Verified answers from your business data
Every number your AI gives you, computed in code and backed by a receipt.
Grounded sits between your people's AI clients and your business systems. It computes and validates the figures in every answer from the source data, shows exactly where they came from, and gives your company one control plane for connections, people and policy.
Works with Claude, ChatGPT, Cursor, VS Code, Codex and any MCP client.
“What was our revenue quarter over quarter for 2025?”
| quarter | sales | change |
|---|---|---|
| 2025-Q1 | 4,716,749.19 | |
| 2025-Q2 | 4,776,220.61 | +1.26% |
| 2025-Q3 | 7,089,514.66 | +48.43% |
| 2025-Q4 | 7,076,477.19 | -0.18% |
- Computed
- sum of amount where type is sale, by quarter, with change from the prior quarter
- Source
- list_transactions, 20,000 rows held, all pages fetched
- Receipt
- signed and chained, verifiable offline
Why it matters
Language models are not calculators.
Ask a model a question over a large result and it will stop and ask how to proceed, take minutes to add up rows by hand, or, when pushed, estimate. None of those answers can be checked. Grounded returns the figure in seconds, with its derivation and a receipt.
Without Grounded
305 s · $0.85
The model added 1,302 amounts by hand. It was right, and there was no way to show it.
With Grounded
11 s · $0.03
The same figures computed in code over every row, with provenance and a signed receipt.
One recorded run of our demo, one model, one question. Not a benchmark.
How it works
Three steps, and no new habits.
People keep using the AI client they already know. Grounded does its work between the model and your systems.
- 1
Connect once
A company admin adds MCP services (QuickBooks, a CRM, an internal API) and, if you want, Gmail and Google Calendar. Keys and Google sign-ins are sealed inside the gateway and never shown again.
- 2
Install in your AI
Each person installs Grounded in the AI client they already use. One install gives them every connection they are allowed to use.
- 3
Ask, and get verified answers
When a tool returns a large result, Grounded holds the rows and computes over them. The answer comes back as a card that shows what was computed, from what, with a signed receipt.
Content validation
Answers you can check, not just trust.
Grounded is a verification layer for AI answers over business data. It computes figures on request, validates the ones the model claims, and records how every number was produced.
Computed in code, not estimated
Totals, counts, groupings, comparisons and formulas are calculated by Grounded from the source rows, not by the model, and the card says which rows were used.
Validate what the model claims
Send a figure the model stated and Grounded recomputes it: match or mismatch within a set tolerance, with the correct figure when it is wrong.
Provenance on every card
Each answer names the source tool and call, when it was fetched, how many rows were held and used, what filters excluded, and the date range covered.
Prose checks
In clients that route the model through Grounded, the written answer is checked too. Any number no card accounts for is flagged as unverified, even if it happens to be right.
Signed, hash-chained receipts
Every tool call, computed answer and admin action is signed and chained. Download them and verify offline, without trusting us. An edited receipt fails verification.
Large results stay out of the prompt
Big result sets are held by Grounded instead of being pasted into the model's context, so answers over thousands of rows are fast and cheap.
Control plane
One place to govern how your company uses AI.
The Grounded portal is where admins connect systems, decide who can use what, and see everything that happened. Every change is recorded and receipted under the person's name before it takes effect.
Connections, managed centrally
Company-wide, per-group or personal connections. Use one shared key set by an admin, or a key from each person. Keys are encrypted with AES-256-GCM and never returned by any API.
People, groups and roles
Invite people, organise them into groups and assign admin, manager or user roles. Each person sees only the connections they are entitled to.
Kill switch and read-only mode
Stop all AI tool access for your company in one click, or allow reads while blocking every write.
Per-tool policy and approvals
Turn individual tools on or off, mark them read or write, or hold them for approval by an admin or manager before they run.
Activity and metrics
Every tool call, grounded query, prose check and admin action, with who asked, what was computed and the result behind it.
Sign-in security
Two-step sign-in for anyone, required company-wide by an admin, and SAML single sign-on linked per company.
Works where your people work
Install once in the AI client you already use.
- Claude
- ChatGPT
- Cursor
- VS Code
- Codex
- Any MCP client
Google accounts and your data
Gmail and Google Calendar are read-only and per person.
When you connect Gmail or Google Calendar, Grounded asks Google for read-only access to that one service in that account and nothing else. Mail and calendar data are fetched only when your AI asks a question, are held in memory for the duration of the answer, and are never written to our database, sold, used for advertising or used to train models. You can remove an account from Grounded at any time, which also revokes Grounded's access at Google. The details are in the privacy policy and the terms of service.
Put a receipt behind every number.
Sign in to the control plane to connect your systems, invite your team and install Grounded in your AI client.