HUB · 5 GUIDES
AI agents
See where AI agents genuinely cut response times, and where deterministic automation is safer. Talk to National Digital.
Quick answer: AI agents combine language-model understanding with deterministic automation; the deciding factor is handoff design and ongoing monitoring, not which model powers the bot.
Last updated
Quick answer
What is an AI agent, and how does it differ from a scripted bot?
An AI agent uses a language model to interpret intent and draft responses; a scripted bot follows fixed decision trees. Most Australian deployments combine both for cost and control.
Additional Context
The distinction matters for scoping: language-model assistants handle open-ended enquiries well but need guardrails and human escalation for anything involving money, legal obligations or complaints.
Sources
- Australia's AI Ethics Principles
Eight voluntary principles for responsible AI design, including transparency, human oversight and accountability.
- OAIC guidance on artificial intelligence and privacy
Explains how the Australian Privacy Principles apply when AI systems collect or process personal information.
How AI Agents Fit Into Operations
The term "AI chatbot" covers two different tools that get bundled together far too often. One is a decision-tree bot that matches keywords to scripted replies, cheap to build and predictable but brittle the moment a question falls outside its branches. The other is a language-model assistant that reads intent, holds context across a conversation and drafts a response rather than selecting one from a list. Business automation projects that succeed treat these as separate build decisions rather than a single product to procure.
What counts as an AI agent
A working definition: if the system can handle a question it wasn't explicitly scripted for, it's doing AI-assisted work rather than simple workflow automation. That capability is genuinely useful for enquiries that vary in phrasing but follow predictable patterns, such as order status, appointment changes or account access requests. It's a poor fit for anything requiring a judgement call on liability, refunds outside policy, or a complaint that's already escalated once.
Where the technology earns its keep
Three areas consistently justify the build cost. Front-line support is the obvious one: support automation localisation work usually starts here, because the volume of repetitive enquiries is high and the cost of getting tone wrong is low. Sales teams get similar value from AI lead scoring that triages inbound enquiries before a rep sees them, cutting the time spent on leads that were never going to convert. The third is internal: businesses running several customer touchpoints, a web chat, a Shopify storefront and a Facebook page, increasingly want one assistant behind cross-channel support automation rather than three disconnected bots with no shared memory of the conversation.
- Order status, account access and appointment changes
- Initial sales enquiry triage and qualification
- Internal IT and HR request intake
- After-hours cover for common questions
A Pattern Worth Planning Around
The Escalation Gap Most Deployments Miss
Problem
Businesses deploy an AI agent to cut response times, then find satisfaction drops because the handoff to a human agent loses the context of everything the bot already asked. Agents re-ask the same questions, customers repeat themselves, and the bot gets blamed for a design decision nobody scoped.
Common Pain Points:
- Context loss at handoff Agents receive an escalated chat with no visibility into what the bot already asked or promised.
- Unclear ownership of edge cases Nobody has decided who handles a request the bot can't classify, so it sits unresolved.
- No measurement of what's working Teams launch a bot and rarely check containment or escalation quality after week one.
Business Impact:
Time Wasted: Repeated data-gathering every time a conversation escalates to a person
Cost Implication: Ongoing duplicated handling between the bot and support staff
Opportunity Cost: Complex or high-value enquiries stall in the handoff instead of reaching someone who can act on them
Solution
Scope deterministic tasks and open-ended judgement calls separately, define the handoff data package upfront, and monitor escalation quality before expanding what the assistant handles.
Our Approach:
- Map the current enquiry mix
Categorise existing tickets or calls by whether they need policy lookup, judgement or genuine escalation.
- Design the handoff package
Define exactly what context, history and confidence score passes to a human agent when the bot exits a conversation.
- Pilot on a narrow enquiry set
Launch against one or two enquiry types before widening scope, watching escalation rates closely.
Expected Outcome: Fewer repeated conversations at handoff and clearer ownership of enquiries once they reach a person.
Key Takeaways
What Actually Determines AI Agent Success
- Scope deterministic tasks separately from judgement callsImportant
Order status lookups and account changes suit rule-based logic; refund disputes and complaints need a person, and blending the two in one bot confuses both.
- The handoff package matters more than the bot's intelligenceImportant
A highly capable assistant that hands off with no context creates more work than a basic bot with a clean, complete handoff to the agent who picks up the conversation.
- Australian English and local phrasing need explicit tuningImportant
Off-the-shelf language models default to American idiom and spelling; without deliberate correction, responses read as noticeably foreign to local customers.
- Escalation quality should be measured from week oneImportant
Containment rate alone hides problems; tracking what happens after escalation shows whether the handoff is actually working for customers and staff.
AI agent projects succeed or fail on scoping and handoff design, not on which language model sits behind them. Measuring escalation quality early catches problems before they become customer complaints.
Where AI Agent Deployments Meet Regulation
AI agents that collect or process personal information sit inside existing Australian regulatory obligations, independent of which vendor platform is used.
APPs apply regardless of automation
Privacy obligation trigger
Significance: highAny AI agent collecting, storing or disclosing personal information falls under the Australian Privacy Principles whether a human or an AI model processes that data.
Source: Office of the Australian Information Commissioner, AI and privacy guidance
8 principles
Voluntary AI governance framework
Significance: mediumAustralia's AI Ethics Principles set out eight voluntary principles, including transparency, human oversight and accountability, that responsible AI agent deployments increasingly reference in design.
Source: Department of Industry, Science and Resources
Under Privacy Act reform
Automated decision-making scrutiny
Significance: mediumThe federal Privacy Act Review flagged a right for individuals to request meaningful information about automated decisions, directly relevant to AI assistants that triage or reject customer requests.
