Can an AI agent reconcile your data? What works and what does not

"Can't AI just do my reconciliation?" Yes and no — and the people who get burned are the ones who don't know which. AI is genuinely great at the judgment parts and genuinely dangerous at the arithmetic. Here's exactly where that line sits, so you get the speed without ever trusting a number you can't defend.

What AI reconciliation actually means

There are two very different things people mean by AI reconciliation. The first is using a model to help set up and interpret a reconciliation — mapping fields, guessing the key, explaining a difference, drafting a rule. The second is letting a model do the matching and math itself. The first is where todays models shine. The second is where they quietly cause damage.

Where AI agents genuinely help

• Field mapping. Given two messy exports, a model is good at proposing which column maps to which — that Order # and order_id are the same field. You confirm; it saves the tedium.
• Key suggestion. A model can spot that SKU alone is not unique and suggest SKU plus location as the composite key, then you validate it.
• Classifying differences. Sorting a list of discrepancies into timing, fee, and real by their patterns is pattern recognition, which models do well.
• Explaining in plain language. Turning row 4471 differs by 12.50 into this order was refunded after settlement — drafting the narrative a human reviews.
• Drafting rules. Translating ignore differences under a dollar into a concrete rule you then run deterministically.

Where AI agents fail

• Doing the arithmetic. A language model predicts text; it does not compute a reliable sum over ten thousand rows. Matching and totaling must be deterministic code, not model output.
• Silent confidence. A model will produce a plausible reconciliation that is wrong, with the same fluent tone as a correct one. There is no I am not sure unless you engineer for it.
• Reproducibility. The same prompt can give different answers. A reconciliation has to produce the same result every time it runs on the same data.
• Auditability. The AI said so is not evidence. A close needs the row-level trail, which a freeform answer does not provide.

The pattern that works: AI for setup, code for truth

The reliable division of labor is to let the model do the judgment-heavy, language-heavy work at the edges, and let deterministic logic do the matching and math in the middle. The model proposes the mapping, the key, and the rules; you approve them; an engine applies them identically on every run; the model then helps explain the output. The number is computed by code and is reproducible; the AI made getting there faster without ever being the source of the answer.

Questions to ask before trusting an AI reconciliation

1. Is the matching done by deterministic code or by the model? (It should be code.)
2. Does the same input always produce the same output?
3. Can I see the row-level evidence behind every difference?
4. Did a human approve the key, the mapping, and the rules?
5. Where the model classified or explained, can I override it?

If the answer to the first two is "the model" and "no," you don't have a reconciliation you can defend. You have a confident guess. Keep the AI on setup and explanation, keep the arithmetic in code, and you get both. The prompt patterns that keep a model in its lane are the practical version of this.

← All writing