Product walkthrough

From a job you redo every month to one that runs itself.

Seven steps, and you can see exactly what happened at each one. AI handles the ambiguous parts once, you approve the decisions that matter, and the engine produces the numbers on every run after that.

01

Upload and describe

Source

Drop in the workbook you actually use and say what needs doing, in your own words. No template to fill in, no fields to configure first — the file can be as involved as your real one, because that is the one you want automated.

  • Excel workbooks, formulas, formatting, and all
  • Described in plain language, not a rule builder
  • The file is parsed in your browser and never uploaded

02

Read the workbook

AI

CellWise works out what is actually in front of it: where each table starts, which rows are headers, what the period columns run on, which cells are formulas and which are typed in, and which parts of the sheet look unfinished.

  • Tables, header rows, and repeated blocks located
  • Formula patterns and dependencies traced
  • Structure summarised compactly, so even a large model fits

03

Propose a plan, ask what is unclear

AI

You get the job written back to you as a sequence of steps in plain language, along with the questions that genuinely change the answer — whether a threshold is a hard floor, whether an already-filled cell may be revised, what to do at a cutoff.

  • Each step explained, with the ranges it reads and writes
  • Assumptions and unknowns stated rather than buried
  • Your answers are remembered, not asked again next run

04

Approve the plan

Human review

Nothing has been computed yet. You read the plan, answer the open questions, and correct anything the model misread. Only when it matches what you would have done yourself does anything get built.

  • Review before any cell is touched
  • Answers fold back into a revised plan
  • Approve when it reads like your own process

05

Generate the Work

AI

The approved plan is compiled into a Work: a fixed list of operations on named ranges, the constraints you stated turned into machine-checked assertions, and anchors that let it find the same tables in a later file. CellWise then runs it against your workbook and shows you the checks before you accept it.

  • Operations, checks, and anchors written down as data
  • Verified against your own file, not a sample
  • A Work that fails its own checks cannot be finalised

06

Run it

Engine

Running a Work involves no model at all. Our spreadsheet engine applies the operations, searches for values where a constraint has to be satisfied, and recalculates every dependent cell the way Excel would — in well under a second.

  • Deterministic: same file, same Work, same output
  • Constraints enforced while values are chosen, not after
  • Your workbook comes back patched, with formatting intact

07

Next month, run it again

Engine

Drop in the new file and CellWise checks it against the Work's anchors before doing anything. Tables that shifted are re-anchored and reported to you. Anything genuinely ambiguous stops the run and asks, rather than guessing at your numbers.

  • Moved tables located and reported — “found at A5, was A2”
  • Ambiguous or missing anchors refuse instead of guessing
  • One recipe can seed Works for a whole family of files
Where the line sits

The transform is not a black box.

A model is good at working out what an unfamiliar workbook means. It is the wrong tool for producing numbers you have to stand behind every month. So CellWise splits the work: AI proposes, you approve, and the engine does the actual writing — the same way, every run.

AI reads what is ambiguous

Working out what a workbook means, what you asked for, and which parts are unfinished. Judgment work, done by a model, once.

You approve what matters

No cell is written until you have read the plan and answered the questions. Those answers are kept, so the same decision is never asked of you twice.

The engine produces the output

Every run after that is deterministic and reproducible. Run it again next quarter, on another machine, and you get the same result.