Skip to main content
Request Demo

AI Agents · Agent Actions

Give Agents Tools to Act Without Giving Them Uncontrolled Access

Agent Actions are the tools that let a Process Agent change the outside world: write to SAP, call an API, send an email or SMS, generate a document, trigger a workflow, or use an approved website when no suitable API exists. Read tools observe. Action tools change systems, and can require human approval before they run.

Request a DemoSee how actions are governed
  • Read vs Action tools
  • Approval-gated
  • Credentials isolated
  • Result confirmed
ACTION TOOL · SAP SUPPLIER INVOICEready
Post Supplier Invoice

external system change · approval gated

Action
  1. Done: Read current SAP PO / GR statefresh
  2. Done: Validate case against policypass
  3. Done: Human approval recordedapproved
  4. Running: Post supplier invoice to SAPexecuting
Action safeguard

This tool changes SAP. Confirm the approved case should be posted using the current live data.

Execute actionCancel
  • Separate Read tools that observe from Action tools that change external systems
  • Gate sensitive actions with runtime-enforced human approval before execution
  • Keep SAP and other system credentials inside platform integration accounts, never in prompts or case data
  • Require a fresh live read before a write when policy says cached data is not sufficient
  • Log the action attempt and confirm the external result instead of assuming success because a tool was called
Read toolsObserve safely
Action toolsChange external systems
ApprovalGate sensitive writes
ConfirmNever assume success

Action anatomy

A governed action has a clear path from evidence to confirmed result.

The process should know what it is about to change, whether the data is current, whether approval is required, and what the external system actually returned.

  • 01 · TOOL ASSIGNMENT

    What action is available

    The agent configuration determines which Action tools exist for the process. If a tool is not assigned, the agent cannot choose it.SAP writeEmail sendAPI actionWeb Agent
  • 02 · PRECONDITIONS

    What must be true first

    Rules, required data, policy, permissions and current system state should be satisfied before the action is eligible to run.Validation passFresh dataPolicy allowRequired fields
  • 03 · HUMAN APPROVAL

    Whether a person must sign off

    Runtime policy can require approval before an Action tool executes, keeping sensitive commitments under human control.ApproveRejectWaiting for youAuthorized user
  • 04 · EXECUTION

    The external system change

    The tool uses the configured integration account or controlled browser session; credentials remain isolated from prompt and case context.Integration accountCredential isolationGoverned callWeb session
  • 05 · CONFIRMATION

    What actually happened

    The agent should record the system response and confirm success or failure rather than assuming success simply because it attempted the action.Response capturedSuccess confirmedFailure routedAudit trail

What agents can do

Turn reasoning into controlled business execution.

Tools are explicit capabilities. The agent only sees and uses the tools assigned to its configuration, and Action tools can be guarded more tightly than Read tools.

  • SAP & ERP

    Create or update enterprise transactions

    Use approved integration tools to read current data and perform governed writes to SAP or other systems of record.live read → validate → approve → write
  • Communication

    Send email or SMS

    Use dynamic process data to notify, request approval, send documents, communicate exceptions or update a customer or supplier.prepare → approve if required → send → record
  • Workflow & documents

    Trigger process actions and generate outputs

    Start downstream workflows, create PDFs or forms, schedule work or move a case to the next operational step.generate → trigger → observe result → continue
  • Web Agent

    Act through approved web interfaces

    When no suitable API exists, use controlled browser automation to navigate, enter validated data or download permitted documents, with logging and human checkpoints.login → navigate → act → capture result
Tool model

Read tools observe. Action tools change external systems.

This distinction lets an agent gather evidence freely within its permissions while applying stronger controls to operations that create commitments or change systems of record.

Read & lookupClassify & extractAction toolApproval gateExternal result

At execution time

Fresh data, approval and confirmation make external actions safer.

The runtime can require a live read immediately before a write, then gate the Action tool with a person and capture the actual external response. Failed or ambiguous outcomes remain visible in the case.

