Redact, cap, and audit every AI request

Keep one set of compliance rules for every AI request across your Delalify workspace. Beryl Guard redacts personal data before a model receives a request, pauses requests at a team budget cap, blocks restricted topics, and records every decision in an audit log.

Get early accessView all products

The checks that Guard makes on each AI request

Beryl Guard checks each AI request before Beryl sends it to a model. Guard first redacts personal data, financial identifiers, and health information, so the model works only with the redacted text. Guard then blocks the request if the person who makes it, their team, or their project is at a budget cap, or if the request matches a topic that your compliance team restricts. Guard applies the same rules whether a request comes from a draft in Channels, a search in People, or an app that Beryl generates.

Set your AI rules once and enforce them on every request

Beryl passes every AI request to Guard before any model processes it. Guard applies the compliance profile, the budget caps, and the topic restrictions that you set, then records its decision in the audit log, which you can search afterward. The steps below show how you set up each of those rules.

What Guard checks on every AI request

PII redaction across every request illustration

PII redaction across every request

An employee who pastes a customer complaint into an AI tool can include the customer's credit card number, while a sales rep who forwards patient notes for a summary can include health information. If the tool sends that text to a model unchanged, your compliance team will learn about the exposure only once the data is outside your workspace.

Before Beryl forwards a request to a model, Guard masks every credit card number, social security number, passport number, email address, and phone number in it, along with any custom pattern that you define. The model receives the masked version, while Guard stores the original in your workspace logs.

Industry-specific compliance illustration

Industry-specific compliance

Regulations such as HIPAA for health records, FERPA for student records, the GDPR for personal data about people in the EU, PCI DSS for card payments, and SOX for the financial reporting of public companies each require their own redaction patterns, retention rules, and audit records. If several of them apply to your organization, your compliance team will have to set up each one in every AI tool that your employees use.

Guard includes compliance profiles for HIPAA, FERPA, GDPR, PCI DSS, and SOX. When you turn on the profile for your industry, Guard applies that profile's redaction patterns, retention rules, and audit requirements to every AI request in your workspace. For a regulation specific to your jurisdiction, your compliance team can build a custom profile.

Budget controls per team, project, and user

A prompt loop can repeat the same request overnight until your organization owes $3,000 in model usage, while the monthly bill for a new feature that goes live without a budget cap can come to four times the forecast. If you set no cap, your finance team will first see that spending on the next invoice.

Guard pauses requests as soon as spending reaches a hard budget cap, which you can set for each team, project, user, or workflow. Guard also gives your finance team a usage dashboard that shows spending by product, team, and use case, so they can match every part of the AI bill to the team and the use case that incur it.

Topic restrictions and content policies illustration

Topic restrictions and content policies

In a customer-facing chat, you will want a qualified person on your staff to answer legal, medical, and investment questions. A school will want the same for questions about self-harm, so that a counselor answers them in place of the assistant.

You can restrict medical advice, legal counsel, financial recommendations, and any other subject that you name for your product and audience. When a request matches a restricted topic, Guard blocks it, then either returns the fallback message that you write or routes the conversation to the colleague whom you assign in Teams.

Full audit trail on every decision

Questions such as why an AI tool blocks a particular request, or which prompts include sensitive data in a given quarter, can come from your compliance officer or from an auditor. If the only record is a set of raw server logs, someone on your security team will have to piece each answer together by hand.

For every Beryl request, Guard writes an audit log entry that records the person who makes the request, the text that Guard redacts, the policy that applies, the provider that returns the response, and the cost of the request. Your compliance team can search the log by date, user, project, or policy and export the matching entries for an auditor.

Inherited governance for generated apps

When a team uses a Beryl Agent to build an internal tool, that tool sends AI requests of its own. If those requests bypass the rules that apply everywhere else, your compliance team will have to audit each generated tool for gaps such as missing PII redaction.

