CellWise turns a spreadsheet job you do by hand into a Work — a checked, reusable automation. AI plans it with you once. After that it runs on our own spreadsheet engine in under a second, with no model in the loop.
8 operations · 32 checks · 10 tables
A public benchmark built from spreadsheets people actually work in — not toy files.
The ten tasks reported here are the benchmark's Financial_Model set: multi-sheet models where a single task can require hundreds of dependent cells to land exactly right. A task counts as solved only when every required cell matches and nothing else in the workbook moved — partial credit is reported separately as the 93% figure.
Four steps, and you can see exactly what happened at each one.
Drop in the workbook and say what needs doing, in your own words.
CellWise reads the file, proposes the steps, and asks about what is genuinely unclear.
Approve the plan and it compiles into fixed operations and checks, proven against your file.
Under a second, no model involved. Download the workbook with its formatting intact.
The work itself is not hard. It is just long, fiddly, and due again in four weeks — and the person who knows the rules is the bottleneck every time.
A Work: the job compiled into fixed operations, the constraints you care about written down as checks, and a result you can re-derive on demand. Run it on this month's file in under a second — and on next month's without asking anyone how it was done.
Two kinds of team feel this every month: the people who plan against hard constraints, and the people who keep a model current. Both are doing work a machine should be repeating for them.
Two different jobs, two different tools. Working out what a workbook means needs a model. Producing numbers you will act on needs something that behaves identically every single time.
The real file — its own layout, formulas, and formatting.
Reads the structure, proposes steps, asks what it cannot infer.
The plan in plain language, before anything is computed.
Fixed operations and checks, verified against your file.
Same file, same Work, same result — every run, no model.
The real file — its own layout, formulas, and formatting.
Reads the structure, proposes steps, asks what it cannot infer.
The plan in plain language, before anything is computed.
Fixed operations and checks, verified against your file.
Same file, same Work, same result — every run, no model.
Reading an unfamiliar workbook and working out what you meant is judgment work, and that is where a model belongs. Producing the numbers is not. Once you approve the plan, the model is out of the loop entirely.
The constraints you stated become machine-checked assertions. CellWise recomputes the whole workbook and tests them cell by cell before you ever see a result — and a Work that fails its own checks cannot be finalized.
We patch the workbook you uploaded rather than rebuilding it, so styles, column widths, borders, and external links survive exactly as they were — with every value already calculated when you open it.
Chat assistants are getting genuinely good at spreadsheets, and that is not the thing we are arguing with. They are simply the wrong shape for work you repeat: you re-explain the job every time, and every answer arrives with unknown provenance.
CellWise does one job: take a spreadsheet task you repeat, and turn it into something that runs itself — the same way, every time, with the checks attached.
The second run of a Work matches the first, on the same file, on any machine. That is a property of the artifact, not a promise about a model's mood.
“Cover never drops below two months” is either true of the output, or the run names the cells that broke it. Nobody has to take the software's word for it.
A generated Work has no dependency on a model at run time. It executes on the spreadsheet engine, in your browser, in well under a second.
If you hand a plan to a colleague, a supplier, or an auditor, you need to be able to say where each figure came from — and produce it again on demand. CellWise keeps that answer available by default.
Your workbook never leaves your browser. Only a short structural summary is sent for planning.
The run is deterministic: same file, same Work, same output — on any machine, with no model involved.
Every check is reported cell by cell, so you can show exactly which constraints held and which did not.
Planning runs on EU-based servers with industry-standard encryption. Your files are never used to train models.
Product updates, company news, and articles on data cleaning, validation, and file standardization.
The questions that come up before a team lets new software touch the file its month depends on.
It is the automation CellWise compiles once you approve a plan: a fixed list of operations on named ranges, the constraints you stated written down as checks, and the anchors that let it find the same tables in a later file. It is data you can read, not a black box — and running it does not involve a model.
CellWise matches the new file against the Work's anchors before it runs anything. Tables that moved are re-anchored and reported to you — “found at A5, was A2.” If something is genuinely ambiguous, such as an anchor it cannot find or two that match equally well, it refuses and asks you to regenerate rather than guessing.
No. AI plans the Work with you, asks about anything ambiguous, and tests its own draft against your file. After you approve, the numbers are produced by our spreadsheet engine — deterministic, reproducible, and identical on every run.
Every Work carries the constraints you stated as machine-checked assertions. After a run, CellWise recalculates the workbook and reports each check with the cells behind it. If a constraint cannot be satisfied, it says so and names the cells, rather than quietly writing a number that breaks it.
It survives. CellWise patches the file you uploaded instead of regenerating it, so styles, column widths, borders, merged cells, and links to other workbooks are preserved. Computed values are written into the file too, so it opens correct without Excel having to recalculate anything.
It depends on how many Works you build and how often you run them — and running an existing Work costs almost nothing, because there is no model involved. We would rather quote you honestly than post a number that does not fit your case. Book a demo and we will walk through it.
Bring the one that normally costs you an afternoon. We will build the Work with you, run it while you watch, and show you what happens when next month's file arrives.
You can also email us at info@cellwise.io.
