ReconLayer
A middle office break-triage tool for small finance-ops desks.
It reconciles a desk's internal book against broker files, flags the breaks, and drafts the dispute email — with the AI drafting the message, not the decision. Tested with five people who reconcile books for a living.
Role
Product Design
Tools
Google Apps Script · Gemini
Type
Self-initiated
Status
0 → 1
Timeline
Jan - Mar 2026
Problem
On a small finance-ops desk, someone reconciles the firm's internal book against broker files every morning by hand, in Excel, before a settlement cutoff. A missed or misclassified break can feed a wrong margin call, and the analyst's name is on the decision.
The tools that automate this are built and priced for large institutions, so the small desk just lives in spreadsheets. I spent over a decade on institutional trading desks, with asset management and settlements before that. I built this to test whether a small tool could take the worst of a reconciliation morning off a desk's plate without pretending to be an enterprise platform.
The Approach
"My name is on that message. If it flips a sign or misstates a decimal under pressure, that's a direct liability."
Operations Analyst, Derivatives Desk · Research Interview
The boundary between rules, draft, and send isn't an implementation detail — it's what keeps the analyst able to defend every send.
What I Built
I built an add-on that lives inside the desk's own sheet, next to the trades. Nothing leaves. It reconciles the two books, flags the breaks, and drafts the counterparty email from the rules engine's output — never from the raw trade data.
The sidebar surfaces three things: the net dispute at the current classification, the ranked value breaks with their variance and reasoning, and the one-sided population breaks a simple sort would have missed. Trades within the $50,000 MTA tolerance are filtered out as noise before anything is shown. The analyst sees only what needs a decision.
The draft email is the AI's only job. It reads the classified breaks and writes a counterparty-facing message from that output — the language layer sitting on top of the rules layer, never touching the trades themselves.
Two rules do the sorting. A value break is a variance past the $50,000 MTA tolerance — a real disagreement on price. A population break is a trade in one book but not the other — one that shouldn't be missing at all. Every line gets sorted into one, the other, or "within tolerance, ignore." That sort is the decision, and it's the part the AI never touches.
The sidebar after Reconcile runs. Value breaks are ranked by variance; population breaks are pinned separately. MTA tolerance filters the noise out before anything reaches the analyst.
The AI writes the message from the classified breaks. It never sees the raw trades. If the language layer is unreachable, the drafting step fails cleanly and the classification still holds.
On the synthetic test set, the tool surfaces a net dispute of +$327,500 across four value breaks and two population breaks. The arithmetic isn't the point; the classification is.
What Testing Showed
Twenty-three people ran the tool against a synthetic dataset — nineteen completed the core flow, reconcile to draft, and six finished the full exit survey. Unmoderated and time-boxed. The core testers were finance-ops analysts who reconcile books like this for a living. Directional and qualitative, not statistically significant.
Three findings stood out.
The break labels are where comprehension breaks first. Half misread "population" and "value" on first read — one confused them entirely, another described them imprecisely and re-ran the test on her own before getting them right. Once you can't read the classification, you can't defend the send. This is the clearest fix, and it leads what I'd do next.
Nobody would send the drafted email without checking it. Two would edit first, two would send as-is — but the "send as-is" testers still named the check: double-check the figures and breaks make sense; review it at least. The check happens whether or not the interface asks for it. That's the line the tool was designed around, and it held on its own.
Trust in the tool and trust in the send are separate. The tester who wouldn't send the AI's email rated the tool 5 out of 5 — and was the one who misread the labels. A subset of analysts will always draw the line at sending an AI-drafted message. The tool's job isn't to talk them across it, but to make sure the ones who do send have something they can defend..
"I'd at least still double-check the figures and breaks make sense."
Operations Analyst, Usability-Test Participant
What's Next
Clearer break labels. "Population" and "value" are middle-office jargon that read as interchangeable to newcomers. Testing named this as the first fix — and the tester who misread the labels was the same one who wouldn't send the AI's email. The comprehension fix and the trust question are closer than they look.
An audit view. Testing showed the trust boundary the tool was built around held — but the trust question moved down a layer, onto the rules engine itself. Every tester named some version of the same Monday blocker: a false positive, a missed break, a calculation they couldn't verify. An audit view is how you answer that.
Broker-specific email formatting. The testers who said they'd send the drafted email as-is both named the same check first: does this match the format the broker expects? Formatting isn't cosmetic — it's what makes the send-as-is path defensible.
Longer term: an Excel version. Google Sheets was the fastest surface to prove the pattern. Excel is where these desks actually work.
Where the pattern generalizes. The rules-engine-decides, AI-drafts, human-sends structure isn't specific to reconciliation. It applies wherever a middle-office decision gets communicated to a counterparty — trade confirms, allocation breaks, corporate actions notices, margin call disputes. Reconciliation was the first workflow because it's the one I know cold. It won't be the only one.
What I'd carry forward. On a desk, someone's name is on every decision. A tool that moves money can help with the work, but it can't absorb the accountability — so the design question was never "how much can the AI do," it was "what can an analyst still stand behind after it does." Keeping the AI on the drafting side of that line is what makes the output defensible. For a solo builder in a domain big vendors already own, the opening isn't a new feature. It's the unglamorous last mile everyone else hands back to the analyst
Go Deeper
ReconLayer is a Google Sheets add-on, so it runs inside your own Google account rather than as a public demo.
Open the sheet → File › Make a copy
In your copy: Extensions › ReconLayer › Open
Google will ask you to authorize the script — because it's a personal, unverified build, you'll see a "this app isn't verified" notice. Click Advanced › Go to ReconLayer, then allow. (It runs on synthetic data only; nothing is sent anywhere but the Gemini draft call.)
Click Reconcile, then Draft email.
Previous
Next