It should be uncomfortable and survivable, because its job is to answer a specific question: what would we do, and when would we know? A downside so catastrophic that the answer is "nothing would help" teaches you nothing. The planning case is the worst outcome you can still steer through.

Which assumptions to flex, and which to hold constant

The commonly missed discipline: change only a few variables between cases, and hold everything else identical. If your three scenarios differ in growth AND pricing AND costs AND hiring AND churn, you have three unrelated models, and when reality arrives you won't be able to tell which assumption drove the divergence. Flex the two to four drivers with real uncertainty. Keep rent, existing salaries, and known contracts the same in all three. The cases should be siblings, not strangers.

The drivers worth flexing

For most SMEs the uncertainty concentrates in a handful of drivers. Typical flex ranges look like this (calibrate to your own history, not to this table):

Two things about this table. Collection timing barely improves in the best case on purpose: customers rarely start paying faster because your product is doing well, so modeling an upside there is fantasy. And hiring is a driver you control, which makes it less an assumption than a dial; the downside case should show it dialed to zero, because that's the lever you'd actually pull.

What AI does here

Concretely, in a tool like Reportly, AI does four jobs in this workflow. It detects patterns in your actuals that manual forecasting tends to miss or oversimplify: seasonality across the year, the real (not contractual) payment lag per customer, the drift in your growth rate over recent quarters. It proposes base-case assumptions from that history instead of from a blank cell. It builds the three cases in minutes, because flexing drivers across scenarios is mechanical once the model exists. And it regenerates everything automatically when new actuals land at month-end, which is the step that manual scenario models almost never survive.

Now what it does not do. AI cannot know you're about to sign your largest-ever customer, that a competitor is launching against you next month, or that you've privately decided to change pricing. It reads history; it cannot read intent or news. Which means the division of labor is clean: the machine supplies the trend-based skeleton and keeps it current, and you supply the judgment, the named upside drivers, the strategic downsides, the context history doesn't contain. AI forecasting that skips your judgment is just extrapolation with better marketing.

Using scenarios to make decisions

Scenarios pay for themselves through trigger points: pre-agreed, measurable thresholds tied to pre-agreed actions. The format is a plain sentence:

• "If we're tracking below base for six consecutive weeks, the two Q3 hires wait."
• "If churn exceeds the base assumption for two months, the retention project jumps the roadmap."
• "If we're above best case on pipeline by end of April, we bring the second salesperson forward."

The reason to decide these in advance is that mid-quarter, with real numbers wobbling, every decision gets argued from scratch and optimism usually wins the argument. A trigger point moves the decision to a calm moment and reduces the mid-quarter question from "what should we do?" to "did the threshold trip?" Review triggers against actuals monthly, as part of the same close-and-review loop as your variance analysis.

Keeping scenarios current

Scenarios built once and never refreshed are worse than none: decisions keep anchoring on assumptions that reality has already contradicted. Tie the refresh to the monthly close, since that's when new actuals exist. The monthly pass is short: re-anchor the base on updated trailing data, check whether any named best-case drivers have actually happened (if the deal closed, it moves from upside to base), test whether the downside is still plausible or needs a new threat, and read the trigger points against the new actuals. Fifteen minutes when the model updates itself; the reason manual versions die is that this same pass starts with two hours of re-pasting actuals. For cash-specific scenarios, the same monthly rhythm applies to your 13-week cash flow forecast, which is effectively a short-horizon downside model run weekly.

Frequently asked questions

What is scenario planning in finance?

Building multiple versions of a financial forecast under different assumption sets, typically best, base, and worst cases, so decisions are made against a range of outcomes instead of a single estimate.

How many scenarios should you build?

Three. Fewer gives no range; more than three or four and nobody can hold them in their head, which means nobody uses them.

How accurate is AI financial forecasting?

For trend-driven lines over a quarter or two, generally as good as or better than manual extrapolation, because it captures seasonality and payment behavior consistently. It cannot foresee discrete events like a big deal or a market shock, so its output needs your judgment layered on top.

What's the difference between a forecast and a budget?

A budget is a fixed plan set at the start of the year and used as a yardstick; a forecast is a living estimate of what will actually happen, updated as actuals arrive. Scenarios are variations of the forecast, not the budget.

How often should scenarios be updated?

Monthly, tied to the close, when fresh actuals land. Quarterly at absolute minimum; annual scenario planning is a strategy offsite, not an operating tool.

If rebuilding three cases by hand is what's kept you at one fragile forecast, see how Reportly's budgeting and forecasting generates and refreshes all three from your live actuals. SaaS teams can see churn- and MRR-specific scenarios on the SaaS industry page.