Data loss prevention decides whether sensitive data may cross a boundary, but once a file has left, DLP has nothing more to say; information rights management travels with the file and keeps deciding who can open it. The two are often sold together and configured from the same console, which hides the fact that they answer different threats, fail in different ways and cost users different amounts of friction. The arrival of AI assistants that read whatever the user can read has sharpened the question, because encryption and usage rights now also decide what a copilot can summarise. This guide compares IRM and DLP as control models: threat model, architecture, failure modes, key management, user friction, AI tool access and a decision method. It is written for data security leads, architects and anyone deciding which files deserve persistent protection. Platforms are listed alphabetically where named, and naming is illustrative, not a recommendation.
Key takeaways
- DLP is boundary control; IRM is persistent control: DLP inspects content at egress points and blocks, warns or audits, while IRM encrypts the file and binds usage rights to identities so protection survives exfiltration.
- Each blinds the other if you are careless: strongly encrypted files cannot be inspected by content scanning DLP, so plan DLP rules on labels and encryption status, not only on content patterns.
- Usage rights now govern AI access: in Microsoft Purview, the VIEW right lets an AI assistant reference a protected item but only EXTRACT lets it summarise; Google states Gemini does not retrieve files where IRM restricts download, print and copy.
- Whoever holds the key holds the control: provider managed keys keep service features working; customer held keys (Double Key Encryption, client-side encryption) keep the provider and its AI services out, at the cost of search, eDiscovery and DLP inspection.
- Use IRM for a small, high value set: apply persistent protection to the files whose leak would matter after they leave, and let DLP cover the broad estate.
The short answer
Choose DLP when the risk is data leaving through channels you control (email, web upload, cloud sharing, endpoints, AI prompts) and you want broad coverage with tunable friction. Choose IRM when the risk is what happens after the data leaves: a forwarded board pack, a design file with a supplier, a laptop that walks away, a former employee's personal copy. Most mature programmes run both, with classification labels as the shared language. The mechanics of each are covered in the Data Loss Prevention (DLP) Guide and the Information Rights Management and Persistent File Protection Guide; this page covers the choice and the seams between them. For posture scanning versus enforcement, see DSPM vs DLP.
| Dimension | DLP (boundary control) | IRM (persistent protection) |
|---|---|---|
| Core question | May this content leave through this channel to this destination? | May this identity open, edit, copy, print or forward this item, wherever it is? |
| Enforcement point | Egress points: mail flow, cloud sharing, endpoint agent, browser, network proxy | The application that opens the file, after obtaining a licence from the rights service |
| Protection after exfiltration | None once the copy is out | Continues: the copy is ciphertext without a valid licence |
| Detection basis | Content inspection (patterns, dictionaries, fingerprints, classifiers), labels, context | Label or template chosen by user or policy; no inspection at use time |
| Typical actions | Audit, policy tip, block with override, block, quarantine | Grant or deny rights such as view, edit, copy, print, forward; expiry; revocation |
| User friction | At the moment of sharing, scaled by rule | At every open for recipients, especially external and on unsupported apps |
| Coverage | Broad: whole estate and many channels | Narrow: supported file types and apps |
| Main failure | False positives drive overrides; unmonitored channels leak | Unsupported apps, broken collaboration, key and licence outages |
DLP in depth: threat model, architecture and failure modes
Threat model: DLP targets accidental and careless disclosure first (wrong recipient, public sharing link, personal cloud upload, pasting into a public AI tool) and deliberate insider exfiltration second. It is weakest against a determined insider or attacker who uses a channel without inspection, or who transforms the data so patterns no longer match.
Architecture in words: a policy engine evaluates content and context at each enforcement point. Microsoft Purview DLP, for example, covers Exchange, SharePoint, OneDrive, Teams, Windows and recent macOS endpoints, on-premises shares through a scanner, some non-Microsoft cloud apps, Microsoft 365 Copilot and Copilot Chat, and inline web traffic to generative AI sites through the browser or network. Detection combines keywords, regular expressions with validation functions, proximity of supporting evidence and machine learning classifiers. Actions range from policy tips and audit to block with override and block, and policies can run in simulation mode, where no action is applied, before enforcement. Google Workspace DLP policies can apply IRM controls to sensitive files that match a rule.
