For MSPs and MSSPs

A managed AI security service, starting with observation.

Start with AI visibility. Help a customer understand how AI is being used across their business. Then build a recurring managed service, expanding into monitoring and response as the customer is ready — with Behavry operating the platform underneath and you managing the customer.

ObserveEstablish visibility into AI activity within the coverage you agree. Useful on its own.
MonitorAdd ongoing review and alerts on the signals you agree to watch. Nothing is blocked.
ProtectWith customer approval, enable controls on selected actions — allow, hold for a named human, or deny.
ProveA signed Decision Trace for each governed action, verifiable by the customer.
/001 · TWO OPPORTUNITIESYOUR AGENTS · THEIR AGENTS

Two opportunities. One control plane.

The first is about how you deliver service. The second is a service you can sell. Both run on the same policy model and the same evidence format. Labor savings and new revenue are outcomes to measure in a pilot, not figures we promise in advance.

Opportunity 01 · Your operations

Govern your own AI technicians

Behavry governs provider-owned agents acting inside customer systems.

Your AI technicians, your ticket queue, your partner access into many customer tenants. Behavry enforces each customer's boundary and the authority it delegated on every governed action, so one agent can serve many customers without one customer's authority leaking into the next. When a customer asks what your AI did in their tenant, you hand them a signed record instead of a log export.

  • Per-customer scope on shared agents and shared credentials
  • Delegated authority mapped from the contract into policy, not inferred from the admin role
  • Escalation to your service desk lead, not to the customer
  • Evidence for the customer, the cyber insurer, and your own counsel
Opportunity 02 · Your service catalog

Sell managed AI governance to customers

Behavry discovers and governs the customer's own AI tools, agents, MCP servers, browser AI usage, and privileged integrations.

Your customers are already running copilots, agents, MCP servers, and browser AI with access to their mail, files, and admin consoles. Often nobody in their organization is watching it. Behavry finds what is visible through the agreed coverage, supports ongoing monitoring of it, and, where the customer approves, puts selected high-risk actions behind enforcement. You package that as a recurring service.

  • Discovery across agreed sources: IdP, SaaS admin APIs, endpoints, browsers
  • Prioritized findings the customer's leadership can read
  • Optional enforcement on write-capable and privileged AI, once the customer approves it
  • Evidence reports on a cadence agreed with the customer
/002 · THE SERVICE MODELOBSERVE · MONITOR · PROTECT

Begin with observation. Add protection only when the customer approves it.

The buying path is a progression, not an all-at-once commitment, and a customer can stay on observation or ongoing monitoring without ever enabling blocking. Observation gives visibility and findings within agreed, supported coverage. Monitoring adds ongoing review, agreed alerts, evidence summaries, and defined response responsibilities. Protection enables selected enforcement controls through an explicit approval and change process.

Reviewing is not blocking. In observation and monitoring, an alert is reviewed by a person, who may recommend an action; the customer or its named approver decides what happens next. Nothing is blocked, and no customer system is changed, automatically. Blocking or holding an action happens only where protection has been enabled for it.

Proposed packagingThe stages below are a proposed way to deliver that progression. Availability, inclusions, commercial terms, and delivery responsibilities are confirmed with each partner before anything is offered as a purchasable service. Report cadence, alert scope, and response responsibilities are agreed during scoping.

Stage 01 · Observe Observation

Visibility and findings within agreed, supported coverage.

  • AI apps and copilots
  • Agents, hosted and local
  • MCP servers and browser AI tools
  • Privileged and write-capable integrations
  • Prioritized findings and recommended authority boundaries
  • Does not block anything

Customer getsAn inventory and findings for what the connected sources can see, with the limits stated

Your roleFixed-scope discovery and assessment

How to scope and deliver it, with templates →

Stage 02 · Monitor Monitoring

Ongoing oversight, without enabling blocking.

  • Ongoing review of the covered inventory and activity
  • Alerts on the signals you agree to watch
  • Evidence summaries for the customer
  • Defined response responsibilities: who reviews, who recommends, who approves
  • Alerts are reviewed, not acted on automatically

Customer getsContinuing visibility and recommendations, with their systems unchanged

Your roleRecurring review and reporting. Cadence and response scope agreed during scoping

Stage 03 · Protect Protection

Selected controls, enabled through explicit approval and change.

