Every week, somewhere in a SaaS company, a VP Revenue is asked: "What's the number?" And every week, they give one back. $3.2M. $4.7M. Whatever the pipeline roll-up said when they last looked at Salesforce. The board writes it down. The quarter ends. The number is wrong by 18%. The explanation begins.
This is not a forecasting problem. It's a framing problem. The question "what's the number" presupposes that revenue outcomes for a given quarter are knowable to single-point precision, and that if you're not hitting that precision, you're doing something wrong. That presupposition is false. Revenue outcomes have irreducible uncertainty, and pretending otherwise doesn't remove the uncertainty — it just removes the mechanism for having an honest conversation about it.
What a Confidence Interval Actually Means in Revenue Forecasting
A P10/P50/P90 forecast isn't a hedge. It's a description of the distribution of likely outcomes given what you know right now. The P50 is the median expected outcome — the number you'd hit if you ran the quarter a hundred times and took the middle result. The P10 is the outcome at the 10th percentile — things go worse than expected, some late-stage deals slip, a large renewal comes in smaller than modeled. The P90 is the 90th percentile upside — deals close early, an expansion comes through that wasn't in the model, nothing slips.
The width of the P10-to-P90 band tells you something specific and useful: how much genuine uncertainty exists in your pipeline right now. A narrow band — P50 of $4.0M with P10 at $3.7M and P90 at $4.3M — means you have high-confidence closed pipeline and your model has good predictive signal. A wide band — P50 of $4.0M with P10 at $2.8M and P90 at $5.4M — means you have a lot of late-stage deals with high face value but uncertain close probability, and your board should plan accordingly.
The VP Revenue who walks into a board review and says "I'm forecasting $3.7M–$4.3M for Q4, with $4.0M as the median" is giving a more honest and more useful forecast than the one who says "$4.2M" and then has to explain the $3.8M actual result six weeks later. They're also having a fundamentally different conversation — one about probability and risk rather than one about optimism.
Why Single-Point Forecasts Break Under Pressure
There's a social dynamic at work in single-point forecasting that makes it reliably inaccurate. When a forecast is a single number, it becomes a commitment. Reps shade their estimates based on what they think their manager wants to hear. Managers roll up to something defensible for the VP. The VP anchors on something achievable that won't make the board uncomfortable. Each step in that chain introduces systematic bias — typically optimism bias in growth-focused organizations, which means single-point forecasts tend to be set too high and then missed.
The consequence is that organizations develop a forecast culture where the number submitted each week reflects negotiated expectations rather than actual pipeline probability. When we ask RevOps teams why they submit the number they do, a common answer is something like "that's what we need to hit" or "it's what the business plan says." Neither of those is a forecast. They're targets — which is a fine thing to have, but a different thing entirely.
We're not saying you shouldn't have targets or that a single committed number for operational planning is wrong — it clearly isn't. Planning for headcount, marketing spend, and capacity requires a number to build a plan around. What we're saying is that the forecast and the target should be explicitly distinct objects, and conflating them produces a culture where actual pipeline probability never gets surfaced honestly.
How to Build the P10/P50/P90 From Your Pipeline
The mechanics of building a probabilistic forecast require a few things. First, you need deal-level close probabilities that are model-derived rather than stage-mapped. Most CRM configurations map probability as a fixed percentage to each pipeline stage — Stage 3 = 40%, Commit = 75%, etc. These are blunt instruments. What you want is a probability for each specific deal that reflects its actual signals: engagement velocity, deal age, competitive situation, rep forecast history, product usage if it's an expansion, billing behavior if it's a renewal.
Second, you need to account for correlated risk. If five of your large deals are all waiting on procurement approval at the same company type (say, public sector or large financial institutions), they might all slip together if that procurement category has a seasonal slowdown. Treating each deal as independent when they share systematic risk factors will produce a P10-P90 band that's too narrow — you'll look more confident than you should be.
Third, you need historical calibration. A model that consistently produces P90 estimates that the actual result never exceeds is too conservative — your P90 should be exceeded roughly 10% of the time, by definition. Calibration testing on your own historical quarters will tell you whether your model is over- or under-confident, and in which direction it tends to err.
In practice, this means your forecast model needs to be trained on your historical data, not on generic industry benchmarks. Deal velocity and conversion rates vary substantially by product, go-to-market motion, deal size, and segment. A model trained on someone else's patterns will be wrong in systematically predictable ways that are specific to your company's dynamics.
What Changes When You Present a Range
The first time a RevOps team presents a P10/P50/P90 forecast to their board, the common reaction is something like discomfort. Boards are used to single numbers and tend to anchor on the P90. This is worth preparing for explicitly.
The reframe that works: a confidence interval doesn't mean you don't know your business — it means you're being precise about what you know. A weather forecast that says "70% chance of rain" is more useful than one that says "it will rain" or "it won't rain." The 70% gives you enough to decide whether to carry an umbrella. A P50 forecast of $4.0M with a P90 of $4.5M gives your board enough to decide whether to approve an incremental headcount hire contingent on hitting the upside scenario.
Over time, presenting ranges also changes the conversation from "why was the forecast wrong" to "was our forecast well-calibrated." If your actual results consistently land between your P10 and P90, that's a sign your model is working, even if any given quarter missed the P50. That's a healthier conversation to have — one that builds trust in the forecast process rather than eroding it every time reality diverges from a single aspirational number.
The Commit Call Problem
One common objection to probabilistic forecasting is practical: ops and finance need to plan against something. You can't build a budget around a range. This is true, and it's why we don't advocate for eliminating the commit call — the number the team is actually planning to execute against. What we advocate is separating the commit from the model output.
The model says: based on current pipeline signals and historical patterns, P50 for Q4 is $3.9M. Your VP Revenue then applies business judgment: we have a strong sales development motion this quarter, there are two late-stage deals the model doesn't have full visibility into, and the team is more motivated heading into year-end. Commit is $4.1M.
That's a legitimate exercise of judgment on top of a model baseline. What's not legitimate is arriving at the commit call first and then reverse-engineering confidence. The model output should be the anchor, with human judgment applied as an explicit adjustment rather than substituted for it entirely.
There's a version of this where the commit call and the P50 are the same — where the team has high confidence, the pipeline is well-qualified, and no significant adjustment is needed. That's fine. The point is that the P50 is the ground truth the business is planning against, not a number arrived at by managerial negotiation.
Revenue forecasting has always been probabilistic. Modeling it as deterministic doesn't change the underlying reality — it just removes the language for talking about risk honestly. A range is not a sign of uncertainty about your business; it's a sign of precision about what you actually know.