Guide · For MSPs and MSSPs
Managed AI observation, without enabling blocking.
How a managed service provider can scope, deliver and explain an observation-first AI service: what the connected sources can and cannot show, who is responsible for what, how a customer stays on monitoring indefinitely, and how selected protection can be added later, only with the customer's approval.
Can an MSP sell AI observation without turning on blocking?
Yes. In Behavry, each customer tenant has an enforcement posture of monitor or enforce. In monitor posture, policy still evaluates each governed action and records what it would have done, but automatic enforcement lets the action through. Discovery connectors only read inventory. So a provider can deliver a findings-led observation service, then an ongoing monitoring service, without blocking anything. Protection is enabled later, action by action, only through the customer's approval and change process.
Three qualifications matter for how you scope and sell it:
- Coverage is only what you connect. Admin-API and identity-provider connectors detect AI services, licences and app grants. Browser and desktop agents show usage on the devices where they are deployed. Individual actions are recorded only on a governed path that an agent is configured to use.
- Detected is not used. A licence or an OAuth grant shows access, not activity. Report it that way.
- "Not blocking" has defined exceptions. Identity, session and tenant-ownership checks, an administrator's emergency stop, and browser-extension local blocking still apply in monitor posture. They are listed in section 006.
The customer cannot govern what it cannot see.
Most customers already run AI they did not formally approve: copilots switched on for a pilot group, third-party apps granted mail and calendar scopes, AI websites used from managed browsers, desktop assistants wired to MCP servers, and small apps built on citizen-code platforms. Often nobody owns the list.
Starting with enforcement in that situation is risky for everyone. A provider that blocks before it understands normal use breaks workflows, loses the customer's trust, and still cannot show what was there. Observation turns the problem into a reviewable inventory and a set of findings, written for the customer's leadership, with the limits stated.
That is useful on its own. Many customers will want to stay there, or move to ongoing monitoring, and never block anything. That is a legitimate end state, not a stalled sale.
What the customer gets from observation
- An inventory of the AI services, grants, apps and agents that the connected sources can see.
- Findings that separate "detected" from "used", with a risk view and a recommendation for each.
- A written statement of what was covered and what was not.
- A clear choice of next step: stay, monitor, or discuss protection for specific actions.
What supported discovery sources can reveal.
Each row is a source type implemented in the Behavry codebase. Which specific connectors are available for a given engagement is confirmed during scoping; see integrations for current status labels.
| Source | What it can show | Access it needs | Highest level reached | What it cannot show |
|---|---|---|---|---|
| Identity provider app grants Okta, Entra ID, Google Workspace | Third-party and AI apps with OAuth grants: app, scopes, principal, consent type, first and last seen | Read-only directory or application access, granted by the customer's admin | Service detected | Whether the app was used. Sign-in and audit logs are not read. |
| SaaS admin APIs Microsoft 365, Google Workspace, Slack, Salesforce, ServiceNow, Zendesk, Atlassian, GitHub | AI feature licensing and enablement, enablement scope, enabled and total user counts, last activity where the platform reports it | Admin-API credentials per platform, as configured | Service detected; usage only where the platform reports activity | Prompts, content or per-user actions |
| Managed browser extension | Visits to AI services, paste and text-input events, DLP pattern findings, and injection findings | Managed deployment of the extension | Usage observed | Unmanaged browsers, personal devices, native apps. The user identifier is a claim unless verified at check-in. |
| Desktop agent (Warden) | LLM API call metadata (provider, model, token counts, whether tools were offered) and MCP server configuration exposure | Device enrolment | Usage observed | Devices without the agent. Configuration exposure does not show that a tool was called. |
| Citizen-code platforms Replit, v0, Vercel, Netlify | Deployed apps, creator, whether the app has an AI agent, database or external API | Platform API access | Service detected | Platforms that are not connected |
| Governed path MCP or API proxy, desktop proxy | Each governed action: agent identity, tenant, tool, arguments digest, policy decision, and a Decision Trace | The agent's tool configuration pointed at Behavry | Action recorded | Actions taken by any path that is not configured through Behavry |
Not a Behavry discovery source
Network, DNS, web-proxy and firewall logs, and identity-provider sign-in or audit logs, are not read by Behavry's discovery connectors. If a customer needs that view, it comes from their existing tools and is stated as outside this service's coverage.
Detecting a service is not observing usage. Neither is recording an action.
