The customer feels the handoff, even when the system hides it
A customer asks a website assistant whether a service suits their situation. The assistant gives a broad answer, asks for their contact details, then reaches its limit. It offers a phone number. When the customer calls, they have to explain the problem, repeat the details, and work out whether the first conversation created a record at all.
This composite pattern is a useful test. The customer experiences the AI system, form, CRM, and support queue as one journey. Weak joins make it feel careless.
AI customer experience design starts with the customer moment, not the chatbot. The aim is to reduce effort while preserving a clear route to judgement, accountability, and recovery. That matters across initial enquiries, self-service, onboarding, support, personalisation, and follow-up.
The Australian Government's 10 voluntary AI safety guardrails connect accountable ownership, risk assessment, testing, disclosure, human control, and stakeholder needs. While voluntary, they offer a practical deployment backbone for deciding what a customer-facing system may do and when someone must intervene.
Map the customer moment before choosing the AI role
Begin with what the customer is trying to achieve. A person checking opening hours has a different need from someone disputing a charge, sharing health information, or asking for advice that could materially affect them.
For each moment, record the customer goal, consequence of an error, permissions required, data sensitivity, reversibility, urgency, and need for human judgement. These inputs lead to five useful AI roles.
| Role | Suitable customer moment | Boundary |
|---|---|---|
| Answer | explain a stable, approved fact | use current source material and show a human route |
| Assist | help a customer understand options | distinguish guidance from a binding decision |
| Collect | gather the minimum details for the next step | explain data use and avoid unnecessary sensitive information |
| Route | send the request to the right person or process | preserve intent and permitted context |
| Stop | pause when risk, uncertainty, or permissions exceed the design | make the alternate route immediate and usable |
The NIST Generative AI Profile supports defining the specific tasks a system will perform, documenting knowledge limits and human oversight, and evaluating output against known evidence. That makes the role a design constraint, not a label added after launch.
If the business has not yet established whether the underlying process is ready, start by choosing and scoping the AI-assisted workflow. Keep the customer-moment map focused on the interaction in front of the person.
Set boundaries around consequence, permissions, and knowledge
A confidence score can't decide whether an answer is safe. It may indicate how a model or classifier rates an output, but it can't account for sensitive information, reversibility, source currency, or the business's permission to act.
Write four boundaries before designing the conversation.
- Define what the system knows and which approved sources it may use.
- Define what it can say, collect, recommend, change, or send.
- Define what it must never imply, including certainty, professional authority, approval, or a binding outcome it cannot provide.
- Define the triggers that stop automation and give the next decision to a named person or team.
Consider a composite service enquiry. An assistant can answer from approved service pages, collect a suburb and preferred timeframe, and route the request. It should stop if the person asks for a firm quote without the required inspection, shares sensitive information outside the approved intake, disputes a previous decision, or needs urgent help. The person receiving the enquiry owns the judgement and the customer receives a clear expectation about what happens next.
Source quality is part of this boundary. Clear service facts, labelled options, current FAQs, and accessible form fields make the permitted knowledge easier to control. The same discipline used for making service information clearer and more structured helps an assistant retrieve reliable business information.
The Australian voluntary guardrails call for clearly defined acceptance criteria, testing before deployment, monitoring after launch, and meaningful human intervention across the lifecycle. The NIST AI RMF adds a connected Govern, Map, Measure, and Manage process. Businesses that need the ownership, approval, incident, and review mechanics can use the separate guide to put governance controls around customer-facing AI.
Tell customers when AI is involved and what it can do
Disclosure should help the customer make a decision in the moment. A long policy hidden in the footer leaves them guessing whether the conversation is automated, what information is needed, and how to reach a person.
A concise interface pattern can cover five points.
You're chatting with an AI assistant that can answer questions from our service information and help route your enquiry. Please don't share sensitive information here. We may save the details you choose to provide so our team can follow up. You can ask for a person at any time or use our contact form.
The exact wording must match the real system, its data handling, and the organisation's obligations. A disclosure that promises a human route is only useful when that route works.
The OAIC guidance for commercially available AI products says public-facing AI tools such as chatbots should be clearly identified. It also calls for due diligence, clear privacy policies and notices, human oversight, and accuracy controls when covered entities handle personal information. Privacy Act coverage and sector obligations depend on the organisation and activity, so a business should confirm its position rather than assume every Australian business has identical duties.
Two Australian Government standards offer more detailed interface benchmarks. AI Technical Standard Statement 10 asks agencies to identify AI interactions, explain limitations, provide alternate channels, design for accessibility, and define human oversight. The Standard for AI transparency statements asks agencies to explain why AI is used, where the public interacts with it or is significantly affected, how effectiveness is monitored, and how impacts are managed. These are government requirements, not universal duties for private businesses. They are valuable design benchmarks because they turn abstract trust into information a customer can see and act on.
