AI in Commerce1 min read
AI-drafted reconciliation reports
LLMs are good at turning structured data into readable prose. Reconciliation summary reporting is a well-shaped use case, with narrow guardrails.
Reconciliation produces structured data. Someone has to read it, summarise the exceptions, and decide what to escalate. Turning structured data into readable prose is one of the things LLMs actually do well. Applied narrowly, this is a good use case.
Written April 2026 from a pilot with an internal reconciliation tool.
What works
- Summarizing a day's exceptions. "Of 12,400 transactions today, 6 exceptions: 4 timing-related, 2 amount mismatches. Full details below." That paragraph, generated from the exception list, saves an analyst a bootstrap step.
- Categorising exceptions. With a labelled exception taxonomy, the model can propose a category for each; the analyst confirms or corrects.
- Drafting responses to specific exception types. For well-understood categories (timing skew on a known corridor), the model can draft a resolution note. The analyst edits and sends.
What doesn't work
- Autonomous resolution. The model proposes; the analyst decides. Fully automated resolution of reconciliation exceptions produces incidents.
- Anomaly detection. The model sees the data through a narrow context window; it can't compare to baseline patterns without explicit help.
- Handling novel exceptions. The model is best at what it's seen many examples of. First occurrences of new exception classes should be escalated to human review.
The discipline
- Every AI-drafted output is labelled as such in the tool.
- Every AI-drafted output has an "edit before send" step.
- The prompt and the model version are logged with the output, so a mislabeled draft can be traced.
AI-drafted reconciliation isn't a labour replacement. It's a productivity boost for the analyst who's already doing the work.