  • Controls chosen action by action, approved by the customer
  • Enforcement on supported paths: Warden desktop proxy, MCP and API, browser
  • Allow, hold for a named human, or deny
  • Decision Trace on each governed action

Customer getsEnforcement on the actions they chose to protect, and evidence for each decision

Your rolePolicy operations and exception handling, within the agreed change process

Across customers Operating view

Run the service across your customers from one view.

  • Parent and child tenants
  • Findings and alerts across your customers
  • Held and blocked actions, where protection is on
  • Reports, per customer

Customer getsA provider who can see and act across their environment, within agreed access

Your roleThe operating console behind the service. Cross-tenant capabilities confirmed per partner

Plus · Autonomy Review Autonomy Review

Proposed partner serviceBegins in pilot and shadow mode

A review of whether a customer's agent is ready for a broader, scoped grant: the evidence from its Decision Traces, the criteria the customer set, and a recommendation the customer approves or declines. Delivered through the operating model above, not as a separate platform product.

Customer gets
An evidence assessment and a recommendation for a scoped grant
Your role
A recurring review, or a separately scoped readiness engagement
What it does not do
Change permissions. A recommendation is not an approval, and the first milestone runs in shadow mode

How Earned Autonomy works →

Buy what you need/Observation and monitoring stand alone/Protection only with customer approval/Autonomy Review is optional

/002B · THE FIRST 30 DAYSOBSERVATION · NOTHING BLOCKED

What an MSP delivers in the first 30 days of AI observation

An observation engagement in three steps. Nothing is blocked and no customer system is changed during these 30 days. The schedule and responsibilities below are a proposed starting point, agreed with each partner and customer during scoping.

Step 01 · Days 1–5

Agree scope and confirm reporting sources. Agree which users, groups and devices are in scope, which sources will be connected, and what each one can and cannot show. Record it in a coverage statement the customer signs.

Owners. The customer’s admin approves the scope and grants read-only access. The MSP leads scoping and writes the coverage statement. Behavry provisions the tenant and supports the connectors.

Step 02 · Days 6–20

Review observed AI use and coverage gaps. Review the inventory and activity the connected sources report. Keep three levels apart: a service detected, usage observed, and an action recorded. Log what is not covered, such as devices without the agent, unmanaged browsers and sources that are not connected.

Owners. The MSP’s reviewer does the review. The customer confirms which tools are sanctioned. Behavry operates the platform and gives technical support.

Step 03 · Days 21–30

Present a customer readout and recommend next steps. Present the findings, the coverage limits and the recommendations. The customer chooses to stop, continue observation, move to monitoring, or ask to scope protection for selected actions.

Owners. The MSP presents the readout and makes the recommendations. The customer decides.

Enforcement is a separate decision. A readout can recommend protection; it cannot enable it. Moving any action to enforcement needs the customer’s explicit approval, action by action, through the change process you agree. Until then, nothing is held or denied.

What each source can observe

Only connected sources produce findings. Which connectors are available for an engagement is confirmed during scoping.

SourceWhat it can showWhat it cannot show
Identity provider app grants
Okta, Entra ID, Google Workspace
AI and third-party apps with OAuth grants: the app, its scopes, who granted it, consent type, first and last seen. Level: service detected.Whether the app was used. Sign-in and audit logs are not read.
SaaS admin APIs
Microsoft 365, Google Workspace, Slack and others
Which AI features are licensed and enabled, for how many users, and last activity where the platform reports it. Level: service detected; usage only where the platform reports activity.Prompts, content or per-user actions.
Managed browser extensionVisits to AI services, paste and text-input events, DLP pattern findings. Level: usage observed.Unmanaged browsers, personal devices and native apps.
Desktop agent (Warden)LLM API call metadata (provider, model, token counts) and MCP server configurations on the device. Level: usage observed.Devices without the agent. A configured MCP server does not show that a tool was called.
Governed path
MCP, API or desktop proxy
Each action routed through it: agent identity, tool, policy decision and a Decision Trace. In observation, actions are recorded, not blocked. Level: action recorded.Actions taken by any path not configured through Behavry.

An example day-30 readout

Illustrative, with synthetic data. A customer with 120 staff, in scope through the identity provider, Microsoft 365, the managed browser extension and the desktop agent. Not a product screenshot or a customer result.

