For analysts, associates and whoever signs off

The AI can’t
plug your model.

Not shouldn’t — can’t. Every write goes through one path, and that path refuses a number typed over a formula.

Build a model See pricing

Tenant-isolated by org Append-only audit trail Never trained onExports a real .xlsx

The precedent trace panel: Closing ARR, 91% of it is Opening ARR

Integrity strip: 1,632 formulas, no plugs

Click any number and it tells you what it’s made of, which typed drivers move it, and what breaks if you change them.

The strip above runs on every keystroke: how many live formulas, and whether anyone has plugged one.

Built for

Corporate Finance

M&A Advisory

Private Equity

Credit & Restructuring

Equity Research

FP&A

1,632

formulas recalculated in ~20ms, in the browser

typed numbers allowed over a calculated cell

Every cell

change recorded with before → after

~20ms

for a driver edit — no model call, no cost

Measured in the product, not projected. Customer counts appear here once there are customers to count.

How it works

Build it by talking to it.

One thread, one model. “Run it at 12% WACC” edits the workbook in front of you — through the formula layer, not a new file.

Prompt library grouped by deliverable — Named the way the desk names the job.

Integrity strip showing 1,632 formulas and no plugs — The check a reviewer runs first, run on every keystroke.

Formulas stay formulas

Change a driver, 1,632 cells recalculate in ~20ms. In the browser.

Structure edits repoint references

Insert a row and =SUM(B10:B20) still means the same block. The silent killer in AI-built models.

Circularity is solved, not flagged

Iterative calculation, like Excel. A three-statement model is circular by nature.

Model integrity

Every number shows its work.

Nobody asks “what is D38”. They ask why is free cash flow negative in 2027 — and in Excel that’s Ctrl-[ until you’re lost.

Audit trail rows showing before and after values per cell

A trace, not a guess

Read off the dependency graph we already build on every recalc. No model call — an inferred breakdown is worth nothing to whoever signs off.

The hardcode scan a reviewer runs by hand

A literal sitting inside a row of formulas is a plug: the model still computes, it's just wrong from that column on.

Errors point at their source

A broken cell is almost never broken where it shows #VALUE!. We name the cell the error actually starts in.

What a competitor’s audit log can say

“A workbook was produced at 14:32.”

What ours says

Assumptions!B30 4.5% → 6.0%

Honest comparison

What we do, and what we don’t.

A modelling desk doesn’t want a data platform bolted on.

Capability	Supasheet	Rogo, Copilot, add-ins
Refuses a hardcode over a formula	Yes	Generated, then reviewed by you
Cell-level audit trail	Yes	A file is produced
Precedent trace with driver breakdown	Yes	Explanations are model-generated
Repoints references on structural edits	Yes	Depends on the generated formula
Live market data	No	Capital IQ, FactSet, LSEG, PitchBook
Citations over a document corpus	No	Their core product

They are in Excel; we replace Excel for the editing loop. That fork is what makes the guarantees above possible — and it’s the trade: we’re a web grid, not the thing already open on your second monitor.

Security

Enterprise-grade by architecture.

The audit trail isn’t a bolt-on. Every change is a recorded operation, not a regenerated file.

Tenant isolation on every route

Every API route pulls org_id off the authenticated user; data is scoped to your organization at the query, not the application layer.

Append-only audit trail

Every change to every model is recorded with before and after values, the actor, and the instruction. Rows are never updated or deleted.

Your model is not a training set

Your workbooks are used to answer your prompts and nothing else. No training on customer data.

Scoped API keys

Machine access runs through revocable, org-scoped keys, metered and logged like any other actor.

On the roadmap, not yet certified

SOC 2 Type IIISO 27001SSO / SAMLSingle-tenant deployment

Who it’s for

Start where you are.

Solo analyst

Build and defend your own models. No one to wait on.

Start free

Deal team

One thread per model. Everyone edits the same live workbook.

Start free

Firm

The trail, the isolation and the controls your review process needs.

Talk to us

Questions we get asked

Is this an Excel add-in?+

No. The grid lives here, which is what makes the guarantees possible — an add-in writes into a workbook anyone can plug a number into, with no record. You export a real .xlsx whenever you need to send one, formulas intact.

What does “the AI cannot hardcode” actually mean?+

Every change goes through one write path, and that path refuses a literal value written over a cell that holds a formula unless it is explicitly forced. It is a guard in the code, not an instruction in a prompt, so it holds regardless of what the model decides to do.

Can I still type over a formula myself?+

Yes — deliberately, exactly as you can in Excel. It is your model. The difference is that it is counted as a hardcode in the integrity panel and recorded in the audit trail, so nobody discovers it by accident three weeks later.

Do you have market data?+

No. Rogo has Capital IQ, FactSet and LSEG; Hebbia has PitchBook. We read the filings, trial balances and models you attach. We would rather say that plainly than pretend otherwise in a demo.

What happens to my models?+

They are scoped to your organization on every query and are never used to train models. Export the .xlsx and walk away at any point — the file is yours and it recalculates in Excel.

The AI can’t plug your model.