Approval
A configured action exceeds Claire’s authority and requires an accountable owner to approve or reject it.
Home / Human in the loop
Human control by design
AI should know when the work belongs to a person.
Claire handles repeatable coordination without pretending every decision should be automated. Approvals, judgment, empathy, expertise, accountability, and exceptions remain with the people responsible for them.
Claire preparesGather context, retrieve records, complete routine checks, and
identify the decision required.
A person decidesReview the exception or approval with the relevant context
already assembled.
Claire continuesRecord the decision, complete permitted follow-up, and keep the
requester informed.
Escalation model
A handoff should move work forward.
Weak automation sends a transcript or creates a generic ticket. A useful handoff explains what happened, what has already been checked, why the workflow stopped, what decision is needed, and what happens after the person responds.
Exception
The request falls outside a normal policy path, contains conflicting information, or needs a discretionary accommodation.
Judgment or empathy
The situation requires professional expertise, negotiation, reassurance, clinical or legal judgment, or relationship management.
What people receive
Context, not another administrative burden.
The human should enter at the decision point—not restart the workflow from the beginning.
- Request summaryWhat the person needs, including urgency and relevant history.
- Work already completedChecks, records, documents, and system actions already performed.
- Reason for escalationThe precise rule, ambiguity, exception, or approval threshold that stopped autonomous execution.
- Required next decisionA clear question or action for the accountable person.
- Continuation pathWhat Claire may complete after the human responds.
Operational examples
Human control changes by workflow.
Healthcare access
Claire can collect scheduling and coverage context. Access staff handle clinical concerns, policy exceptions, sensitive complaints, and cases requiring judgment.
Legal intake
Claire can gather matter information and support conflicts workflows. Attorneys and authorized staff decide representation, legal advice, and professional obligations.
Financial servicing
Claire can collect documents and apply configured servicing checks. Authorized personnel handle exceptions, regulated decisions, and required approvals.
Five control patterns
Human involvement is a workflow state—not a vague promise.
A serious implementation defines what triggers human involvement, which context travels with the request, who owns the decision, how long the workflow waits, and what Claire is allowed to do afterward.
Exception escalation
Claire handles the normal path until required data is missing, a record cannot be matched, an external action fails, a threshold is crossed, or a policy rule says a person must take over.
Context-rich handoff
The person receives the requester, collected facts, current workflow state, rules evaluated, actions attempted, failure reason and the specific decision or action needed.
Human takeover
A person can own the live interaction when empathy, negotiation, sensitive judgment or professional responsibility matters more than automated continuity.
Decision capture
The outcome of the human step must become usable workflow state: approved, declined, changed, needs more information, assigned elsewhere or closed.
Return to Claire
After the human action, Claire can continue only the permitted follow-up: update a record, send an approved confirmation, wait for another event or close the workflow.
Timeout and fallback
A handoff is not complete merely because a notification was sent. Define acknowledgement, wait intervals, escalation ladders and the truthful message given when no person responds.
Design the boundary
Decide where Claire acts and where your people step in.
The boundary should reflect consequence, ambiguity, policy and customer expectation—not a blanket percentage of automation.
Questions to answer before launch
- Which actions may Claire perform without review?
- Which facts or confidence conditions require escalation?
- Who owns each exception during and after business hours?
- What information must accompany the handoff?
- What counts as human acknowledgement?
- What may resume after the decision?
- What happens when the person or system never responds?
Operational examples
The right human checkpoint depends on the work.
HVAC emergency intake
Claire can collect the customer, address, symptom and configured urgency indicators. An on-call dispatcher owns safety exceptions, promises outside policy and technician reassignment.
Freight service risk
Claire can capture load, stop, ETA and delay context, update permitted status fields and notify routine stakeholders. A dispatcher owns missed appointments, recovery decisions and commercial exceptions.
Property maintenance
Claire can identify the resident, unit, issue, access constraints and approved vendor path. A manager owns ambiguous emergencies, unapproved spend and resident-sensitive decisions.
Restoration loss intake
Claire can structure the loss and alert the correct branch. A response leader owns safety judgment, crew commitment, program exceptions and arrival promises.
Insurance FNOL
Claire can assemble the reporting party, policy context, event details and documents. Licensed or authorized people retain coverage, liability, severity and settlement decisions.
Legal client intake
Claire can gather parties, dates, matter facts and documents. Lawyers and authorized staff retain conflicts, advice, merit and representation decisions.
Failure-path acceptance test
Test the moments the polished demo avoids.
Before launch, test a missing identifier, conflicting records, an unavailable API, an interrupted conversation, a caller who asks for a person, an after-hours exception, a declined human decision and a handoff that nobody acknowledges. The result should show what Claire knew, what rule fired, what action was attempted, who owned the exception and what the requester was told.
Never fake completion
If a system write or human decision did not happen, the workflow must preserve that state and avoid sending a success message.
Never lose the thread
The person should not need the requester to repeat everything Claire already collected simply because ownership changed.
Define where AI stops.
Bring us a workflow and we will map its authority boundaries, exception paths, and human ownership.