FindingSource · levelRecommended next stepOwner
4 AI apps hold OAuth grants to mail or files; 2 were granted by individual usersIdentity provider · service detectedReview the 2 user-granted apps and keep or revoke eachMSP prepares the list; customer IT lead decides
An AI assistant feature is enabled for 38 of 120 usersMicrosoft 365 admin API · service detectedConfirm the enabled group matches the approved groupMSP checks; customer approves any change
212 visits and 14 paste events to 3 AI services; 2 DLP pattern findingsBrowser extension · usage observedAgree an acceptable-use note and which signals to alert on in monitoringMSP recommends; customer approves
2 developer laptops have MCP servers configured, one pointing at a file-share serverDesktop agent · usage observedConsider routing those agents through a governed path so their actions are recordedMSP recommends; customer decides
31 devices have no desktop agent; personal devices and unmanaged browsers are not coveredCoverage gapDecide whether to extend coverage before monitoring startsCustomer decides; MSP scopes

Recommendation in this example: move to monitoring, with alerts on the user-granted apps and the browser DLP findings. No enforcement is proposed. If the customer later wants protection for a selected action, that is scoped and approved separately. For the coverage statement and readout templates, discuss an observation engagement.

/003 · THE WORKREAD · VERSUS · EXECUTED

Reading a ticket carries risk. Acting on it adds privilege.

AI that reads and summarizes a ticket can still expose sensitive information, which is why it belongs in what you observe. AI that resolves the ticket can save far more time, but it adds privilege and operational risk: each step is an administrative change inside an environment you do not own.

Action the AI needs to takeWhere it runsWhat it needs to have
Reset a passwordEntra ID, Active DirectoryCredential write on a customer identity
Disable a stale accountEntra ID, Google WorkspaceAccount lifecycle rights, and the offboarding context to use them
Update Microsoft 365 settingsM365 admin center, Exchange OnlineTenant-level admin scope, delegated through partner access
Quarantine an endpointRMM, EDRIsolation rights on a production machine
Modify groups and licensesEntra ID, M365Group membership and license assignment across the tenant
Triage a phishing ticketMailbox, EDR, PSAMailbox read, message purge, and ticket write
Update an RMM policyRMMFleet-wide policy write across an entire customer
Recommend or execute remediationEDR, firewall, RMMChange rights with no ticket-bound limit
Close the ticketPSAThe authority to say the work is done and billable

Every row is privileged. Every row is a different customer. The model does not know the contract.

The agent that resets a password at Acme has the same credentials, the same prompt, and the same blind spots when it moves to Smith Financial ten seconds later. Nothing in the model knows where one customer's authority ends and the next one's begins. That is not an AI safety problem. It is a service delivery problem with a liability attached.

/004 · THE CONTROL LAYERALLOW · ESCALATE · DENY

Behavry sits between the AI technician and the customer system.

Not the prompt. The action: the Graph API call, the PowerShell, the MCP tool invocation, the admin-center write. Each governed call is evaluated before it runs against the boundaries you configured, and gets one of three answers.

Allow

In scope, in authority, in policy. The action runs, and the record shows it did.

Escalate

Risk above the delegated line. Held for a human with the ticket, the diff, and the reason attached.

Deny

Outside the customer boundary, the contract, or the credential scope. Stopped on the governed path before it runs.

Ticket, contract, and customer-scope checks work from explicit mappings you configure. Behavry does not read prose contracts. Which PSA, RMM, and identity systems can be connected is confirmed for each pilot.

Works with/MCP tool calls/API proxies/Browser AI/Desktop agents/integrations →

/004B · A WORKED EXAMPLEILLUSTRATIVE · ONE ACTION, THREE STAGES

One password reset, through each stage.

An AI IT technician performing an administrative action — a password reset — and what each stage of the service does with it. This is an illustrative example, not a product screenshot or a customer result. What can actually be observed, monitored, or enforced depends on the supported integrations and the paths you configure.

The actionWhat the stage does
ContextCustomer tenant Acme; ticket #4821, with the requester as identified by the ticket system; requested action: reset the password for j.rivera@acme.com in Entra ID, attempted by the provider's AI IT technician.
Observation recordsThat the agent attempted entra.reset_password on j.rivera, from ticket #4821, under the partner's delegated access to Acme — identity, target, tool call, time, and tenant. Observation records the attempt; on its own it does not block anything.
Monitoring could flagA reset outside the ticket's named user, a burst of resets across many accounts, a reset on a privileged or executive account, or an attempt against a tenant outside the agreed scope — raised as an alert for the provider's reviewer.
Enabled policy couldPermit the reset when the target matches the ticket's named user, as mapped from the ticket system; hold for a named human when the account holds a privileged role; deny when the target is outside the ticket, the customer boundary, or the credential scope.
Evidence retainedA signed, hash-chained Decision Trace: the request, the policy decision and the rule behind it, the delegation chain back to a human, and the outcome. The customer can verify its signature and hash chain with no Behavry account; what verification proves and does not.
Human approvalRequired wherever policy holds the action — a named operator approves or declines before it runs. A recommendation or a held action is never a self-approval.

