FELLOW

For people done with confident nonsense

Thousands of workbooks. Answers you can check.

Excel workbooks carry the numbers. Plans, contracts and approvals explain them. FELLOW connects the records, resolves repeated versions and keeps the source behind each claim.

Pre-launch design partners. No spam, or email hello@cortadogroup.com

Sound familiar?
  • Upload → wait → wrong answer
  • Sounds smart. Isn’t.
  • You check the spreadsheet cells yourself anyway
01. The pain

You did the work. The model still lies to you.

A polished answer still leaves you checking workbook versions, reporting periods and the documents behind the figures.

Repeated versions, shifting layouts and source errors need explicit checks before the figures can be trusted.

you, 9:14am

“Summarize this MSA.” Gets a retention number that was never in §7.

you, 11:02am

“Who’s the economic buyer?” Names the wrong person with total confidence.

you, 2:40pm

Re-uploads the same deck for the third time. Answer still soft and wrong.

you, 4:55pm

Opens the workbook again. Checks the tab, the formula and the source figure.

02. Why it keeps failing

“Similar” is not “true.”

What you’re stuck with today

Upload. Embed. Pray.

  • Chunks that sound related win the ranking
  • A 2023 deck beats the 2026 contract
  • Conflicts get blended into one smooth lie
  • You re-check every answer by hand

Great for vibes. Catastrophic for a deal, a clause, or a buyer’s name.

What you actually need

Stop guessing. Show the receipt.

  • Claims tied to the line they came from
  • Org, persona, stage, not a wall of prose
  • Honest “we don’t know yet” instead of fiction
  • Answers you can take into a real conversation

That’s FELLOW. Not another chatbot. The layer under the crap. Why RAG alone fails →

03. The fix

Same workbooks and documents. Answers you can check.

No more re-upload roulette. Feed it once. Get something you can actually use.

  1. Ingest

    Start with Excel workbooks and tabs, then the plans, contracts, transcripts and approvals around them.

  2. Extract

    Identify figures, people, plans and periods. Keep their units and source evidence attached.

  3. Prove it

    Retain the source behind the record. Keep unresolved matches and missing evidence visible.

  4. Use it

    Query the connected records and inspect the evidence behind the answer.

Pilot proof

About 5,900 workbooks. Figures you can trace.

In a FELLOW pilot for a mid-sized services company, FELLOW turned Excel workbooks and their supporting plans into one queryable database. The first year passed automated testing in seven days.

Excel workbooks in the estate
~5,900
Sheets with a recorded outcome across both years
~37,000
First-year captured numbers in named columns
99.95%
Days for the first year to pass automated testing
7

What an estate looks like

  • Monthly Excel workbooks, many tabs each, with earlier months copied forward.
  • Hidden sheets, shifted columns, mixed units and formulas replaced with numbers.
  • Word plans, PDFs and a PowerPoint deck around the spreadsheets.
  1. Find and capture

    Inventory workbooks and hidden tabs. Retain values, formulas and source locations. Record unreadable files.

  2. Type and reconcile

    Read headings. Label units. Match people and plans. Keep ambiguous links visible.

  3. Resolve and check

    Resolve repeated versions. Separate detail from totals. Compare payment records and test coverage.

  4. Query and trace

    Use current-version views. Follow a stored figure to the workbook, tab and cell it came from. Keep unresolved questions visible.

Adding every version produced about 2.5 times the current-version total across 203 tracked tables. FELLOW made that duplication explicit before the figures were used.

Pilot results. A recorded sheet outcome can mean loaded, summarised, retained for reference or excluded with a reason. Both years passed automated testing.

Apply with your workbook estate Click a figure, see its cell
05. Why this is different

Chatbots will die on these two hills. We won’t.

Anything can summarize. What wrecks you is the fake certainty, and the silence when something is missing.

  fct_04J1: claim with receiptauthority: contractual

“Retention overrides the standard 90-day policy for this account.”

Prove it

  • msa_v3.pdf: §7.2, Data Retention[1]
  • renewal-call · 2026-05-14 · 00:41:20[2]

Spreadsheet figures

  • Source: workbook, tab, row and cellcell
No floating claims

If it can’t show the source, it doesn’t get to sound sure.

Follow a claim to its supporting clause, message or workbook cell. In the spreadsheet pilot, a stored figure traces back to the workbook, tab and cell it came from, so reviewers can check it against the original.

  • Trace a figure to the workbook, tab and cell behind it
  • Old stuff gets superseded, not ranked as “still relevant”
  • Corrections stick; they don’t vanish into a new chat
  acct_4471: known gapsdiscovery

What we don’t yet know

  • Signing authority above $50k: unconfirmedgap
  • Incumbent vendor & contract end dategap
  • Security review requirement: implied, unverifiedgap
No silent holes

It says when it doesn’t know instead of freestyling.

Half the damage isn’t wrong answers. It’s answers that paper over missing information. FELLOW surfaces the gaps as the next thing to go find before a client does it for you.

  • Missing facts flagged, not invented
  • Gaps become a discovery list, not a vibe check
  • Agents can ask for what’s absent, not fill it in
06. Who this is for

If you’ve been burned by
“just ask the model,” this is for you.

01

People who still check the workbook

Because you can’t walk into a renewal with AI fan fiction. See the account (org, stage, persona) and click any claim back to proof before you say it out loud.

02

Agents that must not freestyle

Same account truth over the API. Scoped, current, grounded, so automation doesn’t invent a signer or a stage while you’re not looking.