Use these three terms consistently in findings and conversations. Most disagreements with a customer about "what you saw" come from mixing them up.
The service, licence, app grant or deployment exists. Example: Copilot is enabled for 40 users, or an app holds Mail.Read.
Does not establish that anyone used it. Do not infer active usage from licensing or access alone.
A source shows people or devices using a service. Example: 12 visits and 3 paste events to an AI site from managed browsers.
Shows activity on covered devices. Does not show individual tool calls or what an agent did in another system.
An individual action was recorded on a configured, governed path, with the policy decision and a Decision Trace.
Only for traffic on that path. Discovery-only connectors never produce tool-call records or action-level Decision Traces.
Prerequisites, required access and coverage limits.
Before the observation window starts
- Agree the sources. Choose which of the sources in section 003 will be connected, for which users, groups and devices.
- Get the access. The customer's administrators grant the read access each connector needs. Record the scopes granted, and the date.
- Set the posture. Confirm the customer tenant is in monitor posture. New tenants default to monitor. Posture is set per tenant and is not inherited by child tenants, so check each one.
- Sign the coverage statement. List what is connected, what each source can show, and what is not covered. The pack includes a template.
- Name the people. Who at the provider reviews, who at the customer receives findings, and who at the customer approves any change.
Coverage limits to state up front
- Only connected sources produce findings. "Available" is not "connected".
- Browser and desktop visibility covers only managed browsers and enrolled devices.
- Some citizen-code platforms have no working connector yet. Connected platforms today are Replit, v0, Vercel and Netlify.
- The 30-day "still ungoverned" alert applies to citizen-code apps only.
- Behavry has no built-in inventory export. The inventory in the pack is a manually prepared deliverable assembled from the product's list views. Audit events can be exported as JSON, CSV, CEF or LEEF.
What observation and monitoring do not prevent.
Say this plainly in the proposal and in the coverage statement. A customer who believes they are protected when they are being observed is the worst outcome for everyone.
- They do not stop a risky action. In monitor posture, if a policy would deny or hold a governed action, Behavry records the would-be verdict and raises a
monitor_would_blockevent, and the action proceeds. - They do not revoke, disable or change anything. A finding about an over-scoped app grant is a recommendation. Revoking it is the customer's decision and action.
- They do not see what is not connected. They cover nothing outside the connected sources, on devices without an agent, or on paths that are not governed.
- They do not fail closed. In monitor posture, if the policy engine is unavailable, governed actions continue and are recorded where the path allows.
What still applies in monitor posture
Monitor posture relaxes automatic policy enforcement. By design, it does not relax:
- agent authentication, session and tenant-ownership checks on a governed path, which are identity, not policy;
- calls to an MCP server that is not registered, which have nowhere to be forwarded;
- an administrator's own stop, suspend, quarantine or emergency lockdown;
- the browser extension's local blocking, which is configured in the browser.
If any of these are in place for a customer, list them in the coverage statement, so that "observation" is never described as "nothing is ever refused".
Behavry operates the platform. The provider manages the customer. The customer approves.
This is the intended split. The exact responsibilities for each engagement are agreed during scoping and written into the monitoring handoff.
| Activity | Behavry | Provider (MSP / MSSP) | Customer |
|---|---|---|---|
| Operate and maintain the hosted platform | Responsible | Informed | Informed |
| Connect and maintain the agreed sources | Supports | Responsible | Grants the access |
| Review findings and alerts | Responsible | Receives | |
| Recommend an action | Responsible | Accepts or declines | |
| Approve any change in a customer system | Consulted | Approves (named approver) | |
| Make the change | Only if the customer asks, under its change process | Accountable | |
| Access the evidence | Yes | Yes, retained by the customer |
Inside the product, a provider works across its customers through partner and child tenants, with roles that include an MSP operator role. Behavry does not have a dedicated "approver" role. Approval of a held action can be given by any user with a writable role, so name the customer's approver in the handoff and set roles accordingly.
How a customer stays on ongoing monitoring.
Monitoring adds agreed alerts, a review cadence, evidence summaries and defined response responsibilities to observation. It changes who looks, and how often. It does not change what is allowed.
- Keep the posture on monitor. Changing a tenant's posture is an administrative action recorded on the signed audit chain. Treat it as a change that needs the customer's approval.
