Most organisations are not there yet. The National AI Centre's SME AI adoption tracker found that 43% of Australian small and medium businesses used AI in some form between December 2025 and February 2026. The most common uses were content generation and data analytics, at 54% of adopters each. Agentic AI was described as "largely untapped". Among businesses not adopting, around 65% cited distrust of AI decision-making or a preference to keep humans in control, and 19% said they simply didn't know how to use AI in their business.
The risk of choosing badly is real. Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. It also estimates that only about 130 of the thousands of vendors marketing "agentic AI" offer the real thing. This guide is for Australian CIOs, COOs and finance, operations and service leaders who want a defensible shortlist. It sets out 40 use cases by department, a scorecard for ranking them, what agents cost to run in AUD, and the Australian rules that apply.
What makes a use case "agentic" rather than plain automation?
An AI agent is software that is given a goal, decides which steps to take, uses tools (APIs, searches, other systems) to take them, and checks the result. Automation follows a fixed script. An assistant answers a question when asked. The difference matters because agents cost more to build, run and govern, so they should only be used where the extra judgement earns its keep.
Gartner's guidance is a useful rule of thumb: use agents where decisions are needed, automation for routine workflows, and assistants for simple retrieval. In practice:
| Approach | Best for | Example | Typical tools |
|---|---|---|---|
| Assistant | Finding or summarising information on request | "Summarise this customer's last five cases" | Microsoft 365 Copilot, Rovo Chat, Gemini |
| Rules-based automation | Same steps every time, structured inputs | Create a Jira issue when a Salesforce case is escalated | Power Automate, Jira automation, n8n, MuleSoft |
| AI agent | Messy inputs, several possible next steps, judgement inside clear limits | Read supplier emails, work out which purchase orders have changed, update the ERP and draft replies | Agentforce, Copilot Studio, Dynamics 365 agents, Rovo agents, custom agents on n8n or a model API |
Many good "AI" outcomes are still won with plain automation. If the steps never vary, a script is cheaper, faster and easier to audit than an agent. Gartner makes the same point: many use cases positioned as agentic don't need an agentic implementation.
How do we decide which use cases to start with?
Score each candidate out of 25 on five criteria, then start with the two or three highest scorers that don't trip a hard stop. This is the scorecard we use to frame first conversations with clients. It is deliberately simple, so a business owner and an architect can fill it in together in under an hour.
| Criterion | Score 1 when… | Score 5 when… |
|---|---|---|
| Volume and value | A few cases a month, minutes each | Hundreds or thousands a month, or each case worth real money (cash, margin, churn) |
| Judgement needed | Fixed steps that a script could do | Unstructured inputs (emails, PDFs, free text) and several valid next steps |
| Data reachable | Information lives in inboxes, spreadsheets or people's heads | Systems of record have APIs and the data is reasonably clean |
| Reversible and low harm | A wrong action moves money, changes someone's entitlement or reaches a customer unchecked | A wrong action is easy to spot and undo, and affects no individual's rights |
| Process documented | Nobody can write down the rules or the exceptions | Steps, business rules and exception paths are written down and agreed |
Hard stops override the score. Don't let an agent act on its own, at least at first, where it would:
- make or substantially inform a decision that significantly affects an individual's rights or interests, such as credit, pricing for a person, hiring or eligibility. From 10 December 2026 these decisions bring new Privacy Act transparency obligations (see our guide to automated decisions and the Privacy Act)
- release payments or change bank details
- send external communications without review in a regulated context, such as financial advice or health information
- change production systems, security settings or access rights.
Then set the autonomy level. We describe three levels, and recommend every new agent starts at level 1:
- Draft: the agent prepares the work (a match, a reply, a record update) and a person approves each one.
- Act with exceptions: the agent completes routine cases and routes anything unusual, low-confidence or above a threshold to a person.
- Act and report: the agent completes the work and people review a sample and the exception log.
Move up a level only when you have measured the agent's accuracy at the current level over a meaningful volume, such as several hundred cases. The autonomy column in each table below is where we suggest starting, not where you should end up.
