The Pipeline Stage Was a Commitment Device
A stage-weighted forecast was never a measurement instrument. It was a commitment device: a named person asserting a probability to a manager who could ask…
Multiply the pipeline by a stage probability and the output looks like measurement. Nothing in it was measured. The probability was a convention and the stage was an assertion, made by a named person who could be asked about it on Thursday. Why is this at commit. Who signed off on budget. The forecast review was not a reporting meeting. It was an interrogation. The usual complaint is hygiene, which assumes the field was an instrument discipline could repair. A rep opened the CRM and picked a value from a list whose weights were set years earlier against deals that no longer exist. A pick from that list is a survey response, and putting a dollar sign on it is how a belief acquires the costume of arithmetic. Nobody measured a buyer. They measured a claim. The problem is not occasional. It is structural. A rep moves six deals to a later stage the day before the forecast call and the weighted number gains a few hundred thousand dollars. Assume one moved because a buyer opened a security review and five were housekeeping. The real event and the five acts of tidying are indistinguishable. The manager has one instrument for telling them apart. Asking. A model that reads login frequency across the buying committee, support contact volume, and movement on a redlined contract beats a picklist. A rep's bias points the same way across every deal they own, so summing forty deals sums forty copies of one distortion. A model gets each account wrong for its own reasons, and the errors partly cancel. The model wins. A stage-weighted forecast was a stack of assertions, each with a name on it. A behavioral forecast is a single output with a confidence interval and no author. When the model says a deal closes and it does not, nobody was wrong. The deal review that was an argument between two people becomes a conversation about feature weights. Accountability did not transfer to the model. It evaporated. Run the model as the forecast and keep the rep's number in a separate field, never blended into it. Two columns, not one. Say the model puts a deal at 62 percent and the rep calls it commit. Averaging makes the aggregate worse, because the rep number carries a directional bias that blending reintroduces. It makes the assertion worse, because a claim absorbed into a formula is no longer one anyone can be judged on. Store the assertion with a timestamp and the reason given, then score it against what happened. Over four quarters the record shows which commit calls hold and which rep runs three weeks early. Being wrong becomes a measured property of a named operator. A team closing eighty deals a year across six reps will never fit a model worth showing a board. There the second column is the instrument. The signals that matter live outside the CRM, and connecting them is Intelligence-layer work. Connection presupposes the source is emitting, and before a close it usually is not. Product usage during the trial. Support contact from someone senior. Whether a security review has been opened. Check whether any system at your company would know a champion changed jobs. Most learn it from the rep, after the deal slips. Nothing downstream ever asked those systems to publish. Product analytics answers questions on a dashboard no revenue system can query. The APIs exist and the events are reachable. What does not exist is a normalized stream landing against an account identifier. The identifier is the expensive part. Usage is keyed to a workspace, support to the address that opened the ticket, billing to the signing entity, the CRM to a name a rep typed. Those four keys do not join. The forecasting model is not the project. The instrumentation is the project. Build the model first and you get the stage-weighted forecast with more math on it. We build the instrumentation layer before the model, for the same reason every deployment we have run spent four weeks of architecture work for every one week of model work. The opening weeks of a revenue engagement go to event definitions and account resolution. The forecast is the last thing built.