Decide exactly what AI can do with your apps

Every time AI reaches into one of your apps, like searching the CRM, sending an email or opening a pull request, Dtwo policies check the request and the response. Block what's risky, narrow what's too broad, hide what's sensitive, and react to what AI has already done.

Free Trial

Block destructive actions

Let AI read and draft. Stop it from deleting.

Some actions can't be undone. Turn them off for AI with one policy, and everything else keeps working: search, summaries, drafts, updates you're comfortable with.

  • Salesforce delete_record(Opportunity, 006…) Blocked

    AI can look up and update deals, but it can't delete CRM records.

  • GitHub delete_repository(acme/billing) Blocked

    Agents can create branches and pull requests, not wipe out a repo.

  • Microsoft 365 delete_onedrive_file(Q4 Board Deck.pptx) Blocked

    AI can find and read files. Deleting them stays a human job.

Inspect tool inputs

Look at what AI is asking for, then decide

Policies can read everything AI sends to a tool: the query, the recipient, the channel, the branch. Use it to narrow a request before it runs, or stop it outright.

Add criteria to queries

Dtwo adds your conditions to the request, so AI only gets back what it should.

  • Salesforce Only your own deals

    AI sends The name and amount of every opportunity

    Dtwo runs The name and amount of every opportunity owned by the signed-in user

  • Jira Keep HR and security tickets out

    AI sends Tickets that mention “laptop refresh”

    Dtwo runs Tickets that mention “laptop refresh” outside the HR and SEC projects

Block calls based on what's in them

Same tool, different answer, depending on the details.

  • Email outside recipients

    to: jo@acme.com Allowed

    to: jo@gmail.com Blocked

  • Slack public channels

    post to #team-sales Allowed

    post to #general Blocked

  • GitHub protected branches

    push to fix/login-bug Allowed

    push to main Blocked

Redact tool outputs

AI gets the answer, not the sensitive parts

Before a response goes back to the model, Dtwo finds sensitive values and replaces them. AI can still summarize the ticket, the chart or the code. It just never sees what it doesn't need.

  • PII

    Names, SSNs, phone numbers, card numbers

    AI asks “Summarize ticket #4821”

    AI receives

    Zendesk ticket #4821

    Email: [REDACTED]

    Phone: [REDACTED]

    Issue: “I was charged twice on Oct 12…”

  • PHI

    Patient identifiers, record numbers, diagnoses

    AI asks “When is this patient’s follow-up?”

    AI receives

    Discharge summary

    Patient: [REDACTED]

    MRN: [REDACTED]

    Follow-up in 2 weeks with PCP.

  • Secrets

    API keys, tokens, passwords, connection strings

    AI asks “What log level does prod use?”

    AI receives

    config/prod.env

    AWS_ACCESS_KEY_ID= [REDACTED]

    DB_URL= [REDACTED]

    LOG_LEVEL=info

Dynamic safeguards

Rules that remember what AI already did

A single action can look harmless on its own. Dynamic safeguards look at the whole session, so what AI is allowed to do next depends on what it has already done.

One session in Claude

  1. Step 1

    Reads “M&A Plan.docx”

    Labeled Confidential in SharePoint

    Allowed
  2. Dtwo notes it

    Session now has confidential data

    The rules for the rest of the session tighten.

  3. Step 2

    Posts a summary to #general

    Fine on its own, not after step 1

    Blocked
  4. Step 3

    Sends it to the deal team's private channel

    Allowed

More ways to use them

  • After reading untrusted content

    Once AI has read a public GitHub issue or an inbound email, write actions are off for the rest of the session. A hidden instruction in that content can't make AI change anything.

  • After pulling customer records

    Once AI has exported contacts from the CRM, it can't email or upload files outside the company. Customer data stays in the company, even if someone asks AI to send it out.

  • After touching patient data

    Once AI has opened a patient record, it can only share what it found with people inside the company. Sending to the wrong address doesn't become a privacy breach.

Getting started

Start from the catalog, or say what you want

Write it in plain English

Describe the rule the way you'd say it to a colleague. Dtwo turns it into a policy you can review, test and turn on.

You

Don't let AI send email to anyone outside acme.com, and never let it delete files in OneDrive.

Dtwo
Don't let AI send email to anyone outside acme.comNever let it delete files in OneDrive

Ready-made bundles

Pick a bundle for your team's stack or for a framework you follow, choose the policies you want, and turn them on. No policy code required.

Browse the policy catalog

FAQs

Do I need to know how to write code to create policies?

No. You can turn on ready-made bundles from the policy catalog, or describe a rule in plain English and Dtwo turns it into a policy you can review before it goes live.

What happens when a policy blocks something?

The action doesn't run, and the AI client gets a message saying it was blocked by your organization's policy.

Do policies apply to every AI client?

Yes. Policies run in Dtwo, not in the AI client, so the same rules apply whether someone uses Claude, Cursor, Codex or anything else connected to Dtwo.

Can different teams have different policies?

Yes. Policies can act on the groups you already define in your identity provider, like Microsoft Entra or Okta. For example, engineering can push to feature branches while sales can't touch GitHub at all.

How does redaction know what's sensitive?

Policies look for common patterns like SSNs, card numbers, API keys and patient identifiers. You can also add your own, like internal project names or customer IDs.

Does redacting data make the AI less useful?

Usually not. AI still gets the rest of the response, so it can summarize the ticket, explain the code or answer the question. It just doesn't see sensitive values it doesn't need.

What's the difference between a regular policy and a dynamic safeguard?

A regular policy looks at one action at a time. A dynamic safeguard looks at what AI has already done in the session. Posting to Slack might be fine on its own, but not right after AI has read a confidential file.

Can I see what my policies did?

Yes. Every tool call is logged with who made it and what the policy decided, so you can see what was allowed, blocked or redacted. The data sent to the tool isn't stored.

Say yes to AI, with constraints, and mean it.

Start from a bundle or write your own policies in plain English.

Free Trial
Get in touch

Tell us what you need to govern.

A few lines is plenty — what you're deploying, and what you need control over. We read every one.