With the problem reframed as an internal last-mile layer, ideation came down to choosing which manual loop to attack — and staying disciplined about what a solo builder could actually ship.


Property Value
🏷️ Status In progress
📅 Project ReconLayer · Middle-Office Break-Triage Tool
🔖 Phase Ideate
🧩 Input The four Discover findings · the three hard constraints
🗂️ Output Concept selection · form-factor decision · MVP cut-line

The wrong axis, then the right one

An early concept list ranked ideas the way most product exercises do: by customer value and AI differentiation. That axis was misleading. For a solo builder, the binding constraint isn't value — it's integration surface: how many live systems a tool must connect to in order to function. Re-scored on that axis, the "highest-value" ideas were exactly the least buildable and the most already-owned.

Concept Customer value Real blocker
AI exception investigator Very high Needs live feeds from the OMS, custodian, prime broker, and SWIFT — and it already exists as SmartStream Smart Agents
Settlement-failure predictor Very high Needs live settlement + counterparty history; owned by DTCC / Clearstream / Meritsoft
Operational decision copilot Very high "One assistant across twelve systems" is a platform and an integration wall, not a plugin
AI knowledge search Medium Horizontal RAG owned by Glean / M365 Copilot; nothing middle-office-specific to build
Reconciliation assistant High Runs on two sheets the analyst already has — no live integration
Counterparty scorecard Medium Computable from the desk's own history file — buildable
Auto-drafted comms Medium A generation feature, best folded into the recon assistant

Everything at the top of the value ranking required exactly the multi-system integration a solo builder can't do and incumbents already had. Everything buildable clustered lower — because it ran on data the analyst already exports.

The selection

The winner was the reconciliation / exception assistant, with the counterparty scorecard and auto-drafted email folded in as panels rather than separate products. It sat at the one intersection that mattered: high value and buildable without live integration. It's the buildable cousin of the "AI exception investigator" — same job of explaining and acting on a break, minus the twelve-system dependency that makes the enterprise version unshippable by one person.

The form-factor decision

The reframe from "product" to "layer" also settled the shape. Not a standalone app, not a system of record — an add-on inside the spreadsheet the desk already lives in. That choice inherits the incumbent's distribution and data, competes with nothing, and needs no migration. It's the difference between asking a lean desk to adopt a new system and simply adding a button to the tool they open every morning.

What got deliberately cut

Ideation is as much about subtraction as selection. Cut, on purpose, for a first build:

Each was cut because it either reintroduced the integration wall or replaced the auditable core with a black box. The cut-line itself became the MVP scope, carried forward into Prototype.

Slide-ready takeaway: Scored on integration surface rather than raw value, the "best" AI ideas were the least buildable and the most already-owned. The reconciliation assistant won because it ran on data the analyst already had — high value, zero live integration.