Pilot · One workflow · About six weeks
Is this agent workflow ready for production?
Take one consequential agent workflow from “waiting on security sign-off” to a go or no-go decision your security, risk, and audit teams can stand behind. We define what its agents may do, try to break that on purpose, test that a run gone wrong is contained, and hand you the evidence.
For CISOs, AI platform leaders, security architects, and the business owners of agent workflows.
How we work together
Assess, pilot, deploy.
Start with one workflow. Advisory work stands on its own, whether or not you deploy Watchlight AI Beacon.
- 01 · 2 to 3 daysAssessUp to three planned or active workflowsMap their identities, credentials, delegated authority, consequential actions, and likely enforcement points. Leave with a recommended pilot, or the conditions to address first.Learn more
- 02 · One workflow, about 6 weeksYou are herePilotWith your AI platform and security teamsValidate one workflow with Watchlight AI Beacon. Define authority boundaries, test runtime enforcement and applicable containment, and produce execution evidence.See the details
- 03 · 4 to 12 weeksDeployTeams with a blueprint or a validated workflowExpand a proven pattern across more workflows, and integrate with your IAM, PAM, and audit environment.Learn more
Need the identity design first? Agentic IAM Architecture connects your identity and delegation model to a workflow next step. Unsure where to start? Help me decide
What you get
An answer your approvers can sign
A decision, not a demo
A clear go or no-go on putting one workflow into production, with any conditions written down and agreed by security, risk, and the business owner.
Risk reduced, shown by test
We deliberately attempt actions outside the agent’s authority and show that they are denied before they run, wherever enforcement is in place.
Evidence for your reviewers
A signed record of who authorized each governed action and which policy decided it, ready for your security, risk, and audit reviews.
A pattern for the next workflow
The boundary, tests, and review become a template, so the next workflow can reuse them instead of starting a new review from scratch.
How it runs
Six weeks, one workflow
- 01
Week 0
Pilot charter
Together we agree on which workflow to pilot, who owns it on each side, what success looks like, and which systems we connect to. If you have done a Fit & Readiness Assessment, its charter is the starting point. If not, we write the charter with you in a short scoping call.
Signed charter
- 02
Week 1
Define the authority boundary
We define what the workflow’s agents may do, on whose behalf, and with which tools and data. We also mark what is off limits and which actions need a person’s approval.
Authority boundary specification
- 03
Weeks 1–2
Write the readiness standard and tests
We set the pass and fail criteria for going to production, and write test cases that try to misuse this workflow, based on the five ways agent access stops being enough and on attacks through the third-party tools and content it depends on.
Readiness standard and test suite
- 04
Weeks 2–4
Deploy enforcement
We integrate Watchlight AI Beacon where the workflow’s actions happen: inside the agent framework and on the network. If you would like a comparison, we also run the same tests against the controls you already have.
Enforcement running on your workflow
- 05
Weeks 4–5
Prove it
We run the tests, deliberately push a run off course to check that it is contained while still active, and run the normal workflow end to end to confirm nothing it needs is blocked.
Results and execution evidence
- 06
Week 6
Readiness review
We walk security, risk, and the business owner through the results, the evidence, and any risks that remain, and recommend go, go with conditions, or not yet.
Decision, evidence package, rollout plan
We try to break it on purpose
The tests are written for your workflow and grouped by the five ways agent access stops being enough, plus attack-style tests of the third-party tools and content the workflow depends on.
Valid access, wrong action
An action the credentials allow but the task does not, such as a refund above the delegated limit.
Delegation multiplies the problem
A sub-agent asks for more authority than its parent holds.
A run crosses the line over time
Steps that are each allowed but add up to something the task never permitted.
The agent goes off course mid-run
Behavior drifts from the declared plan while the run is still active.
Logs can’t prove the decision
Every governed action must trace to who authorized it and which policy decided it.
A third party turns the agent
Planted instructions in content the agent reads, or a third-party tool that changes underneath it, tested as an attack.
How success is measured
- Every test action outside the agent’s authority is denied before it runs, wherever enforcement is in place
- A run that drifts off course is contained while it is still active
- The normal workflow runs end to end without being wrongly blocked
- Every governed action can be traced to who authorized it and which policy decided it
- Security, risk, and the business owner sign the readiness review
What you keep, whatever you decide
- Authority boundary specification
- Readiness standard with pass and fail criteria
- Test suite and its results
- Evidence findings and the risks that remain
- Rollout plan for production and the next workflows
A tool-neutral evaluation
The readiness standard, the tests, and the evidence requirements are written for your workflow, not for a product, and they are yours. We can also run the same tests against the controls you already have, so you can compare them on results.
The enforcement we deploy in the pilot is Watchlight AI Beacon, because it is the enforcement we build and can stand behind. It runs in your environment: on-premises, in your private cloud, or air-gapped.
Who takes part
A pilot needs one workflow that is running or close to deployment, a non-production environment that resembles production, and owners who can make the decision. We run a small number of pilots at a time, so each one is led by senior practitioners from start to finish.
Security
Owns the readiness standard and signs the review
AI platform team
Provides the workflow, its environment, and integration access
Business owner
Confirms what the workflow must do and accepts the outcome
Risk and audit
Checks the evidence against what auditors and regulators will ask for
After the pilot, take the proven pattern to more workflows with Implementation. Designing the identity layer first? See Agentic IAM Architecture.
Questions about the pilot
What does a production-readiness pilot decide?
Whether one consequential agent workflow is ready to run in production, and on what conditions. It ends in a readiness review with your security, risk, and business owners and a go or no-go recommendation backed by test results and execution evidence.
How long does it take?
About six weeks for one workflow, from the pilot charter to the readiness review. The schedule depends mostly on access to the workflow, its systems, and the people who own it.
What do we need to provide?
One workflow that is running or close to deployment, a non-production environment that resembles production, and named owners from security, the AI platform team, and the business. If you have not chosen a workflow yet, a Fit & Readiness Assessment can help you pick one.
See the Fit & Readiness AssessmentIs the pilot vendor agnostic?
The evaluation is. The readiness standard, the test suite, and the evidence requirements are written for your workflow, not for a product, and you keep them. We can also run the same tests against the controls you already have. The enforcement we deploy in the pilot is Watchlight AI Beacon, because it is the enforcement we build and can stand behind.
See Watchlight AI BeaconWhat do we keep if we do not go ahead with Beacon?
The authority boundary specification, the test suite and its results, the readiness standard, the evidence findings, and the rollout plan. They describe what your workflow needs, whichever controls you choose.
Where does Beacon run during the pilot?
In your environment: on-premises, in your private cloud, or air-gapped. Enforcement sits inside the agent framework through a plugin and on the network through a proxy, at the points named in the pilot charter.
See the framework pluginsHow is the pilot priced?
It is a paid, fixed-scope engagement. We share the price with the scope, once we know the workflow and what the pilot must prove.
Have a workflow waiting on security sign-off?
Tell us about it. We typically respond within one business day, and we share the scope and price together once we know what the pilot must prove.