What changes when customers see the answer directly?
With a staff assistant, the answer is a draft for an employee who knows the business. They can spot an outdated fee, an odd tone or a policy that does not apply, and fix it before it reaches anyone outside. With a customer chatbot, nobody stands between the model and the customer. The Australian Government's Guidance for AI Adoption uses a chatbot to illustrate this: one that answers simple questions during business hours while a staff member monitors it is a low-risk use, and the risk expands if it runs around the clock without oversight. The level of supervision is the main variable you control.
Which questions suit each audience?
Customer questions tend to be repetitive and public: opening hours, how to lodge a claim, what documents to bring, where an order is. Staff questions are broader and draw on internal material such as procedures, templates, past decisions and product specifications. Pull a month of enquiries from your inbox, phone notes or ticketing system and sort them by source: could the answer come from a page you already publish, from an internal document, or only from a person's judgement?
- Customer-facing fits when answers come from published, stable content and a wrong answer is easy to correct.
- Staff-facing fits when answers need internal documents, judgement or material customers should not see.
- Neither fits when the answer depends on individual circumstances the system cannot check, such as an eligibility decision.
What does a customer chatbot need that a staff tool does not?
A public chatbot meets people you cannot vet, and some of them will test it on purpose. It needs controls that a signed-in staff tool can often do without, and each control needs an owner and a test before launch. Several of these also appear in Australian government and regulator guidance. Budget staff time to review conversations closely in the first weeks after launch. The list below is the minimum to plan for.
- Disclosure: the Guidance for AI Adoption says customers receiving services and information from a chatbot should know it is not a human.
- Handover: a clear route to a person, with the conversation passed along so the customer does not repeat themselves.
- Narrow scope: a fixed list of topics and approved sources, and a polite refusal outside them.
- Misuse defences: prompt injection is the first entry in OWASP's Top 10 for LLM applications, including indirect injection through content the model reads.
- Privacy notice: the OAIC expects privacy policies and notifications to explain how AI is used.
How do the data and privacy questions differ?
A staff assistant mostly reads your documents and employees' questions. A customer chatbot collects personal information from members of the public: names, account numbers, health details or complaints typed into a chat box. For organisations covered by the Privacy Act, the OAIC says privacy obligations apply to personal information input into an AI system and to the output data it generates. Where the conversation is stored and processed also matters. If customer conversations must stay in Australia, check the deployment settings described in our guide to LLMs in Australian cloud regions before you choose a platform.
How do you measure success for each?
Pick measures that show whether the tool helps, and compare them with a baseline taken before launch. For a staff assistant, useful measures include time to find an answer, the share of answers staff accept without edits, and the number of repeated questions sent to subject experts. For a customer chatbot, look at conversations resolved without handover, handovers that reached a person, complaints that mention the chatbot, and repeat contacts within a week. Avoid relying on conversation counts alone, because a chatbot that fails to resolve questions also raises the conversation count.
Can one system serve both audiences later?
Yes, if you plan the separation from the start. A common pattern is one retrieval platform with two collections: a public collection of approved customer content and an internal collection with permissions. The customer chatbot is given the public collection only, enforced in the retrieval layer instead of by instructions to the model. Staff can use both. Starting with the staff assistant lets you test retrieval quality and source maintenance with a forgiving audience, then publish a narrower customer version once the content owners have their update routine working.