Design privacy and accessibility into the route
Privacy notices work best at the point where a person decides what to share. Before each field or conversational prompt, ask whether the information is necessary for the current role. A system that routes an enquiry may need a service category and contact preference. A detailed personal history is likely unnecessary.
The OAIC states that, where the Privacy Act applies, personal information entered into an AI system and personal information generated by it must be handled under the Australian Privacy Principles. It recommends as best practice that organisations do not put personal information, particularly sensitive information, into publicly available generative AI tools. It also warns that inferred or incorrect output about an identifiable person can still be personal information.
Use a stop condition when the customer begins sharing information the system is not approved to handle. Explain why the conversation is pausing, do not repeat the sensitive text unnecessarily, and offer a suitable private channel. Obtain specialist privacy or legal advice when the use involves sensitive information, significant decisions, regulated services, children, or people experiencing vulnerability.
Accessibility belongs in the same route design. Identify controls clearly, preserve keyboard operation and focus order, announce status changes to assistive technology, use plain language, and avoid relying on colour alone. Keep a non-AI route available at a comparable point in the journey. If navigation, forms, analytics, CRM capture, or fallback channels cannot support that route, check the wider website journey before adding AI.
Make the human handoff carry the conversation forward
A safe handoff transfers responsibility and permitted context. A button that sends the customer elsewhere only changes the channel.
- Detect a defined trigger such as a request for a person, a prohibited topic, sensitive data, urgency, repeated failure, or an action outside the system's authority.
- Explain that a person needs to take over and obtain any consent required to transfer the relevant details.
- Summarise the customer's goal, the steps already tried, and the approved facts collected.
- Transfer only the information permitted for that team and purpose.
- Confirm that the receiving person or queue has accepted the handoff.
- Set an expectation for channel and timing without making a promise the team cannot keep.
- Provide a recovery route if the transfer fails or the customer cannot use the proposed channel.
The difference is visible in the journey.
| Bot dead end | Designed escalation |
|---|---|
| The customer asks twice for help | The first request for a person triggers the route |
| The bot asks for details without explaining why | The assistant explains purpose and collects only necessary details |
| A phone number appears with no transfer | The system sends a permitted summary to an accepted queue |
| The customer repeats the story | The person confirms the goal and continues from the summary |
| Failure leaves no next step | A contact form or phone route remains available |
This composite comparison gives a design team concrete states to prototype and test. It doesn't represent measured results.
Launch with a scorecard that measures customer effort and trust
Measure whether the journey works for the customer and the team. Deflection on its own can hide abandonment, repeat contact, incorrect answers, or extra work after a poor transfer.
Capture a baseline before launch, then compare like with like. Verify samples instead of treating every completed chat or closed ticket as a successful outcome.
| Measure | Signal | Decision |
|---|---|---|
| Verified resolution | the customer's goal was correctly completed | expand only when quality holds |
| Escalation | AI routes a case to a person | inspect whether triggers fire at the right time |
| Repeat contact | the same issue returns through another channel | find lost context or false resolution |
| Correction | a person changes an AI answer or summary | improve sources, scope, or review |
| Abandonment | the customer leaves before a useful next step | reduce effort and expose the human route |
| Customer effort | steps, repetitions, and channel changes | remove avoidable work |
| Handoff rework | staff rebuild missing or unusable context | change the transfer payload |
| Privacy incident | information crosses an unapproved boundary | pause, investigate, and remediate |
| Accessibility failure | a user cannot complete or exit the route | fix before expanding access |
Response time is useful beside these measures. Speed can't redeem a wrong answer or instant dead end. The full method for baselines, verified outcomes, exception rates, review effort, and scale, fix, or stop decisions sits in the guide to measuring AI workflow outcomes and reliability after launch.
Choose the next design decision
Take one customer moment and make five decisions. Clarify the customer's goal. Choose whether AI should answer, assist, collect, route, or stop. Set the knowledge, action, data, and consequence boundaries. Design an accessible human route that carries permitted context. Record a baseline that will show whether effort and trust improve.
That diagnostic often finds the constraint elsewhere. It may be unclear service information, a form that collects the wrong details, a disconnected CRM, an inaccessible contact route, or nobody assigned to accept the handoff.
Fix the weakest join before exposing the journey to more customers. That may mean designing a clearer website and customer journey so the AI interaction, forms, content, measurement, and human follow-up share one coherent route.