Grounded · by CodelessOps
Ask anything.
Verify everything.
An audit-grade retrieval layer for financial documents. Point it at your filings, schedules and decks, ask in plain language, and get the version-correct answer sourced to the exact figure, or an honest "I can't confirm that." Built for the moment someone asks "are you sure about that number?"
Sold as an implementation, not a licence: the platform plus the customisation around your pack, your questions, your close. Grounded is the engine behind the financial knowledge base build, one of three ways I work with finance teams.
01 · The problem
The costly number is the one nobody had time to check.
When an answer costs an afternoon, questions get rationed. The board pack ones get asked. The "hang on, why did that move?" ones wait for a quieter week. Nobody is at fault here, that is what retrieval by hand costs.
The number that hurts is the one that looked fine and nobody had the afternoon to open. Grounded makes the question cost a minute, so it gets asked.
47%
Almost half of a finance team's effort goes on collecting and validating data.
FP&A Trends Survey 2026. That is the half Grounded takes off the desk.
Source: FP&A Trends, "Designing the FP&A Operating Model for the AI Era" (16 Jul 2026), citing the FP&A Trends Survey 2026.
01
Find the right spreadsheet
retrieval
02
Open and navigate the tabs
retrieval
03
Find the section, make sense of it
retrieval
04
Write the analysis
analysis
02 · What makes it different
Built for the moment someone asks "are you sure?"
Version-correct across restatements
Two vintages, two correct answers, never a blend. Which version is a first-class question.
Shows its exact source
The filing, the schedule, the cell. The citation is harvested from what was retrieved, never written by the model.
The trust chain
Every figure traced to your documents or computed by code from figures that are. The unverifiable is held back, not softened.
Reads the grid, not the words
A number's identity is its position: column, period, scenario. Never its label.
Validated on your golden questions
The questions that must be right every time, validated with you before go-live, and your scoreboard after it.
Your data, on your terms
A private workspace, your own AI keys or AWS Bedrock with zero retention, nothing used for training.
The distinction that matters: a citation shows where a number claims to come from; a verification proves it's right. It's the difference between "the invoice is attached" and "the invoice was matched." Grounded's trust chain runs your documents → extraction → computation → answer, every link checked by code.
03 · Then build on it
Every answer is traceable, so you can put checks on top.
Connect the systems you already use and let the checks run each month, with the source under every flag.
Variance flags
What moved against budget or the last forecast, with the commentary on record.
Shows its working
Reconciliation checks
Two reports, one number. The gap found and traced to the line that explains it.
Shows its working
Board pack prep
Numbers verified before the narrative is drafted, each still one click from its cell.
Shows its working
The chasing goes, the sign-off stays yours.
04 · See it working
Watch one question land.
A board pack question answered in seconds and traced to the cell; citations that open the source at the exact figure; an honest "I can't confirm that"; a board memo drafted from numbers verified first.
In biotech? Watch one $2.1M travel through an entire close
Prefer your own reality? The Golden Questions Sprint: a small fixed-fee week where your team writes 25 questions it already knows the answers to, and I grade whatever AI you use today against them. The scorecard is yours to keep. Nothing to install, no IT sign-off.
05 · Proof over promises
I published an exam for finance AI, then sat my own tool for it.
81/81
The Board Pack Test: 34 realistic finance documents, 25 questions, a deterministic grader. A verified pipeline scores the ceiling; raw frontier chat scored 66 and 71 of the 81 machine-checkable points.
Everything is public: documents, questions, grader, every transcript. Anyone can re-run it, against Grounded or against whatever you're being sold.
06 · Start with a pilot
Proven on your real data, or you hear it from me first.
A fixed-fee, 30-day pilot on your own close pack, in the deployment your security posture allows. Hit the success criteria you wrote and the fee is credited against year one; miss them and you hear it from me first.
Deployed your way
in the setup your security posture allows
Golden questions
co-built and validated on your numbers
Criteria in writing
defined by you before anything starts
Fee credited
hit them and it comes off year one
For your IT and security team: your documents live in a private workspace, isolated from every other client's, never used for model training, on your own AI keys or AWS Bedrock with zero data retention. A fully single-tenant instance on separate infrastructure is there when your security review demands it. Every question and answer is logged, and the code can be inspected before you commit. Bring the hard questions to the call; that's the meeting I want to be in.
07 · Common questions
Asked by every finance leader I talk to.
What is Grounded?
Grounded is an audit-grade AI analyst for finance teams. It reads your budgets, forecasts, variance reports, and board packs, and answers questions in plain English with the exact source cited: the file, the schedule, the cell. When the documents can't support an answer, it says "I can't confirm that" instead of guessing.
Why not just upload the spreadsheet to ChatGPT?
For one file, one question, once, that genuinely works. But a chat upload flattens a spreadsheet into text, forgets everything between sessions, and can't tell you which version a number came from. Grounded reads the grid positionally, keeps your whole reporting estate ingested and current, and cites every answer to the cell so verification takes seconds, not an afternoon.
How do we get started?
With the Golden Questions Sprint: a small fixed-fee week where we write the 25 questions your team already knows the answers to and grade whatever AI you use today against them. The scorecard is yours to keep, and the same question set becomes the acceptance test if we go on to the pilot or the build.
How do we know the answers are right?
Two ways. Every answer carries a citation the system harvested from what it actually retrieved, not a reference the model composed, so checking a number takes one click, not an afternoon. And before you rely on anything, we validate your golden question set together: your questions, your pass threshold, re-runnable any time. Quality stays a number on a scoreboard, not a feeling.
What's the difference between a citation and a verification?
A citation shows where a number claims to come from. A verification proves it's right. Most AI tools stop at the first: a tidy footnote next to a figure the model produced. Grounded does both. Citations are captured from what was actually retrieved, arithmetic runs through code rather than the model, and every figure in an answer is checked against your documents before you see it. The result is shown as a trust chain under the answer; anything unverifiable is held back.
Is our data used to train AI models?
No. Grounded runs hosted, with your documents in a private workspace isolated from every other customer, on your own AI keys or AWS Bedrock with zero data retention. Your data is never shared with other customers, nothing is used for model training, and every question and answer is logged for audit. A fully dedicated single-tenant instance is available if your security policy requires it.
08 · Bring your own pack
Thirty minutes on your calendar.
Bring the ten questions your team gets asked every month, or your last close pack, and watch them answered with the source, or refused, the same way.
- Your questions, live
- No pitch unless you ask for one
If thirty minutes turns into "do it on our pack", the next step is a fixed-fee pilot on your own close pack: success criteria you define in writing, your golden questions as the scoreboard, and an honest report if we miss them.