RunningFresh read completeWaiting for youAction executingCompletedFailed
  • Permissioned

    Only assigned tools exist

    Each agent gets only the tool set needed for its process rather than unrestricted access to every platform capability.
  • Approval-gated

    External changes can wait

    Writing to SAP, sending email or other Action tools can require human sign-off before execution.
  • Confirmed

    Do not assume success

    The agent verifies the system response and records whether the external change actually succeeded.

Action governance

The runtime controls the action even when the agent chose it.

Process Agent governance is strongest around external changes: approval can be required, credentials remain isolated, data freshness can be enforced and the result is logged with the case.

  • Human-in-the-loop

    Human approval on actions

    Sensitive Action tools can require sign-off before they execute.
  • Credentials

    Credentials isolated

    SAP and other login details remain in integration accounts, never in prompts or case data.
  • Freshness

    Live read before write

    Policy can require current affected records immediately before a commitment is made.
  • Audit

    Action history logged

    The attempted action and external response stay associated with the work item.
  • Policy

    Policy decides autonomy

    The runtime decides whether an action may run autonomously or must wait for a person.
  • Limits

    Retry limits

    Tool retry caps prevent an agent from repeating a failing external action indefinitely.

Automating something that isn't on this page? The agents adapt to new document types, new systems and new rules. Tell us what you are automating and we will show you the nearest thing we run.

Talk to our team
Actions designed for SAP shops

Read SAP live, prepare the transaction, get approval, then write back safely.

Process Agents can combine cached feeds with live SAP reads and governed Action tools. Credentials stay isolated, policy can require a live read immediately before commit, and human approval can gate the final write.

Explore SAP Automation

Why this matters

Why governed Agent Actions matter.

  • Moves beyond copilot suggestionsThe agent can complete approved work instead of only telling a person what to do next.
  • Explicit tool permissionsThe agent only has the actions assigned to its configuration.
  • Read and write are separatedObserving data does not automatically imply permission to change it.
  • Human control on commitmentsSensitive external changes can wait for an authorized decision.
  • Credentials never enter promptsIntegration accounts isolate system authentication from agent reasoning.
  • Current data before commitFreshness policy can force a live read before a write.
  • Web fallback when APIs are absentApproved browser automation extends execution to portals that do not expose suitable APIs.
  • Every outcome is traceableThe system response, success or failure state and follow-up stay with the case.

FAQ

Common questions.

What is an Action tool?
An Action tool is a capability that changes an external system or causes an external effect, such as writing to SAP, sending an email or executing an approved browser action.
How is an Action tool different from a Read tool?
Read tools observe or retrieve information. Action tools change an external system and can therefore be governed with stronger approval and policy controls.
Can an Action tool require human approval?
Yes. The Process Agent model explicitly allows tools that change external systems to require human sign-off before they run.
How are SAP credentials protected?
Credentials stay in the platform's integration accounts. They are never shown to the agent, written into a prompt or stored with a case.
Why perform a live read before a write?
Cached data is useful for speed, but policy can require current data immediately before a commitment so the agent acts on the latest state of the affected records.
What if the external action fails?
The agent should not assume success simply because it called the tool. The external response is captured so the case can record success, failure, retry or human follow-up.
What if there is no API?
A governed Web Agent can use an approved browser-based process to navigate, enter validated data or download permitted documents, with execution logging and human checkpoints where required.

Security & compliance

Enterprise governance by design

Agent Actions operate under runtime-enforced policy with permissioned tools, approval gates, isolated credentials, data-freshness controls, retry limits and a traceable record of the external system result.

  • HUMAN-IN-THE-LOOP
  • CREDENTIALS ISOLATED
  • FULL AUDIT TRAIL

Turn decisions into execution

Bring one process where people still copy, post, send or click through systems manually.

We'll show an agent gathering the evidence, passing the controls, waiting for approval where required, and then executing the approved SAP, API, communication or web action with the result logged.

Request a demoTalk to sales