Failure modes:
- Override fatigue: rules that fire on every spreadsheet teach users to click "override with justification" without reading.
- Channel gaps: an unmanaged device, a sanctioned app without a connector, or an AI agent using an API token instead of a browser.
- Encrypted payloads: content already encrypted by IRM, client-side encryption or a password protected archive cannot be inspected; rules must fall back to labels, file type or encryption status.
- Ends at the boundary: once a permitted transfer has happened (a legitimate email to a supplier), DLP has no further say over that copy.
Deeper coverage of endpoint and AI channel DLP is in the Endpoint DLP Guide and GenAI DLP vs Traditional DLP.
IRM in depth: threat model, architecture and failure modes
Threat model: IRM assumes the file will leave your control, through legitimate sharing, forwarding, device loss or exfiltration, and makes the copy useless to anyone who is not authorised. It also limits what authorised users can do: read but not print, or edit but not forward.
Architecture in words: when a user applies an encrypting label, the client encrypts the content and embeds a policy naming who has which rights. To open it, a recipient's application authenticates to the rights service, which checks the policy and issues a use licence containing the content key and rights. In Microsoft Purview's Rights Management service, that use licence is valid offline for a configurable period (tenant default 30 days), after which the user must reauthenticate; label changes take effect on already opened content when the current licence expires. Usage rights include VIEW, DOCEDIT, EDIT (save), EXPORT (save as), EXTRACT (copy, including screen captures), PRINT, FORWARD, OBJMODEL (macros and programmatic access) and OWNER (full control, including removing protection). Microsoft notes that applications implement usage rights differently and recommends testing before production. Google Workspace offers IRM controls on Drive files that restrict downloading, printing and copying, and DLP policies can apply them automatically.
Failure modes:
- Rights are not a screen lock: anyone who can view content can photograph it, and the EXTRACT right explicitly includes screen captures. IRM raises the cost of leakage; it does not prevent a determined viewer from recording what they see.
- Over broad grants: "all employees: co-author" templates give EXTRACT and EXPORT to thousands, which some apps treat as permission to save an unprotected copy.
- External recipient friction: partners without a compatible identity or app cannot open files, so users fall back to unprotected copies through personal email.
- Owner sprawl: the OWNER right lets a holder remove protection; grant it narrowly.
- Revocation lag: revoking access or tightening a label does not reach a copy until the cached licence expires, so long offline validity weakens revocation.
Protection after exfiltration: the deciding question
Ask for each data set: if a copy were outside our tenant tomorrow, would it still matter? Board papers, merger documents, product designs, source exports, legal advice and bulk personal data usually answer yes. Routine operational documents usually answer no, and boundary DLP plus access control is sufficient. The table below maps common leak paths to which control helps.
| Leak path | DLP effect | IRM effect |
|---|---|---|
| Email to wrong external recipient | Can warn or block on content or label before send | Recipient cannot open if not in the policy |
| Forward by an authorised recipient | Applies only if the forwarder is inside your DLP scope | FORWARD right and identity binding still apply |
| Upload to personal cloud or public AI site | Can block on managed device or browser | Uploaded copy stays encrypted and unreadable to the service |
| Lost or stolen laptop | No effect on data at rest | Files need a licence; disk encryption still needed for the rest |
| Departing employee keeps copies | May detect bulk copy to USB or cloud | Access ends when account is disabled and licences expire |
| Attacker with a stolen session | May detect unusual egress volume | Attacker inherits the user's rights while the session is valid |
| Screen capture or photograph | Endpoint controls may restrict capture | Limited: viewing implies the content can be recorded |
The last two rows matter: neither control helps much once an attacker is the user. Identity protection, session controls and token hygiene carry that risk; see the Data Classification Guide for deciding which data sets justify the stronger controls.