Source: Attorney-General's Department, Privacy Act Review Report
Methodology
Reference points draw on published Australian Government guidance and statutory review documents rather than vendor or survey estimates, reflecting obligations that apply regardless of platform choice.
Build Decisions That Actually Matter
Deciding what belongs to rules-based process automation and what belongs to an AI assistant is the single most consequential design decision in any AI agent project. Deterministic workflow automation software handles the same input the same way every time, which is exactly what a regulator or an auditor wants to see for anything involving payments, eligibility or compliance. AI judgement suits open-ended language and ambiguity well, but it introduces variability that needs ongoing monitoring rather than a one-off sign-off.
Deterministic automation versus AI judgement
Agentic AI, where an assistant plans a sequence of actions and calls other systems on its own, is still early for most mid-sized deployments. The more reliable pattern seen across implementations is an AI layer for language understanding sitting in front of deterministic automation vs AI logic for anything that touches a system of record. That combination gets the flexibility of natural language without handing an unsupervised model the authority to approve a refund or close a ticket.
What implementation reveals
The gap that shows up repeatedly once a bot is live isn't the model's accuracy, it's the absence of a feedback loop. Teams launch, watch containment rate for a fortnight, then stop looking. Six months later nobody can say whether the assistant is still answering correctly, because the question mix has shifted and no one is watching conversation analytics for drift. Building that review into a monthly operations task, not a one-off launch metric, is what separates an AI agent that keeps earning its keep from one that quietly degrades.
What an AI automation costs
An automation that takes a repetitive judgement-heavy task off a team - triage, extraction, drafting, routing - wired into the systems the work already lives in. Priced for one production workflow, evaluated against real cases, and delivered with two training sessions for the people who will run it.
| Planning and evaluation | |
|---|---|
| What the task actually is, where the data comes from, and how anyone will know the automation is right often enough to trust. | |
| Process and data auditThe task as it is performed, and whether the inputs it depends on are reachable and clean enough to automate against. | $1,500 - $5,000 |
| Evaluation and integration designAn agreed standard for a correct answer, scored on real historical cases, plus the contracts against the systems the automation reads and writes. Without this there is no way to tell improvement from noise. | $2,000 - $6,000 |
| Build and release | |
| The working automation, and what it takes to run it in production with a human able to see and correct it. | |
| Automation buildThe workflow itself: prompts or models, the retrieval and tool calls around them, and the fallback path for the cases it should refuse to handle. | $5,000 - $21,000 |
| Rollout, monitoring and handoverStaged rollout behind human review, logging that makes a wrong answer traceable, and two training sessions that leave the team able to run and adjust the automation themselves. | $1,500 - $8,000 |
| Total Investment RangeTypical project: $25,000 | $10,000 - $40,000 |
Payment Terms
Staged: the planning and evaluation phase is invoiced on commencement, the build on release. Model and API running costs are billed to the client by the provider, not through us.
Key Assumptions
- One workflow in production, not a platform.
- Model and API running costs are the client's and billed by the provider.
- Every build includes two training sessions for the people who will operate the automation, with the length set against the rollout.
- A human stays in the loop wherever a wrong answer would reach a customer unreviewed.
These are the ranges a project like this usually lands in. Answer seven questions and we will narrow it to yours.
Talk to an AI Automation Specialist
Discuss where an AI agent would help most, and where deterministic automation is still the better fit. A straightforward assessment against your current systems.
AI Agents: Common Questions
What's the practical difference between a rules-based bot and an AI agent?
A rules-based bot matches keywords to a fixed set of scripted replies and breaks down outside those branches. An AI agent uses a language model to interpret intent and generate a response, handling phrasing it hasn't seen before. Most working deployments combine both: rules for anything touching a system of record, and AI for open-ended language, with a clear handoff between the two.
What is agentic AI, and how does it differ from standard automation?
Standard automation follows a fixed sequence someone has already designed. Agentic AI plans its own sequence of steps and calls other systems to complete a task without a human specifying each action in advance. It's a genuinely different capability, but for most operations teams it remains higher-risk than proven, and a supervised AI layer over deterministic automation delivers similar value with far less exposure.
Can an AI agent be trained to sound like an Australian business rather than a generic assistant?
Yes, but it needs deliberate work. Language models default to American spelling, idiom and tone unless explicitly tuned, and Australian customers notice the mismatch quickly. Correcting it means adjusting prompts, reviewing sample conversations for phrasing that reads as foreign, and checking the bot doesn't default to terminology that doesn't match local products, regulations or customer expectations.
Where should call centre automation start if a full AI rollout isn't realistic yet?
Start with the enquiries that are highest volume and lowest risk: order status, appointment changes, account access and basic account queries. These suit deterministic logic or a narrow AI assistant without exposing complaints, refunds or anything involving liability to automated handling. Expanding scope after measuring containment and escalation quality on that narrow set is a safer sequence than a broad launch.
How do we decide between building an AI agent in-house and engaging an AI automation agency?
In-house build works when there's existing engineering capacity and the use case is narrow. An external AI automation consultant earns its cost when the project needs integration across several systems, ongoing model tuning, or compliance review that internal teams haven't done before. The honest test is whether the team maintaining the bot after launch actually has time to keep watching it.
What ongoing work does an AI agent need once it's live?
A launched AI agent still needs someone reviewing escalation quality, checking containment rate against the current enquiry mix, and retraining when the question set shifts. Businesses that treat launch as the finish line typically see quiet accuracy drift within months, because the assistant keeps answering old questions well while the actual enquiries have moved on.