AI Agents · Human-in-the-Loop

Agents Work the Case — and Ask a Person When They Should

Human-in-the-Loop is built into the Process Agent operating model. When evidence is incomplete, contradictory or unusual — or when policy requires a person before an external action — the agent pauses the case, explains its reasoning, gives a recommendation, and waits for accept or reject.

Use a clear Waiting for you state instead of forcing the agent to guess or silently fail

Require human sign-off before Action tools change SAP, send communications or update another external system

Escalate with the evidence, reasoning and recommended action — not just a generic question

Preserve case state while waiting and resume from the correct point after a decision

Keep the human response and subsequent agent steps in the same auditable case history

Human-in-the-loop by default · Recommendation included · Case state preserved · Full audit trail

Invoice Automation

receipt → matching → approval → post to ERP

waiting for you

✓Classified document · Invoicedone

✓Extracted fields & line itemsdone

✓Matched to PO 4500000538done

!Price variance above tolerancewaiting

Waiting for you

Line total is 3% over the PO. Approve the variance and post to SAP, or reject?

Approve & postReject

Pause safely

No forced guesses

Recommend

Reasoning + next action

Approve

Gate external actions

Resume

Continue same case

Where people belong

Put human judgment exactly where the process needs it.

The agent handles the normal path. Human checkpoints are for business decisions, unusual evidence, approval authority or missing information.

Exception review

Incomplete or contradictory evidence

The agent stops when the facts do not support a safe decision and surfaces the evidence behind the exception.

detect → explain → recommend → wait

Approval gate

Sensitive external action

Require sign-off before posting to SAP, sending an external communication or performing another Action tool.

prepare → request approval → accept/reject → execute

Correction

Missing or wrong business data

Ask a person to correct or provide a value instead of inventing information that should come from a document or system.

identify → request input → validate → resume

Business judgment

Policy says a person decides

Keep decisions such as high-value approvals, unusual variances or process exceptions under explicit human authority.

policy → escalate → decide → continue

Case state

Waiting is a first-class process state — not an error.

Process Agents have explicit statuses for work that is queued, running, waiting for a person, waiting on a system, resuming, completed, failed or cancelled. That makes human dependency visible and operationally manageable.

QueuedRunningWaiting for youWaiting on systemResumingCompletedFailedCancelled

Escalation anatomy

A good human checkpoint contains more than an approval button.

The reviewer should receive the context needed to make the decision and the agent should know exactly how to resume afterward.

01 · TRIGGER

Why the agent stopped

A policy, unusual evidence, missing value, threshold, contradiction or external action can create the checkpoint.

Policy requirementExceptionMissing evidenceAction approval

02 · EVIDENCE

What the agent found

Surface the document values, system data, matching result, discrepancy or other facts that caused the escalation.

Source evidenceERP dataVarianceValidation result

03 · REASONING

How it reached the checkpoint

Explain the conclusion reached so far so a reviewer can follow what the agent did before it stopped.

Conclusions loggedNo hidden jumpTraceable stepsCurrent state

04 · RECOMMENDATION

What it thinks should happen

Ask with a recommended next action rather than only presenting an unexplained exception.

ApproveRejectCorrectProvide missing input

05 · DECISION & RESUME

What the person decided

Capture accept/reject or correction, record it with the case, then resume the process from the appropriate point.

Decision recordedCase resumesAction gatedAudit trail

When it runs

The agent waits for you without losing the process.

A case can remain in Waiting for you until the authorized user responds. After the response, it enters Resuming and continues with the same evidence and prior conclusions available.

RunningWaiting for youDecision recordedResumingCompleted

Transparent

Every step logged

The agent records conclusions as it reaches them, so the reviewer can see what happened before the escalation.

Escalates well

Asks with a recommendation

When the evidence is incomplete or unusual, it provides reasoning and a recommended action and waits for accept or reject.

Conservative

Stops rather than invents

It does not invent values that should come from a document or system and does not assume an action succeeded simply because it tried.

Approval governance

Human approval is enforced by the runtime — not merely requested in the prompt.

The Process Agent architecture separates policy from instructions. Approval requirements for external actions can therefore remain enforced even if the prompt changes.

Human-in-the-loop

Approval on Action tools

Any tool that changes an external system can be configured to require human sign-off before it runs.

Policy

Policy over instruction

The approval requirement lives in governed policy, not only in natural-language instructions.

Audit

Decision audit trail

The human response remains associated with the same case and its subsequent execution.

Response target

Human-response target

Runtime policy can define how long a case should wait before escalation or operational follow-up.

Credentials

Authorized system access

The agent still uses isolated integration credentials; the reviewer approves the action, not the credentials.

Thresholds

Threshold-based escalation

Business limits such as tolerances or high-value thresholds can determine when a person must decide.

Human control near SAP

Let the agent prepare the SAP action — then require approval before the write.

An agent can read and validate SAP context, prepare a recommended action, and pause in Waiting for you. Only after an authorized person approves does the governed Action tool commit the change.

Explore SAP Automation→

Why this matters

Why Human-in-the-Loop matters for enterprise agents.

Keeps judgment with peopleAutomate normal work while preserving human authority over exceptions and sensitive decisions.

Prevents guessingThe agent stops when evidence is insufficient rather than inventing missing values.

Improves approval qualityReviewers see evidence, reasoning and a recommendation instead of a context-free request.

Works with process stateWaiting for you and Resuming are explicit states, so the work item remains operationally visible.

Gates external actionsSAP writes, email sends and other Action tools can require sign-off before execution.

Policy enforcedThe runtime holds the approval rule even if instructions are edited.

AuditableThe decision and subsequent action remain part of the same case history.

Designed for exceptionsVariances, missing data, unusual evidence and contradictions can be escalated cleanly.

FAQ

Common questions.

What does Human-in-the-Loop mean in Artificio?

It means a Process Agent can pause a case and ask a person to verify, approve, reject, correct or provide information before the process continues.

When does an agent ask a person?

It can ask when evidence is incomplete, contradictory or unusual, when required data is missing, or when runtime policy requires human approval before an Action tool runs.

What does the reviewer see?

The Process Agent model is designed to provide the evidence, the agent's reasoning and a recommended action rather than only presenting a generic approval request.

Does the case lose its state while waiting?

No. Waiting for you is an explicit case state. After the decision, the case can enter Resuming and continue from the appropriate point.

Can SAP writes require approval?

Yes. Tools that change external systems can be approval-gated before execution.

Is approval controlled by the prompt?

The stronger control is runtime policy. The Process Agent design explicitly separates governed policy from natural-language instructions.

Keep people where they matter

Bring an exception or approval your team still handles manually.

We'll show an agent doing the routine work, pausing with evidence and a recommendation, waiting for an authorized decision, and then resuming the same case safely.

Request a Demo →Talk to Sales