AI MATCH RULE ASSISTANT
How much should AI decide for the user? An assistant that suggests reconciliation rules, explains them, and leaves the final call to people.
Client
SIMETRIK
Year
2024
Category
RECONCILIATION
Role
Senior Product Designer
Simetrik helps finance teams reconcile transactions. First, customers set up matching rules: what to match, with what tolerance, and in what time window.
Customers got stuck on that step. Run, export to Excel, compare, adjust, repeat: it took 2–3 days to reach a valid setup. Our internal implementers did it in under 10 minutes.
Senior Product Designer, from problem to post-launch. I did the research hands-on, set the product direction with the PM, and worked with Engineering and Data Science on how suggestions would be scored and explained.
Users had no mental model. Nothing showed them which rules usually work for each type of reconciliation.
Feedback came too late. You only saw a rule's impact after a full run.
Even experts settled for basic rules. We saw it in all 5 clients we interviewed, from a bank to an e-commerce business. The knowledge existed, but it lived in our implementers' heads.
/// 4 DECISIONS
The real question wasn't how to build a rule assistant. It was: how much should AI decide for the user?
01 — AI suggests. It never applies on its own.
Why: These rules decide which money matches. A silent mistake is worse than a slow setup.
Trade-off: One more confirmation step, even when the suggestion is right.
Result: People used the suggestions because the final call stayed theirs.
02 — Every suggestion explains why it exists.
Why: A confidence score alone doesn't tell an analyst whether to trust a rule.
Trade-off: More text on screen, and extra work with Data Science to turn scoring into plain language.
Result: "Why this rule" was one of the most used parts of the assistant.
03 — Preview the impact before applying.
Why: Late feedback caused the Excel loop. Showing the result first breaks it.
Trade-off: The preview had to feel instant, which limited what we could compute.
Result: More rules applied on the first try, and much less exporting to Excel.
04 — Manual adjustment is always possible.
Why: Experts know things the model doesn't. Taking control away would make them drop the feature.
Trade-off: Two paths to maintain: assisted and fully manual.
Result: Experts kept editing before applying, and novices got a reliable starting point.
Rule suggestions ranked by confidence, each with a clear "why".
An instant impact preview before applying.
Adjustable tolerances and time windows, always editable.
Rule templates per reconciliation type, plus an audit trail of who changed what.
Users: Setting up a valid ruleset went from 2–3 days to ~5 minutes (median).
Business: 95% of target users adopted the assistant, and it became a paid add-on, a new revenue line.
Operations: Fewer configuration tickets, with precision staying within threshold.
I'd bring our internal implementers in as co-designers from week one. They solved in 10 minutes what took customers days, and their knowledge ended up being the templates anyway. I'd also measure the Excel loop before starting, so the "before" is a number and not a pattern we saw in replays.
