The reconciliation nobody volunteers for

Two systems both say they're right. The gap between them is where your month-end goes to die, and where most AI projects quietly stall.

Every operations leader can list their tools. Far fewer can tell you whether those tools agree with each other. That second question is the expensive one.

Here is the pattern, and it is boringly consistent. A deal closes in the CRM. An invoice goes out from accounting. A project opens in the workspace tool. Each record is correct on its own. None of them point at each other. So a person becomes the bridge: they export from one system, match it against another, and re-key the difference. Usually on a Friday. Usually against a deadline.

Put a number on it. Ledge's 2025 benchmark of 100 finance teams found cash reconciliation is the single most time-consuming task in the close, eating 20 to 50 hours a month, with most teams stitching together 3 to 5 systems to do it. Half of those teams take more than five business days to close at all. The reason is not the reporting. It is everything before the reporting: matching fragmented data by hand.

Call it 30 hours a month, the figure one finance manager in that survey gave. That is 90 hours a quarter. More than two full work weeks of a capable person doing nothing but matching rows. At a conservative $60 an hour fully loaded, that is about $5,400 a quarter, or north of $21,000 a year, to answer a question the systems should answer themselves.

Here is the method I use to kill it. Four steps, in order.

1. Name the source of truth per object, not per tool. The customer master lives in the CRM. Revenue lives in accounting. Delivery status lives in the workspace tool. Decide who wins for each field before you touch anything.

2. Give every record one shared key. One customer ID that travels across billing, accounting, and the CRM. Half of reconciliation is archaeology because the same customer has three different names and no common number. Fix that and the join gets boring, which is the goal.

3. Make the reconciliation observable before you automate it. Build a daily diff that simply shows where the two systems disagree, in plain view. You want to watch the drift for a few weeks before you let anything act on it.

4. Replace the person doing the join, not the judgment behind it. Automate the matching. Keep a human on the exceptions, where the actual decisions live.

The before and after is the whole argument. Before: 30 hours a month exporting and matching, and a close that slips the moment one source is late. After: the match runs on its own, and a person spends maybe 4 hours a month reviewing what got flagged. That is roughly 26 hours back every month, about 78 a quarter. The work that was invisible becomes a short review meeting.

Now the part that matters beyond the close. This is the same wall every AI project hits. The demo that reconciles two systems takes an afternoon. The structure underneath it, naming the source of truth, assigning the key, making drift observable, is the other 70%. Skip it and the model has nothing trustworthy to stand on.

The data backs this up, uncomfortably. S&P Global found 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before. MIT's NANDA study put 95% of generative AI deployments at zero measurable return. Informatica's survey of data leaders named data quality and readiness the top obstacle. None of those are model problems. They are reconciliation problems wearing a fancier outfit.

If you want the drift measured instead of guessed, that is what OpScan does: it maps your stack and scores the silos with a Silo Score. Full disclosure, it's mine.

The reconciliation nobody volunteers for is the work that decides whether your AI project ships or just demos well. Do it first.