a deal copilot with a banker’s caution

it does the analytical grind on a live mandate, traces every claim to its source, and asks before anything reaches a client

A boutique merchant bank lives on a handful of complex deals. A mandate either closes for a large success fee or it earns almost nothing. In that world a missed deadline or one unsupported claim in a client memo is a potential zero, and it can cost a relationship that took a decade to build.

Built the copilot that does the analytical grind on a live deal while holding a careful banker’s caution: it understates, it flags, and it sources everything.

what it does

It sits on the firm’s deal data and carries the heavy, detail-intensive work across a mandate. It ingests messy company data (trial balances, debt schedules, cap tables, ERP exports) into clean, auditable inputs, surfacing every ambiguity rather than smoothing it away. It audits the financial model. It drafts and fact-checks the materials. It ranks the right counterparties. It tracks the process so nothing drops.

it reads the model, never edits it

The financial model in Excel is the source of truth, and the copilot never silently changes a cell. It reads the model and checks the things that quietly break a deal: the balance sheet ties, the cash-flow bridge, the debt roll-forward, the covenant math, the places where the deck and the model disagree. When it believes a formula is wrong, it proposes a patch, the current formula, the suggested formula, and the expected impact on the headline number, and a banker decides whether to apply it.

every claim carries its source

In materials, every substantive financial or operational statement maps back to a source document. Unsupported claims get flagged to be sourced, labelled a management estimate, or cut. Projections are marked as projections. The tone understates by default. The target hallucination rate is zero, because a single unsupported number in a CIM is exactly the kind of error that ends a relationship.

what waits for your approval

Internal work runs without interruption: retrieval, drafting, issue detection, scoring, checklists. Anything that touches the outside world waits for a person, sending an email, changing a model assumption, granting data-room access, moving a deal stage. Client information stays on the firm’s controlled infrastructure, and one client’s data never surfaces in another client’s context.

the bottom line

A small team running complex deals has a thin margin for error and a brand built over years. The copilot shortens the distance between taking a mandate and closing it by doing the grind faster and catching the mistakes earlier, while leaving every external decision and every published number to the people whose names are on the deal.

Law, accounting, diligence: the same caution earns its keep wherever the work is detail-intensive, reputation-dependent, and binary, and a single sloppy line can cost the engagement.

One of a family of agentic systems built on a shared, auditable control plane.