Why Your AI Forecast Is Only as Good as Your Reconciliation

A surprising number of finance teams have figured out that AI helps with forecasting and then quietly discovered it does not help with theirs. The model is fine. The vendor demo was real. The problem is upstream, in the data the forecast is built on, and no amount of model quality fixes a foundation that does not tie out.

The survey evidence is consistent on this. Across 2025 FP&A benchmarks, data quality and availability sit at the top of the AI adoption barrier list, named by more than half of respondents, with the difficulty of integrating or replacing legacy systems close behind. The same studies show the upside is genuine: roughly 42% of organizations rate their forecasts as good or great, and that climbs to about 65% among teams using AI or ML. So both things are true at once. AI moves forecast accuracy. And the single biggest thing standing between most teams and that improvement is not the algorithm, it is the state of their numbers.

This is the uncomfortable part for anyone about to sign a second AI contract to fix the disappointing results of the first one. A forecast is a function of its inputs. If the inputs are three exports from three systems that disagree about what last quarter’s revenue actually was, the forecast inherits that disagreement and dresses it up in a confident chart.

Garbage in, confident garbage out

Old-school garbage-in-garbage-out at least looked like garbage. A broken spreadsheet returned a number that was visibly wrong, and someone caught it. AI changed the failure mode. Feed an LLM contradictory source data and it does not stall or flag the contradiction. It produces a clean, fluent, plausible answer. The output looks more trustworthy than the inputs deserve, which is exactly the danger.

Picture the typical mid-market close. Bookings live in the CRM. Recognized revenue sits in the GL. Billings run through a separate billing platform. Cash lands in the bank feed. Each system is internally coherent and none of them agree at the edges, because of timing, because of a credit note posted in the wrong period, because someone reclassed an account in October and nobody told the data team. A human analyst building a forecast spends a meaningful slice of their week silently arbitrating those differences. They know which number to trust for which purpose. That arbitration is institutional knowledge, and it almost never gets written down.

Now hand the same four feeds to an AI and ask for a revenue forecast. It has no way to know that the CRM bookings figure double-counts a renewal, or that the GL number is the one the auditor will hold you to. It picks something, projects from it, and the error compounds across every period it forecasts forward.

Reconciliation is the foundation, not a chore

Reconciliation is usually treated as the boring monthly task that happens after the interesting work. For AI forecasting it is the interesting work. Reconciliation is the act of resolving those four feeds into one set of numbers that ties back to the ledger, where bookings, billings, recognized revenue, and cash each have a defined relationship to the others and the differences are explained rather than ignored.

Do that once, properly, and you get a single reconciled model: one place where revenue means a specific thing, traceable to its source, agreeing with the audited accounts. That model is what an AI forecast should be built on. Not the raw exports. Not a snapshot someone pasted into a sheet last Tuesday. The reconciled layer.

The accuracy ceiling here is structural. A forecast cannot be more reliable than the base period it extrapolates from. If your starting point is off by 4% because of an unreconciled timing difference, every projected quarter carries at least that 4% before the model makes a single assumption of its own. You can buy the best forecasting AI on the market and you have already capped its accuracy with the data you fed it.

This is also why “just connect the AI to our systems” underdelivers so often. Connection gives the model access to the data. It does not reconcile the data. Pulling live figures from the ERP through a protocol is real progress for retrieval, but a forecast still needs one agreed base to extrapolate from, and raw GL fields are not that. The reconciliation has to happen between the connection and the math. We make that case in more detail in MCP for finance is not enough.

What this looks like in practice

The fix is sequencing, and it is less glamorous than buying a tool. First, decide what each metric means and where its authoritative source is. Second, reconcile the feeds into one model that ties to the ledger, with the differences documented rather than buried. Third, point the forecasting layer at that reconciled model. Only then does AI accuracy translate into forecast accuracy you can defend.

A worked example. Suppose recognized revenue in the GL was 12.0 of something last quarter, but the CRM shows 12.5 in closed bookings and billings show 11.7. An AI handed all three has to guess, and whatever it picks, it is wrong against two of them. Reconcile first, and you establish that recognized revenue is 12.0, the 0.5 gap to bookings is deferred revenue not yet recognized, and the 0.3 gap to billings is a timing difference clearing next month. Now the forecast extrapolates from 12.0 with the deferred balance modeled explicitly. The number means something, and you can show your work to the board.

There is a real limit worth conceding. Reconciliation does not make your assumptions correct. If you assume 20% growth and the market delivers 5%, a perfectly reconciled model will be perfectly wrong. What reconciliation eliminates is the other error, the avoidable one, where the forecast is wrong not because the future surprised you but because the past was never agreed. That is the error you can actually fix, and it is the one most teams have not.

Once the base is reconciled, the rest of the FP&A stack gets safer too. Driver math stays deterministic instead of being re-derived each run, which we cover in driver-based forecasting. And continuous re-forecasting stops drifting, because every refresh ties back to the same foundation rather than to whatever the source systems happened to show that morning, the subject of continuous forecasting without losing control. All of it sits under the broader argument in the AI FP&A pillar.

The takeaway

Before you evaluate another forecasting AI, ask a blunter question: does our base period tie out to the ledger, in one place, with the differences explained? If the answer is no, the forecast tool is not your bottleneck and a better one will not save you. Fix the foundation, and the AI you already have starts earning its keep.

That is the layer Rexfin builds, one reconciled model your forecasting AI can stand on. If you want to see what a forecast looks like when every number traces back to source, book a demo.