Beta
Your team builds agents in plain language. Datum checks the finished output against records you confirmed, refuses what does not match, and leaves a signed trail that can’t be quietly changed.
In development, built to run in your environment.
Your prices. Your dates. What’s actually true.
Words you never use. Limits you never cross.
One space per project. Nothing crosses over.
The steps you take every time. Saved once.
Proof of what you checked. Months later.
Integrate it via MCP (Model Context Protocol) with
Gemini
ManusClaude, ChatGPT, Grok and Manus let you add a custom MCP connector on their free plans. Perplexity requires a paid plan, and Gemini support is coming.
How it works
Datum proposes facts, rules and skills from what your team already wrote. Nothing is used until someone confirms it, and from then on every draft is checked against it.
One correction, one approval. Nothing is used until you say so.
The difference
Prompts steer the model. Evals sample it after the fact. Neither stops a specific wrong output from shipping. Datum checks every finished output against records your organization confirmed, blocks what does not match, and leaves a signed receipt for every check. Deterministic, not another model grading a model.
Why this exists
of organizations lack a mature governance model for agentic AI. Only 21% say they have one, even as agents scale faster than the guardrails around them.
Deloitte 2025, 3,235 IT & business leaders, 24 countriesof the time AI saves you goes back into fixing what it wrote. For a daily user, a week and a half a year.
Workday 2026, 3,200 employeesA tribunal held Air Canada responsible for a policy its chatbot invented. It argued the bot was separate. That failed.
Moffatt v. Air Canada, 2024Datum doesn’t make the model correct, and it’s no replacement for human review. It makes every output checkable by someone who wasn’t in the room: an auditor, a regulator, a client. What ships still ships in the organization’s name.
How it works
01 / 06
One value you stated, kept under your account rather than in an assistant’s memory, so it survives switching tools. A source is where it can be read again. Saying it in a chat does not make it a fact.
LikeYour setup fee is $750, raised in March. The handbook page a deductible came from.
02 / 06
A rule is measured against the finished draft, not asked for in a prompt. It counts the characters, finds the banned word, and refuses what does not hold. Write it once, it applies to everything after.
LikeUnder 1,300 characters. Never the word guarantee. Always the disclosure line.
03 / 06
Everything lives in a space, and a check reads exactly one. Your own writing, your business, and each client stay separate, enforced at the check rather than remembered by you. Spaces nest like folders, and nothing crosses between them.
LikeOne client’s figures cannot turn up in another client’s proposal.
04 / 06
Something you do every week, written once in your words and reachable from whichever assistant you are using that day. It holds instructions, not facts, so nothing in it is checked and no receipt is issued.
LikeHow you write a listing description. How you run a renewal review.
05 / 06
The raw thing you work from, kept where the work happens instead of in a folder you forget. A brand guide, a transcript, the notes from a call. Nothing in it is checked and no receipt is issued, which is the point: it is what you write from, not something you are claiming is true.
LikeThe brand guide a client sent. The discovery notes from a first call, kept as you wrote them.
06 / 06
Every check leaves a signed record of what was checked, against which facts, and when. It can’t be edited after the fact, so when someone asks why something went out, you have the answer.
LikeSix months on, proof of which confirmed rate a quote was checked against.
Where it sits
Going out, Datum gives the model only what you’ve confirmed, from one space. Coming back, it checks the output so the model cannot invent what it wrote. Enterprise is being built so both ends run inside your boundary: your data does not leave, and every decision is logged.
The obvious question
Partly, and those parts help. DLP watches for patterns, evals sample quality after the fact, and model guardrails steer tone. None of them check a specific finished output against your own records and refuse it. That is the gap Datum closes, deterministically, with a receipt.
See it
Ordinary outputs, from the assistants your teams already use. The details that are right until the day they quietly aren’t.
“Our uptime SLA is 99.9%.”
Datum blocked itThe signed contract commits to 99.5%. The overstated number never reached the customer.
“As agreed, your terms are net-60.”
Datum blocked itNet-60 belongs to a different account. This one is net-30. It never left the draft.
Gemini wrote
“This solution is fully HIPAA compliant.”
Datum blocked itLegal has a standing rule against unqualified compliance claims. It found it and held the draft.
Pricing
In development. Private deployment, single sign-on, and an audit trail your compliance team can verify, scoped to your organization.
Book a callSmaller than an organization? See every plan.
Billed monthly by card. Cancel any time. Anyone who joined before October 5, 2026 keeps the plan and price that was listed when they joined for a year, with thirty days notice before anything changes.
In development. Early deployments are scoped with your security and platform teams.
The free plan needs no card.
Questions
No. Datum runs alongside whichever models your teams already use. It performs no inference and ships no model of its own. Your assistant writes; Datum decides whether the output matches what your organization confirmed.
Guardrails steer the model and evals sample it after the fact. Neither blocks a specific finished output. Datum measures each output against confirmed records and refuses it on mismatch, deterministically, and that is why the receipt is worth anything.
No, and it never claims to. It knows what your organization confirmed. Point it at a wrong record and it will faithfully hold output to that record, and the receipt names which record it used. What it prevents is the model asserting something no one confirmed.
Your records and content stay under your account, and no one at Riley AI Systems reads them. You can export or delete everything at any time without asking. Enterprise deployments are being built to run inside your own environment.
A signed record of what was checked, against what, and how it came out. If an auditor or regulator asks, the record shows what was checked and proves it hasn’t changed since. That is the part nothing else does.
It refuses things it should not sometimes, and it cannot catch a claim your assistant never declared. Those limits are written here rather than buried, including the ones that are awkward.
Over MCP today, with receipts you can verify from the command line. API access for programmatic and CI use is in development. It sits between your assistants and the systems they read, so existing workflows don’t change. The plans are here.