Enforcement (permit, hold, or deny) applies only once the Protect stage is enabled on a configured path. In an observation-only engagement, the reset is recorded, not blocked. Test these boundaries yourself: a 12-case matrix →

/005 · THE CONSOLEILLUSTRATIVE MOCKUP · EXAMPLE DATA

What the console looks like on a Tuesday.

An illustration of the operating view, drawn with made-up data for a provider with 147 customers. It is not a product screenshot, and cross-tenant console capabilities are confirmed per partner.

Four customers you can call today. With evidence.

Each row is a conversation an account manager can open, with the evidence behind it.

/006 · RECURRING DELIVERYMONITORING · PROTECTION · PROPOSED SCOPE

Behavry runs the platform. You manage the customer.

A recurring engagement that starts with observation, continues as monitoring, and adds protection only where the customer approves it. The scope below is a proposed starting point; report cadence, alert handling, and response scope are agreed with each partner and each customer during scoping. The division of responsibilities is our intended operating model, confirmed in writing before it becomes a contractual commitment.

The customer-facing promise

“We found where AI can reach sensitive data, act with privilege, and operate outside the authority you defined. We watch that activity on the paths we cover, enforce the controls you approved, and bring you the evidence.”

What your account manager says. What the evidence report shows.

Monitoring · recurring · proposed
  • Review changes in the covered inventory and activity
  • Review the alerts you agreed to watch, and recommend actions
  • Deliver an evidence summary to the customer
  • Identify workflows suitable for further assessment or protection

Recommendations are reported separately from approved changes. Monitoring does not block or change customer systems.

Protection · once enabled · proposed
  • Tune the controls the customer approved
  • Handle held actions and exceptions under the contract
  • Include signed traces for holds and denials in the evidence summary
  • Review permitted actions with the customer, and agree any grant changes in writing

An agent's history in one customer's environment never authorizes work in another.

The MSP or MSSP

Owns the customer relationship and delivers the managed service: manages users, reviews alerts and findings, and handles customer-facing responsibilities within the agreed scope.

Behavry

Provides and operates the platform: provisions and configures the hosted environment, supplies access and licenses, maintains the infrastructure, and delivers the agreed technical support. Enforces the configured boundaries and supplies evidence for the supported paths.

The customer

Defines the authority it delegates, and keeps approval rights and evidence access, including through a change of provider.

Named operators

Act within that delegation, and approve or decline what falls outside it. Signed traces can be checked at verify.behavry.ai with no Behavry account.

Deployment · what to confirm in scoping
Customer boundaries
Each customer is its own tenant. Agents, policies, alerts, and audit events are isolated per tenant at the query level (Trust center). Your access to a customer is the access that customer agrees, and the customer keeps approval rights and evidence access.
Hosting
For MSP engagements, Behavry hosts and operates the platform. The current deployment architecture is published on the Transparency page. Hybrid, own-cloud, and self-hosted models can be discussed where a customer's requirements call for them.
Operation and support
Behavry operates, maintains, and updates the platform and gives you the agreed technical support. You support your customer. Escalation lines among the customer, you, and Behavry are set out in writing.
Isolation, recovery, capacity, availability
Agreed during scoping against the customer's requirements, and confirmed in writing. No availability commitment is published on this page.

Proposed packaging/Observation as a scoped engagement/Monitoring as a recurring service/Protection added with customer approval/Terms confirmed per partner

/007 · THE PILOT30 DAYS · 1 ENVIRONMENT · UP TO 3 WORKFLOWS

30-day governance pilot for an AI IT technician.

A scoped technical option for partners who want to go straight to enforcement — not a prerequisite for the observation engagement above. It governs an AI IT technician performing administrative actions, across one customer environment and up to three explicitly supported workflows, so you can measure labor impact, policy outcomes, and evidence before you commit to a contract. Agent and tool compatibility is scoped before anything is signed.