Every app and workflow that Beryl generates sends its AI requests through Guard, so PII redaction, budget caps, topic restrictions, and audit logging apply to those requests in the same way as to requests from the rest of the platform. Because Beryl also builds those rules into the code that it generates, every deployment of a generated app applies them.

What changes for your compliance and finance teams

Models receive only redacted text

Guard masks personal, financial, and health data in each request and keeps the original text inside your workspace. Because Guard records each redaction in the audit log, your compliance team can show which version of each request Beryl sends to the model.

An AI bill that your finance team can explain

Guard pauses requests at the cap that you set for each team, project, and user, so the monthly AI bill comes to at most the sum of those caps. Your finance team can see in the usage dashboard which teams and use cases account for each part of the bill.

An audit log that Guard writes as each request happens

Guard records each request in a searchable log at the time that it happens, together with the decision that Guard makes and the policy that applies. Whoever prepares the quarterly compliance review can search that log for the evidence that an auditor asks for.

One policy for Beryl and the apps that it generates

The rules that your compliance team sets in Guard apply to Beryl's own features and to every app or workflow that Beryl generates. Your compliance team therefore maintains one policy for every AI request that your organization makes.

How organizations use Beryl Guard

A clinic turns on the HIPAA profile for every AI request

A clinic's physicians use Beryl to draft visit summaries, answer questions from earlier notes, and match each new referral to a specialist. The clinic turns on the HIPAA profile in Guard, so Guard redacts patient names, dates of birth, and medical record numbers from every request before a model processes it. Guard also records each request in the audit log, which the compliance officer reviews every quarter.

Result

Physicians can use Beryl on clinical work because the models process only the redacted text of each request. For each quarterly review, the compliance officer searches the audit log, which contains every request alongside the redactions that Guard makes.

A clinic turns on the HIPAA profile for every AI request

A payments company gives each of its four product teams an AI budget

A payments company's four product teams each use Beryl for a different feature: fraud triage, customer support, identity checks during onboarding, and internal search. The finance lead sets a monthly AI budget for each team in Guard, which alerts the team at 80% of its budget and pauses its requests at 100%. The usage dashboard shows the finance lead the spending for each team, feature, and customer segment.

Result

Each team's AI spending ends the month at or below its budget, because Guard pauses a team's requests at the cap. When the CFO asks what makes up the bill, the finance lead can answer from the dashboard by team, feature, and customer segment.

When a student asks the school assistant about self-harm

A school district gives its students an assistant, built on Beryl, that helps them find resources and answers their course questions. The district restricts self-harm, substance abuse, and personal crisis as topics in Guard, so when a student asks about one of them, Guard blocks the request and routes the student to whichever counselor is on call in Teams.

Result

A student who asks about one of those topics reaches a counselor, who receives the conversation so far along with the referral. The counseling team can follow every referral in Teams, while students keep using the assistant for course questions and resources.

When a student asks the school assistant about self-harm

A law firm reports each quarter on AI requests that involve privileged material

The general counsel of a law firm with six offices on three continents reports each quarter on every AI request that involves attorney-client privileged material, by office, matter type, and regulatory regime. For each report, Guard exports the audit log entries that match the firm's privilege policy, with the source citations, user IDs, and timestamps for each request.

Result

The general counsel builds each quarterly report from that export, so the firm's paralegals no longer collect the records office by office. Because Guard redacts privileged material from every request under that policy, the general counsel can show a regulator, request by request, what Guard removes before each model call.

A law firm reports each quarter on AI requests that involve privileged material

Explore more of Beryl

Orchestrate

Beryl Orchestrate performs each step of a workflow and pauses at sensitive steps until someone approves.

Learn more

Agents

A Beryl agent works toward your goal for up to 24 hours with the permissions of the person who starts it.

Learn more

Context

Beryl Context answers your question from People, Channels, OneCloud, and Teams and links each answer to its source.

Learn more
View all Beryl features

Apply your compliance rules to every AI request

Guard gives your compliance team one place to set redaction, budget, and topic rules for Beryl, and one audit log to search for each review.

Get early accessView all products