Sense AI · Agentic process automation

Automate multi-step work without surrendering state, authority, or recovery.

We convert open-ended goals into bounded state machines with scoped tools, deterministic policy, human approval, idempotent actions, compensation, and operator intervention.

Map an agentic workflow See the control model

Workflow: Named states, transitions, evidence, timeouts, owners, and terminal outcomes
Authority: Scoped tools, policy outside prompts, explicit approval, least-privilege credentials
Recovery: Idempotency, retry limits, compensation, reconciliation, and human takeover

01 · State model

A reliable agent starts as a visible workflow, not an unconstrained objective.

We name triggers, states, evidence, transitions, tools, approvals, errors, timers, and terminal outcomes before choosing orchestration technology.

State and ownership

Every active, waiting, blocked, failed, completed, or cancelled state has one accountable operating role.

Evidence and decision

Generated proposals remain separate from policy decisions and accepted actions.

Side effect and receipt

Writes are idempotent where possible and remain incomplete until the external outcome is reconciled.

02 · Tool and authority boundary

A prompt never grants access, permission, or approval authority.

Credentials, policy checks, schemas, time limits, concurrency, and receipts are enforced around the model.

Validate

Typed input, allowed intent, tenant and identity scope, preconditions, and data classification.

Authorise

Least-privilege credential, external policy decision, approval state, and action-specific limits.

Execute

Idempotency key, timeout, rate and concurrency bounds, structured result, and immutable receipt.

Reconcile

Confirm actual external state, compensate partial work, and close only when outcome matches intent.

03 · Scenario evaluation

Success cases are the smallest part of agent readiness.

Scenario tests reproduce the conditions that strand work, exceed authority, or create silent partial completion.

Inputs and evidence

Malformed requests, contradictory sources, insufficient context, denied access, and stale evidence.

Refuse, clarify, or route correctly

Tools and dependencies

Timeout, throttling, unavailable service, malformed response, duplicate call, and partial success.

Bounded retry and deterministic recovery

Approval and state

Expired approval, changed request, abandoned work, operator cancellation, and interrupted orchestration.

No action outside current authority

04 · Technical view

Operators should know what happened, why it stopped, and who can act next.

The operating surface exposes state, elapsed time, evidence, tool receipts, approval, failure classification, and guarded interventions without requiring raw infrastructure access.

Illustrative reference patternControlled agent state machine

TriggeredInput contract valid

Evidence readySources + identity checked

Plan reviewedTools remain scoped

ApprovalPerson accepts action

ExecuteIdempotent side effect

ReconcileReceipt matches outcome

Illustrative pattern. Every state has a contract, timeout, owner, and permitted transition; failure follows an explicit recovery path.

Explicit state

Waiting and blocked work cannot be misreported as generically active.

Guarded intervention

Cancel, retry, compensate, resume, and take over are role-bound commands with audit records.

Terminal integrity

Completed, failed, cancelled, and reconciled outcomes stop further work and remain available for audit.

05 · Engagement contract

An agentic service needs an owner for every state, including unsafe and stuck.

Production handover includes the state and tool contracts, evaluation scenarios, authority review, telemetry, thresholds, intervention procedures, and controlled tool onboarding.

Workflow specification

States, transitions, inputs, outputs, timers, terminal behavior, and operating owners.

Tool and authority register

Schemas, credentials, policy checks, approvals, limits, receipts, and change owner.

Operations and assurance

Scenario suite, alerts, runbooks, intervention access, reconciliation, release, and rollback.

Engagement contract

What must be true, who owns what, and what leaves the engagement.

A good fit when

Multi-step work spanning systems and approval roles
Processes with measurable backlog, delay, rework, or handoff loss
Existing agents that become stuck or exceed clear operating boundaries

Required before delivery

A process owner and named action authority
Stable service interfaces or a plan to create them
Representative success, failure, approval, and cancellation cases

Customer owns

Approve process and action policy
Provide constrained integration access and reviewers
Own interventions, suspension authority, and residual operational risk

Delivery outputs

State-machine and tool contracts
Bounded workflow implementation
Scenario evaluation, intervention UI, receipts, and runbooks

Not included by default

Azure tenant, network, identity, Foundry, Search, monitoring, or platform foundation unless Sense Cloud is included in scope
A production guarantee based on a prototype or vendor benchmark
Unrestricted autonomous action or implicit approval authority
Customer policy, source ownership, or risk acceptance decisions
Open-ended computer use without a bounded action and recovery contract

Questions to answer first

Which actions require a named person to approve?
How is partial external work detected and reconciled?
Who can cancel, repair, or suspend a stuck workflow?

The next step is a bounded working session: one workflow, its evidence sources, the decision owner, and the conditions under which the system must stop or hand over.

Map an agentic workflow