When should an AI chatbot hand over to a human?
Set practical human handover triggers for an AI customer service bot, with a reusable triage checklist, escalation ownership and a simpler non-AI option.
Stanic & Co.
Perspectives for business leaders
An Australian SME should add a human bypass when a chatbot encounters sensitive information, a consequential decision, a disputed answer, a policy exception or a customer asking for a person. It should also escalate when it cannot verify an answer against approved information. Do not treat a model’s confidence score as proof of accuracy. Make human review available as a service policy, without assuming it is a universal legal entitlement.
The practical decision is not simply whether to put someone “in the loop”. It is which events stop automation, what reaches the reviewer and who remains responsible until the customer receives a response. The following is an illustrative operating procedure for that handover, not a legal standard or a guarantee of correct decisions.
When should an AI chatbot hand over to a human?
Start with a boundary: the bot may explain approved information, but it should not settle a consequential dispute simply because it can produce an answer. For example, explaining where to find an invoice differs from deciding that a disputed charge is valid. Route the second task to someone with access to the evidence and authority to correct the account.
There is evidence for making that route visible. In the OAIC’s 2026 community attitudes survey, 81% of Australian adults identified the right to human review as an essential condition before organisations use AI to make decisions affecting them. That is a statement of public expectations, not evidence that every chatbot interaction carries a statutory right to review.
The OAIC’s commercial AI guidance also calls for consideration of human oversight, product suitability and accuracy. It describes AI use in decisions with legal or similarly significant effects on individual rights as likely to be a high privacy risk activity. Ask a qualified adviser how privacy obligations apply to your organisation and proposed workflow.
Keep this decision separate from the initial choice of which services to automate. If that scope is still unclear, first use the guide to which customer queries are safe to automate. Here, the question is narrower: once a conversation starts, what makes the bot stop answering and hand responsibility to someone else?
Reusable handover checklist: stop on any failed condition
Use the following as an illustrative release checklist. For each conversation, record pass, fail or unknown against each condition. A fail or unknown means stop substantive automated answering and create a human-owned case. The bot may still acknowledge receipt and explain the next step. These are proposed controls, not rules prescribed by the sources.
Condition 1 — Information boundary: PASS only if the request stays within the approved data scope. FAIL if it contains health information, identity documents, payment credentials or other information your business has excluded. Route through an approved secure process; do not invite the customer to paste more details. Health information may be legally sensitive information, while financial and identity information should not automatically be labelled the same way. Have an adviser confirm the relevant classifications.
Condition 2 — Decision boundary: PASS only if the bot is explaining approved information without deciding a dispute or exception. FAIL for contested bills, requests to override policy, disputed eligibility, complaints about an earlier answer or other consequences outside its authority. Condition 3 — Evidence boundary: PASS only if the answer is supported by a current, approved source that covers the circumstances. FAIL if the source is missing, contradictory, stale or silent on the requested exception.
Condition 4 — Customer choice: PASS only while the customer has not requested a person or challenged the answer. FAIL immediately on either event; do not require another attempt at self-service. Condition 5 — Operational readiness: PASS only if the handover destination is working, an owner is assigned and the customer can receive an honest response expectation. FAIL if any element is absent; show the fallback contact route instead.
Example entries: “Opening hours; approved current page; no personal details; no human request” passes all five conditions. “My promotional credit is missing; please override the bill” fails the decision boundary. “Can I upload my medical report?” fails the information boundary. Record the trigger, case identifier, destination, acceptance status and next update due. Do not count a handover as complete merely because the bot says it has transferred the conversation.
Why a 90% confidence score is not enough
A proposed rule such as “escalate below 90% confidence” needs a definition before it becomes useful. What produces the score? Is it the model’s own assertion, a document-matching score or an estimate tested against labelled examples? Without evidence connecting that score to correct answers for your particular service, it should not authorise automated handling.
The OAIC explains that generative AI is probabilistic and can produce inaccurate or false results. Its accuracy guidance calls for reasonable steps proportionate to the risks when handling personal information. It does not prescribe a 90% confidence threshold. A confidently expressed answer can still fail your source or authority checks.
Prefer observable release conditions: an approved source exists, the answer matches it, the question is within scope and no escalation trigger applies. If your supplier provides a confidence score, treat it as an additional signal until evaluation demonstrates what it means. A high score must never override a customer’s request for a person, an information restriction or a policy-exception trigger.
Before launch, assemble a labelled test set containing routine questions, disputes, requests for people, ambiguous wording and missing-source cases. Record whether each should answer or escalate, then compare the bot’s behaviour with those labels. For an illustrative pilot gate, require every critical escalation case in that test set to route correctly. Passing a finite test is not proof that all future conversations will be safe; use the pilot accuracy audit guide to structure broader evaluation.