AI tool access to protected files
AI assistants act with the user's permissions, so the controls you set for people now also set what the assistant can read. The two models behave differently:
- Google Workspace: Google states that Gemini retrieves only data the user can already access, that Drive trust rules can further limit retrieval, and that where IRM restricts download, print and copy, Gemini does not retrieve those files to generate a response. With client-side encryption, Google states that protected data is unreadable to Google and to assistants like Gemini. Admins can review Gemini audit logs showing when Gemini accessed a Drive file. See the Gemini for Google Workspace Security Guide.
- Microsoft 365: Microsoft's usage rights documentation states that VIEW lets Microsoft 365 Copilot and other AI apps reference an item with a link but not summarise it, while EXTRACT lets Copilot summarise the content for the requesting user. Separately, a Purview DLP policy can block Microsoft 365 Copilot and Copilot Chat from processing items carrying selected sensitivity labels. Content protected with Double Key Encryption is not accessible to Copilot at rest. See the Microsoft 365 Copilot Security Deep Dive.
The design lessons are platform neutral. Use usage rights deliberately: a "view only" label is now also an "AI cannot summarise" label on some platforms, which may be exactly what you want for legal advice and unwelcome for a sales playbook. Use a DLP rule on labels to keep AI processing away from defined classes even where users hold full rights. Reserve customer held keys for data that no provider service, AI included, should ever read. AI agents that call APIs with service credentials are not users with labels in a viewer: inventory those non-human identities and scope their tokens, because a broadly scoped agent token can read every file its service account can. The Enterprise AI Copilot Security Guide covers the wider control set.
Key management and its trade-offs
| Key model | Who can decrypt | What keeps working | What stops working |
|---|---|---|---|
| Provider managed key | Provider services on your behalf, plus authorised users | Search, eDiscovery, DLP inspection, AI assistants within rights | Nothing; least separation from the provider |
| Customer managed root key (BYOK) | As above; you control key lifecycle in the provider's key service | Same as provider managed | Service if you revoke or lose the key |
| Customer held second key (Microsoft Double Key Encryption) | Only clients with access to both your key service and the Azure key | Desktop apps on supported clients; protected viewer | Copilot, eDiscovery, content search, SharePoint and OneDrive DLP and search, coauthoring, Office for the web |
| Customer held key (Google client-side encryption) | Only clients authorised by your external key service or HSM | Encrypted editing in supported apps; Vault retention and export | Content DLP scanning, Vault content search, malware scanning, Gemini access |
Availability notes: Microsoft lists Double Key Encryption as included with Microsoft 365 E5 and supported in Microsoft 365 Apps for enterprise on Windows, and estimates it for a small share of data (around 5%). Google lists client-side encryption for Frontline Plus, Enterprise Plus, and Education Standard and Plus editions; Google's options include partner key services, Cloud HSM for Google Workspace and a customer built service using the CSE API.
The trade-off is direct: every service that cannot read the content cannot help you with it. Customer held key data needs its own eDiscovery, retention and incident response plan, and DLP rules for it must work on metadata and encryption status. See the BYOK, HYOK and External Key Management Guide.
User friction and adoption
- DLP friction is bursty: it lands at share time and can be tuned per rule from audit to block. Start in simulation, then policy tips, then block with override, then block.
- IRM friction is constant: every recipient pays it at every open. Prefer automatic or default labels for defined repositories over asking users to choose.
- External collaboration decides success: test opening protected files from each major partner's identity and app stack before launch.
- Collaboration features: coauthoring and web editing behave differently with protected files, so pilot with the actual teams who share them.
Worked scenario: a merger data room and an AI rollout
A company is preparing an acquisition while rolling out an AI assistant to all staff. The deal team shares valuations with two external advisers.
- Classify: deal documents receive a "Restricted: Deal" label that encrypts with named internal users and two adviser domains. Rights: internal team co-author; advisers view and print, no copy, no forward. Offline licence validity is shortened to seven days so revocation lands quickly.
- AI posture: because advisers hold only view style rights, assistants acting for them cannot summarise the files on platforms that tie summarisation to copy rights. A DLP rule also blocks AI processing of the "Restricted: Deal" label for internal users, so the deal team's assistants do not ground answers on the documents during the confidential phase.