- Agree the signals. Typical signals:
- a policy would have blocked or held an action on a governed path (
monitor_would_block); - DLP findings on AI input from managed browsers;
- a new app grant with sensitive scopes;
- a citizen-code app still ungoverned at 30 days;
- a new MCP configuration on a device.
- a policy would have blocked or held an action on a governed path (
- Route them. Alerts can go to the customer's or provider's SIEM (Splunk, Microsoft Sentinel, QRadar, Chronicle, syslog, OTLP or webhook) and to Slack or email. Alerts carry a severity and a status of open, acknowledged or resolved.
- Review on the agreed cadence. A person reviews each alert and may recommend an action. The customer's named approver decides what happens. Nothing is blocked, and no customer system changes automatically.
- Report. Send an evidence summary on the agreed cadence, with the coverage statement re-issued when sources change.
Reviewing is not blocking
The output of monitoring is a reviewed alert and a recommendation. The decision to change anything stays with the customer.
If, and only if, the customer wants it: selected protection.
Protection is optional and narrow. It applies to specific actions on supported, governed paths, and it is enabled through the customer's own change process.
- Propose specific controls. Each proposal names the action, the path, and the outcome: allow, hold for a named human, or deny. Base it on what monitoring showed would have happened.
- Get written approval. The customer's named approver accepts each control through their change process.
- Enable for the agreed scope. Posture is set per tenant, and some controls (for example, delegation checks) have their own observe or enforce switch, so confirm which mode each control is in after the change.
- Agree rollback. Record how to return to monitor posture, who can do it, and how quickly.
- Evidence each decision. On governed paths, enforced decisions produce a Decision Trace the customer can verify. How Decision Traces work.
For an AI technician acting inside customer tenants, see the companion guide: how to test AI technician tenant and ticket boundaries.
Download the example pack.
Four working templates for an observation engagement. They are manually prepared service deliverables, not Behavry product exports. Sample rows and the sample report use a fictional customer and synthetic data, labelled on every page.
- 01 · Customer AI inventory template (CSV)Manual · columns map to product fieldsOne row per service, grant, app or agent. It has a detection-level column (service detected / usage observed / action recorded), and six synthetic sample rows.
- 01b · Inventory field guide (CSV)ReferenceWhere each column comes from (product field name or manual), allowed values, and what it does and does not prove.
- 02 · Sample findings report (Markdown)Synthetic · illustrative layoutSummary, sources reviewed, six findings with detection levels, what the report does not establish, and next-step options.
- 03 · Coverage and limitations statement (Markdown)TemplateConnected sources, access granted, the three visibility levels, what is not covered, and what observation does not do.
- 04 · Monitoring handoff (Markdown)TemplateSignals, routing, and a responsibility matrix for who reviews, recommends, approves and responds. It also covers how protection would later be approved.
What the pack is not
It is not a screenshot of the Behavry console, and not an automatically generated report. It contains no customer result. Report cadence, alert scope and response times are agreed in scoping. None is published as a service level.
The evidence behind this guide.
Each product statement above was checked against the Behavry codebase and its automated tests on 13 November 2026. Implemented and tested is not the same as deployed for a given customer, or validated end-to-end in their environment.
| Statement | Evidence | Limit |
|---|---|---|
| Per-tenant monitor posture records would-be verdicts and lets the call run | Implemented; automated tests passed (monitor mode and proxy-in-monitor suites) | Mocked targets. The exceptions are listed in section 006. |
| Identity-provider, SaaS admin-API, browser, desktop and citizen-code sources | Implemented; connector and service unit tests passed | Live connector syncs against a real customer tenant were not run for this guide. |
| Alert fields, statuses and SIEM / notification routing | Implemented; alert scope and SIEM forwarder tests passed | The routing used for each customer is confirmed in scoping. |
| Partner and child tenants, MSP operator role, tenant isolation | Implemented; partner, role and isolation tests passed | Row-level-security policy tests were skipped in the local run. |
| No inventory export | Code review: list views and audit export exist; no inventory export endpoint | The pack's inventory is a manual deliverable. |
- MSPS & MSSPS
The observe, monitor, protect service model, and the engagement enquiry.
- COVERAGE
The supported agent surfaces, on the paths you connect.
- BOUNDARY TESTS
Test an AI technician's tenant and ticket boundaries before you trust it.
Start with what the customer can see.
DISCUSS AN OBSERVATION ENGAGEMENT →Check supported integrations →