How to Reconcile a Rep Forecast With a Model Forecast

Once a company runs a statistical forecast alongside the rep roll-up, it inherits a new problem: two numbers, both defensible, that disagree. Most teams resolve this politically. The number that matches the target gets presented, and the other one gets described as directional.

That habit wastes the entire investment. The gap between the two forecasts is information, and worked properly it points at specific deals and specific broken assumptions.

Why do the two forecasts disagree in the first place?

They are built from different evidence, and each sees things the other cannot. A rep forecast encodes buyer conversations, procurement noise, and champion behavior that never touches the CRM. A model encodes the historical relationship between deal attributes and outcomes across thousands of opportunities, including patterns no individual rep has enough sample size to notice.

The disagreement is usually concentrated, not distributed. Two numbers that differ by 8 percent at the roll-up rarely differ by 8 percent across every deal. More often the gap concentrates in a handful of opportunities, and finding them takes one sorted report.

Start there every time. Reconciliation is a deal-level exercise that produces a number, not a number-level argument that ignores deals.

Which forecast should win, and when?

Decide in advance, by deal characteristic, based on where each has historically been more accurate. Deciding after you see the gap converts the rule into a preference. Two rows deserve emphasis. A deal that has moved from one quarter into the next is less likely to close even while sitting in commit, and reps carry these forward at full confidence because nothing about the deal looks worse from their vantage point. And the earliest warning on any opportunity is the absence of a signal, meaning no change in stage, close date, or amount. The rep rarely volunteers either pattern, which is exactly why the model call should carry the deal in both cases. See deal slippage for the underlying patterns.

What should the reconciliation process actually look like?

Four steps, run weekly, taking under thirty minutes once the reports exist.

Sort the deal-level differences by dollar gap and take the top ten. Those will typically explain most of the roll-up difference.

For each, classify the source of the disagreement. Either the rep holds evidence the model cannot see, or the model holds a pattern the rep is discounting. Force the classification, because "we just disagree" is where the process dies.

Apply the pre-agreed rule from the table. The point of setting it in advance is that nobody relitigates it in the room.

Log the outcome with the deal, the direction, and which side was applied. At quarter close, score both. After six quarters you will know which side wins in which situation on your business rather than on general principle.

What does a persistent gap tell you?

A gap that never closes points at a broken input, not at a difference of opinion.

If the model is consistently below the rep roll-up across every segment, check deal values first. Most opportunities close for less than the amount recorded in the CRM. A pipeline with an $80,000 average deal size producing $40,000 average closed-won deals will generate exactly this pattern, and the rep roll-up will look optimistic in every period until the value assumption is corrected.

If the gap is confined to one segment, the model probably lacks history there. New products and new territories are the usual cause.

If the gap appears suddenly across the whole business, something changed in the market. A new competitor creating pricing pressure shrinks deal sizes. Tighter capital markets push buyers to cut cost rather than add vendors, and win rates soften. Broad uncertainty slows decisions and stretches the qualified-to-closed cycle. A territory reshuffle distracts the team while coverage ratios still look fine. A model that updates on recent history will register these before a rep roll-up does, because reps are forecasting their own deals rather than the aggregate shift.

Does model forecasting require clean CRM data?

It requires consistent data, which is a much lower bar than clean data. Every company believes its data is uniquely bad and that this is why it cannot forecast. Almost none of them are right. Everyone has messy data, and it matters less than people assume. As long as the mess is consistent, a model trained on your own history learns the pattern including the mess.

What does break a model is inconsistency: stage definitions that changed mid-year, a territory realignment that reassigned history, a CRM migration that dropped fields. Those reset the learning window. Batch structural changes into one annual event and mark the date so everyone knows where the comparison line sits.

Plan for four to six weeks to fully train a model on your own historical sales performance. That is the window before the output is worth reconciling against anything.

What has to be true before a model number reaches the board?

Traceability. A figure that cannot point back to the records that produced it is not usable in a board setting, because validating it costs as much as building the deck by hand. That validation burden is the practical gap in most AI-generated reporting today, and it is more important than the sophistication of the underlying method.

Machine learning and optimization have been doing this work for years, well before the current wave of language models made it fashionable to say so. The value language models add sits mostly in ad hoc analysis, answering a specific question quickly against a semantic layer that already knows what the numbers mean. ORM built Radar for that purpose, exposing the semantic and analytics layer through MCP so it can be queried from whichever model you connect or directly in the Radar interface.

What accuracy should the reconciled number reach?

Better than either input on its own, measured over at least six quarters. Score the rep roll-up, the model output, and the reconciled number separately at every snapshot. If the reconciled number does not beat both, your tie-break rules are wrong and the table above needs rebuilding on your own results.

For a reference point, accuracy around 90 percent on new and expansion business is a common outcome of a heavy manual process, though it takes real effort to produce and stops holding as conditions change. ORM targets 95 percent without manual adjustment, steady from day one to day ninety of the quarter and updating as the quarter progresses. Reaching that range depends less on which number wins any given argument and more on whether the reconciliation is run the same way every week. For the process this sits inside, see sales forecasting best practices and the forecast accuracy definitions.