- Boundary rules: DLP blocks the label from leaving to any domain outside the two advisers and alerts on bulk download of the deal site.
- Customer held keys: the board considered them and rejected them for this phase because eDiscovery and DLP inspection would be lost during a period of likely litigation holds.
- Close: after announcement, the label is relaxed and access for one adviser revoked; their cached licences expire within seven days.
Decision checklist
- Does the data matter after it leaves? Yes: IRM for that set. No: DLP and access control.
- Can every recipient open protected files? If not, fix identity federation first or users will route around it.
- Must provider services, including AI and eDiscovery, read it? Yes: provider or customer managed root key. No: customer held key.
- Is the risk mainly accidental sharing? DLP with policy tips gives most of the value at least friction.
- Small organisation? Start with DLP on two or three data types and one encrypting label for the most sensitive repository; add more only with evidence.
- Large organisation? Treat labels as the shared language for DLP, IRM, AI controls and DSPM, and govern the label taxonomy centrally.
Detection and measurement
- Override rate: share of DLP block with override events that are overridden, by rule; high rates mean the rule needs tuning.
- Protected coverage: percentage of files in defined high value repositories carrying an encrypting label.
- Protection removal events: count of label downgrades and protection removals, with justification review.
- Open failures: rights service denials for legitimate recipients, a direct friction signal.
- AI access to labelled content: count of AI assistant access events on restricted labels from platform audit logs.
- Licence validity: distribution of offline validity settings across labels, against your revocation target.
Common failure patterns
- Encrypting everything: blanket encryption breaks search, DLP inspection and collaboration, and users learn to strip labels.
- DLP rules that only read content: protected files sail through because the scanner cannot see inside them.
- Co-author for all: wide EXTRACT and EXPORT grants that undo the protection and also allow AI summarisation.
- Ignoring the external path: partners cannot open files, so business moves to unprotected channels.
- Forgetting non-human readers: agents and integrations with broad API scopes read protected repositories outside the label model.
Framework mapping
| Framework | Reference | Relevance |
|---|---|---|
| NIST CSF 2.0 | PR.DS-01, PR.DS-02, PR.DS-10 | Protect data at rest, in transit and in use |
| NIST SP 800-53 Rev. 5 | AC-4, SC-7 | Information flow enforcement and boundary protection (DLP) |
| NIST SP 800-53 Rev. 5 | AC-16, AC-21, SC-28, SC-12 | Security attributes (labels), information sharing, protection at rest, key management (IRM) |
| NIST SP 800-53 Rev. 5 | SI-4 | System monitoring for exfiltration indicators |
| ISO/IEC 27001:2022 Annex A | 5.12, 5.13, 8.12, 8.24 | Classification, labelling, data leakage prevention, use of cryptography |
| CIS Controls v8.1 | Safeguard 3.13 | Deploy a data loss prevention solution |
Practitioner checklist
- Agree one label taxonomy used by DLP, IRM and AI controls.
- Identify the high value data sets that matter after exfiltration and protect only those with encryption.
- Write DLP rules on labels and encryption status as well as content.
- Define rights per label deliberately, treating copy rights as AI summarisation rights.
- Set offline licence validity per label to match your revocation needs.
- Choose key models per data set and plan eDiscovery and incident response for customer held keys.
- Test external recipients and collaboration before enforcement.
- Inventory AI agents and integrations that read files with service credentials.
- Run DLP in simulation, then tips, then block, reviewing override rates at each step.
Standards and references
- Microsoft Learn: Configure usage rights for Azure Rights Management
- Microsoft Learn: Double Key Encryption
- Microsoft Learn: Learn about data loss prevention
- Microsoft Learn: Considerations for DSPM for AI
- Google Workspace: Client-side encryption FAQ
- Google Workspace: Enterprise security controls for Google Workspace with Gemini
- NIST: SP 800-53 Rev. 5
- NIST: Cybersecurity Framework 2.0
Related NHI Mgmt Group resources: Information Rights Management and Persistent File Protection Guide, Data Loss Prevention (DLP) Guide, DSPM vs DLP, Google Workspace Security Guide, Microsoft Purview Information Protection Guide.