How it runs
  • Scope the environmentOne customer tenant via your existing partner access. Agent, ticket-source, RMM, and EDR compatibility confirmed in scoping. PSA connections such as ConnectWise, Autotask, or HaloPSA are proposed per pilot, not confirmed integrations.
  • Inventory, policy, observationGoverned actions are recorded and decided in observation first. Nothing is blocked yet.
  • Enforcement, by agreementAllow, escalate, and deny rules built with your service desk lead, switched on through a change process you agree.
  • Up to 3 explicitly supported workflowsChosen from the list, or your own runbooks once their paths are confirmed
  • Decision Trace evidenceSigned and hash-chained for each governed action, verifiable by the customer
  • A1 to A2 eligibility, in shadow modeOnly after the published demonstration gate passes. Permissions do not change.
What the closing report separates Measured labor impact. Observed policy outcomes. Unresolved issues. Coverage limits. Recommendations. A promotion recommendation in that report is not a production authorization.
Example pilot workflows · pick up to three
WorkflowSystemsExample starting policy
Password resetTicket → Entra IDALLOW within ticket scope, requester identity verified against the directory
Stale account disablementEntra ID, ticketALLOW after inactivity check. ESCALATE if the account holds a privileged role
M365 group or license changeM365 admin, Entra IDALLOW for the ticket's named user. ESCALATE on admin groups or license removal
Phishing ticket triageMailbox, EDR, ticketALLOW read and classify. ESCALATE purge and sender block
Endpoint quarantine recommendationRMM, EDRALLOW the recommendation. ESCALATE isolation to a human, every time
Low-risk remediation approval flowRMMALLOW from the approved runbook list. DENY anything not on it

Starting points, built with you in week one. Ticket and contract scope are enforced through mappings you configure; Behavry does not read prose contracts.

DISCUSS AN ENFORCEMENT PILOT →

Bring one customer and the workflows you want to govern. We will scope coverage, authority, evidence, and success criteria together.

/008 · THE PARTNER PROGRAMPROPOSED · PENDING COMMERCIAL APPROVAL

Your customer stays your customer.

These are the program requirements we are putting forward for commercial approval. They are proposals, not terms. We will confirm each one with you in writing before a pilot is contracted.

  • Registration and attributionPartner registration and attribution on pilot requests, with your consent before any customer contact.
  • Ownership rulesDefined rules for account ownership, renewals, expansion, and overlap with direct sales.
  • Separate chargesBehavry platform charges kept separate from your assessment and managed service fees.
  • EnablementDeployment guidance, a sample scope of work, and example evidence reports.
  • Clear support linesEscalation and support responsibilities set out among the customer, you, and Behavry.
  • Customer continuityCustomer-controlled approval and evidence access that survives a change of provider.

Co-branding, white-label exports, deal protection, console inclusion, partner discounts, and specific integrations are proposed until product and commercial owners confirm them. Nothing on this page sets a price or a term.

If removing Behavry means turning off your AI technicians, it is doing its job.

The control layer is what makes labor savings safe to keep. Take it out and the authority problem comes back.

DISCUSS AN MSP ENGAGEMENT →
/009 · DISCUSS AN MSP ENGAGEMENTPARTNER ROUTE · 2 BUSINESS DAYS

Discuss an MSP engagement.

Start wherever you are: AI observation for a customer, ongoing monitoring, or an enforcement pilot for an AI IT technician. Tell us what you want to cover and we will scope it together. You do not need to name the customer yet.

What are you interested in? *

Only an enforcement pilot asks for a candidate workflow.

Optional.

The engagement concerns *

Optional.

The administrative action you want an AI IT technician to perform under enforcement.

Optional. Leave blank if you would rather not name them yet.

Leave unticked and we will not contact them at all.

By submitting you agree we may use these details to reply, as described in our privacy policy.

RECEIVED. WE REPLY TO YOU WITHIN 2 BUSINESS DAYS. DISMISS
/ONWARDBEHAVRY · RECORD
  • COVERAGE

    Eight ways an agent gets in, and where each one is intercepted.

  • EARNED AUTONOMY

    How an agent's authority is reviewed against evidence, and why a recommendation is not an approval.

  • COMPLIANCE

    Control mappings and evidence support, produced as a by-product of enforcement.

  • DECISION TRACE

    The evidence format, with a sample package you can verify yourself.