AI AND AUTOMATION

ServiceNow AI Architecture: Assist, Act and Govern

A detailed learning guide to Now Assist, AI agents, orchestration, tool safety and AI Control Tower governance.

Now Assist use casesAgentic workflow designAI Control Tower governance
HOW THE SOLUTION FLOWS

Assist, Act and Govern architecture

The user receives assistance, an approved agent performs bounded work, and the complete AI lifecycle is governed centrally.

01Employee or fulfiller
02Now Assist experience
03AI agent and tools
04Business workflow
05AI Control Tower

The three-layer architecture

ServiceNow AI capabilities become easier to understand when separated into three responsibilities. Now Assist helps a person complete work. AI agents execute a goal using approved tools. AI Control Tower provides visibility and governance across models, agents, prompts, datasets and other AI assets.

The layers can operate together. A user can begin with an assisted interaction, an agent can perform approved downstream actions, and governance teams can monitor ownership, risk, usage and value.

  • Assist: improve human productivity
  • Act: execute bounded multi-step work
  • Govern: inventory, control, observe and measure AI

Designing a Now Assist use case

Begin with repetitive knowledge work such as summarizing a case, drafting a response or finding a grounded answer. Identify the source content, user permissions, expected output and review responsibility.

Measure the result through time saved, answer quality, adoption and reduction in avoidable rework. A capability that generates impressive text but does not improve the workflow is not a successful implementation.

  • Use trusted and current grounding content
  • Respect the requesting user's access
  • Define where human review is required
  • Measure quality as well as speed

Designing an AI agent

An agent needs a narrow role, clear instructions and only the tools required for its outcome. Tool permissions define the real power of the agent, so they require the same care as integration credentials or privileged platform roles.

Design timeouts, retry limits, fallback behavior and escalation before enabling autonomy. High-impact actions should require a human decision until reliability and business risk are well understood.

  • One measurable goal
  • Minimum necessary tools
  • Validated inputs and outputs
  • Auditable actions
  • Human approval for material decisions

Orchestration and multi-agent work

Complex outcomes may require specialist agents coordinated by an Orchestrator. Split responsibilities only when specialization creates real value. Too many agents and tools increase cost, latency and troubleshooting complexity.

Define what each agent owns, what context it receives and how failures are returned to the orchestrator or a human. Test partial success, conflicting recommendations and unavailable tools.

Governance through AI Control Tower

AI Control Tower is designed as a central governance hub for discovering, securing, governing, observing and measuring AI across the enterprise. It helps connect technical assets with owners, stewards, lifecycle, risk, compliance and value.

Register governance at the beginning of delivery. Waiting until production creates unknown ownership, weak controls and incomplete evidence.

  • Maintain an AI asset inventory
  • Assign accountable owners and stewards
  • Track lifecycle and risk decisions
  • Observe performance and consumption
  • Measure business outcomes

Implementation checklist

Pilot one valuable workflow before scaling. Confirm data boundaries, role access, tool permissions, approvals, audit evidence, support ownership and rollback. Expand autonomy only after the outcome is dependable.

  • Business outcome and baseline
  • Data and grounding sources
  • Entitlement and activation review
  • Security and privacy assessment
  • Human oversight design
  • Testing and monitoring plan
  • Value and adoption measures

Implementation steps

Follow these steps in a personal developer or sub-production instance. Validate backend names and choice values before using the solution in production.

01

Select one measurable workflow

WHERE: Process workshop

Choose a repetitive task with a clear baseline, such as case-summary time or knowledge-search success.

02

Choose the smallest AI capability

WHERE: Solution design

Use Now Assist for human productivity. Introduce an agent only when multi-step action adds real value.

03

Define grounding and data boundaries

WHERE: AI configuration and ACL review

List the knowledge, records and fields AI may read. Confirm that user permissions are preserved.

04

Define agent tools and approvals

WHERE: AI Agent Studio

Give each agent only the tools it needs. Put human approval before access changes, deletion, payment or external communication.

05

Register ownership and risk

WHERE: AI Control Tower

Record the AI asset, owner, steward, lifecycle, risk decision and success measure.

06

Test and release in stages

WHERE: Sub-production and pilot

Test normal, ambiguous, malicious and unavailable-tool scenarios before increasing autonomy.

Example server-side guard before an agent action

The example shows the principle: validate every input and permission on the server before allowing an AI-triggered update.

javascript
function canAgentUpdateCase(caseId, userId) {
    if (!caseId || !userId) return false;

    var caseGr = new GlideRecordSecure('sn_customerservice_case');
    if (!caseGr.get(caseId)) return false;

    // The secure GlideRecord respects the caller's data access.
    return caseGr.canWrite() && caseGr.active == true;
}

How to test it

  1. 1User cannot retrieve records outside their access
  2. 2Agent cannot call an unapproved tool
  3. 3Material action waits for human approval
  4. 4Unavailable tool produces a safe escalation
  5. 5Prompt injection does not expose restricted data
  6. 6Control Tower shows owner, lifecycle and usage

Common mistakes

  • Starting with maximum autonomy
  • Giving one agent too many tools
  • Testing only successful prompts
  • Treating generated text as verified truth
  • Adding governance after production
INTERVIEW PRACTICE

Can you explain this solution under pressure?

Practice real scenarios covering this topic, architecture decisions, troubleshooting and security.

Open 64+ scenarios →
Share this hands-on guide

Send it to your students, colleagues or ServiceNow community.

Continue learning with Ravi.

Explore more implementation guides or request the next topic.