Give the handover an owner, not just a queue
Make the customer service manager accountable for the escalation policy, staffing and response expectations. Give the technical owner responsibility for routing, failure alerts and disabling automated answers. Assign privacy or legal questions to an appropriately qualified adviser. Each live case also needs an identified team or person responsible for its next action.
The handover package should contain only what the reviewer needs: the customer’s stated request, relevant conversation excerpts, the source used, the reason for escalation and any action already attempted. Separate customer statements from AI-generated summaries. For important cases, the reviewer should inspect the original exchange rather than assume the summary preserved every qualification.
Set customer expectations accurately. Say “Your request has been submitted for staff review” only after the case system confirms receipt. Do not say “You are speaking with a person” while the customer is still waiting. The OAIC recommends clearly identifying public-facing AI tools, alongside transparent information about AI use in privacy policies and notifications.
After hours, offer a case submission or another genuine contact route rather than implying immediate attendance. Choose response targets according to staffing and the consequences of delay. If capacity is exhausted, reduce the bot’s scope or switch to the fallback rather than quietly suppressing escalation. The trade-off is explicit: a bypass can increase staff work, but preventing handover to improve containment figures defeats its purpose.
Hypothetical failure: a disputed promotional charge
Consider a hypothetical retailer whose bot answers billing questions using a standard policy page. A customer says a one-off promotion promised a credit that is missing from their bill. The bot retrieves the ordinary exclusion policy, denies the adjustment and closes the chat. The promotion exists in a separate record the bot cannot access.
The failure is not merely an inaccurate sentence. The bot has treated incomplete evidence as authority to settle a dispute, then removed the customer’s immediate route to challenge it. The customer could complain about both the billing outcome and inaccurate information recorded about them. This example illustrates a possible failure; it is not a reported incident or a prediction of a legal finding.
Under the checklist, the first mention of a disputed charge or exceptional promotion triggers handover. The bot acknowledges the issue without accepting or rejecting the credit, creates a case and identifies the relevant conversation. A staff member checks the promotional evidence and account history, then decides within their authority or refers it further.
To detect this failure after launch, review sampled denials and conversations where customers repeat a dispute, challenge an answer or seek another contact channel. Check whether a case was actually created. If an automated denial escaped the boundary, pause that intent, preserve the relevant evidence under your retention policy, correct the account and associated records where appropriate, and contact affected customers through the approved process.
Restoring the feature should require a new test covering the missed promotion, paraphrased disputes and routing failure. Do not simply add the exact complaint wording to a trigger list and assume the underlying problem is fixed. Review whether the bot was allowed to decide something that should always have belonged to staff.
Human review has limits: test the whole service
A human bypass is not a guarantee of correctness. A reviewer can lack the same promotional information, misunderstand the policy or accept the bot’s summary without checking it. If staff cannot access the necessary evidence or change an outcome, the handover creates delay without meaningful review.
Define the reviewer’s authority and evidence requirements. For a disputed bill, that might mean checking the original offer, relevant account entries and the customer’s explanation before reaching a decision. For a privacy-sensitive case, staff may need a specialist pathway rather than broader access to the entire transcript. Resolve uncertainty about permitted information handling with your adviser.
Monitor escalation quality as well as volume. Suggested measures are missed mandatory handovers in sampled conversations, transfers without an accepted case, time to first staff response, repeat contact after apparent resolution and corrections to automated answers. These are proposed operational measures, not regulatory benchmarks. Avoid treating fewer transfers as success without checking customer outcomes.
A useful comparison is business.gov.au’s published AI approach: it says people check information before publication and remain responsible for important decisions. That is an example of a deliberately bounded operating model, not a requirement for SMEs to copy. Your review design should likewise specify what people actually verify, rather than rely on the label “human-in-the-loop”.
When a structured contact form is the better choice
Do not add a conversational bot merely to redirect most conversations to staff. If your common enquiries involve billing disputes, individual circumstances or confidential documents, a structured contact form may be the clearer starting point. It can let customers select an intent and route the request directly to the right team without AI interpretation.
An illustrative form could offer “General information”, “Billing dispute”, “Policy exception” and “Request a person”. Ask only for the minimum details needed to begin. Provide a separate approved route for sensitive supporting material, a receipt reference and realistic response expectations. Keep a general option available so customers are not blocked by categories that do not fit.
The form sacrifices conversational convenience and may require customers to choose a category. The bot introduces a different burden: evaluating triggers, maintaining approved sources, supervising summaries and checking that transfers work. Compare the complete workflows, including rework and staff handling, rather than comparing chatbot response speed with the time needed for a person to resolve a case.
Approve the chatbot bypass only when there is a useful routine-answer scope, a working handover route, staffed ownership and evidence that the trigger tests pass. Otherwise, start with the form or a staff-assisted process. Record the decision owner, excluded intents, review evidence, fallback method and next review date. The immediate next step is to test the proposed handover end to end—including a closed queue—before allowing the bot to answer customers.
Sources
General information only; verify sources before making business decisions.