Trust the numbers. Even the ones AI wrote.
CellGuard checks the edits an AI makes to your financial model and catches the bad ones before you rely on them.
EBITDA: a live formula was silently overwritten with a number.
AI edits your model
CellGuard checks each edit
You keep the proof
Your team gets the speed of AI.
You keep numbers you can sign.
“Avoid using AI-generated outputs for financial reporting, legal documents, or other high-stakes scenarios.”
On the same support page, Microsoft says to use native formulas for “any task requiring accuracy or reproducibility,” because COPILOT “can give incorrect responses” whose results “may change over time, even with the same arguments.”
The author of your models changed. Your accountability didn't.
AI now writes directly into live financial models. In 2026 both Microsoft's Copilot and Anthropic's Claude rolled out to Excel users broadly, editing the workbook itself, with no independent check before a change lands. The controls that catch a wrong number were designed around a human typing one number at a time. Nobody is typing one at a time anymore.
AI ships inside Excel
Copilot and Claude now write straight into live workbooks.
Faster than you can review it
Agents edit at machine speed. “Work on a copy” is a disclaimer, not a control.
You sign the numbers
SOX, DORA and your auditor hold you to every figure, whoever wrote it.
That gap is yours to close,
so we built CellGuard.
Every edit goes through the same gate.
- 01
Propose
Any AI drafts an edit: ChatGPT, Claude, Copilot or your own agent.
- 02
Check
CellGuard checks it independently, no AI in the loop. In agent tools the check runs before the edit is saved; work from other assistants is reviewed the moment you send it.
- 03
Verdict on the record
Block, flag or pass, sealed on a log no one can quietly rewrite. When the auditor asks what changed and when it was checked, you hand them the record.
Why a CellGuard verdict holds up when your numbers are challenged.
Independent
The thing checking the work never did the work, whichever AI proposed it. So when someone asks who checked the AI, the answer is never “the AI”.
Same input, same verdict
Fixed, written rules, not another AI's opinion. Run the check twice, in front of your auditor, and the answer is identical.
Sees the whole model
Every sheet, not just the edited cell. It catches the edit that looks fine on its own but breaks a total three sheets away.
The dangerous edit isn't the one that turns a cell red. It's the number that still ties out but is wrong. Out of the box CellGuard guards the identities every model must satisfy: assets = liabilities + equity, gross profit that ties to revenue and costs. Then tell it the relationships specific to your firm: margins, allocations that sum to 100%. It flags any edit that quietly breaks one, even when the sheet looks clean. Caught as the edit is made, not three weeks later in the board pack.
Out of scope, by design: whether an assumption is commercially wise (we check that it sits inside its range, not that it is right); errors that were already in the model before the AI touched it; anything beyond the 42 named checks. Nothing past the list is silently guessed at. Your Model Risk team can test that boundary, cite it and defend it. Business judgment stays where it belongs: with you.
Forty-two ways a model quietly goes wrong. We check every one, by name.
No black-box scores. Every check is a named, fixed rule with a set severity: block, flag or info (noted on the record, never stops an edit); an edit that trips no rule passes. In the API a flag is returned as warn. The rules are deterministic, meaning the same input always returns the same verdict. When Model Risk asks what exactly gets checked, this is the list you hand over.
Formula integrity
A curated view. The full, versioned catalogue is yours on request: every rule and its severity, plus a plain statement of what we deliberately don't check, and why. Block is reserved for provable breaks; every flag carries the cell, the rule and the reason.
The fastest way to judge it: run the 42 checks on one of your own models.
See it on your model →It runs where the work already happens.
Pre-save blocking is live in agent tools today. Everywhere else, CellGuard reviews the workbook right after the AI is done. Same rules, same verdicts, same record.
Inside your AI toolspre-save check · MCP
In AI developer tools like Claude Code and Cursor, the agent asks CellGuard before it writes the edit. For the full 42 checks, the agent sends CellGuard the workbook itself.
Live
Beside your assistantafter-the-fact review
Send a workbook Copilot or ChatGPT already touched. You get a clear verdict and an evidence record, with no change to how anyone works.
Live
In Excela panel in the workbook
Review verdicts, with the reason for each, inside the workbook. Same rules, same verdicts. One click to check the open workbook.
PlannedSecure, independent, on the record.
Your spreadsheet is held only while the check runs, then deleted automatically within the hour. It is never used to train any model.
What we keep is the audit record: each verdict is fingerprinted (SHA-256) and chained to the one before, so history can't be quietly rewritten. Time, rule, verdict. Never your cell values. Kept 24 months, exportable for your auditor.
CellGuard runs as an independent check around any AI, in our cloud or entirely on your own servers.
Security details →
CSA STARLevel 1 self-assessment · public registry
GDPR-alignedself-assessed · EU data protection
NVIDIA Inceptionprogram member
CellDelta is a member of the NVIDIA Inception program. © 2026 NVIDIA, the NVIDIA logo, and NVIDIA Inception are trademarks and/or registered trademarks of NVIDIA Corporation in the U.S. and other countries. Membership does not constitute an endorsement by NVIDIA.
Full security posture, sub-processors, DPA and a sample evidence pack available on request.
Let AI write the model.
Keep the control.
We use your email only to handle your request. See our Privacy Policy. To accelerate access, email [email protected].
We couldn't submit that automatically. Please email [email protected] and we'll set you up.
Request received
A founder replies, usually within one business day. Meanwhile, the docs cover the connector, verdicts and data handling.
Tailor your walkthrough
All optional. Your request is already in; these just help us tailor the first conversation. Used only for that. See our Privacy Policy.
Thank you
That helps. We’ll be in touch shortly.