Which finance use cases suit AI agents?
Finance is often the best first department because the work is high volume, the rules are written down and every action leaves a ledger trail. Microsoft now ships finance agents for Dynamics 365, including an Account Reconciliation agent and an Expense agent, announced in May 2025.
| # | Use case | What the agent does | Start at |
|---|---|---|---|
| 1 | Accounts payable invoice matching | Reads supplier invoices (PDF or email), matches them to purchase orders and receipts, and queues mismatches with a suggested reason | Level 2 |
| 2 | Account reconciliation at month end | Matches bank, AP and AR sub-ledger entries to the general ledger and flags unexplained differences | Level 2 |
| 3 | Expense report checking | Extracts receipt details, codes expenses to policy and GST treatment, and flags out-of-policy claims | Level 2 |
| 4 | Collections follow-up | Prioritises overdue accounts, drafts reminder emails in the right tone for each customer's history, and logs promises to pay | Level 1 |
| 5 | Variance commentary | Drafts the first pass of monthly budget-versus-actual commentary from the ledger and the prior month's notes | Level 1 |
| 6 | Supplier master data checks | Checks new supplier records for ABN validity, duplicates and missing fields before a person approves them | Level 1 |
Keep payment release and changes to supplier bank details with people, whatever the agent's accuracy. Those are the actions that payment-redirection fraud targets.
Which sales use cases suit AI agents?
Sales agents work best on the administrative load around selling, not the relationship itself. Most of these run inside the CRM, so Salesforce customers will usually look at Agentforce first, and Dynamics 365 Sales customers at Microsoft's sales agents.
| # | Use case | What the agent does | Start at |
|---|---|---|---|
| 7 | Inbound lead qualification | Researches each inbound lead against your ideal customer profile, scores it, and drafts a first reply for the rep | Level 1 |
| 8 | CRM hygiene | Finds stale opportunities, missing next steps and duplicate accounts, and proposes fixes in a weekly batch | Level 2 |
| 9 | Meeting follow-up | Turns call notes into a summary, CRM updates and a follow-up email draft | Level 1 |
| 10 | Quote preparation | Assembles a draft quote from the opportunity, price book and standard terms, and flags non-standard discounts for approval | Level 1 |
| 11 | Renewal preparation | Builds a renewal pack 90 days out from usage, support history and contract terms | Level 2 |
Which marketing use cases suit AI agents?
Marketing already uses generative AI heavily for content. The agentic step is linking that content work to data and channels, with brand and consent rules applied automatically. Consent and opt-out handling under the Spam Act and the Privacy Act should stay rule-based, never left to a model's judgement.
| # | Use case | What the agent does | Start at |
|---|---|---|---|
| 12 | Campaign build checks | Checks a campaign before send for broken links, missing unsubscribe, consent filters and brand rules | Level 2 |
| 13 | Segment suggestions | Proposes audience segments from CDP or CRM data for a stated campaign goal, with estimated sizes | Level 1 |
| 14 | Content localisation | Adapts approved copy and assets for regions or channels and routes them for review | Level 1 |
| 15 | Digital asset tagging | Tags new assets in the DAM with products, usage rights and expiry dates | Level 2 |
| 16 | Performance reporting | Pulls channel results into a weekly report with plain-English commentary on what changed | Level 2 |
Which customer service use cases suit AI agents?
Customer service has the most mature agent tooling, and the most visible risk, because a wrong answer reaches a customer directly. Salesforce's own Agentforce pricing examples use "Where is my order?" as the model case: the agent authenticates the customer, looks up the order and replies.
| # | Use case | What the agent does | Start at |
|---|---|---|---|
| 17 | Order and booking status | Authenticates the customer, retrieves the order or booking and gives a status and next step | Level 2 |
| 18 | Case triage and routing | Classifies incoming emails and web forms, sets priority and routes them to the right queue with a summary | Level 2 |
| 19 | Agent assist | Suggests replies to human agents from the knowledge base and similar past cases | Level 1 |
| 20 | Returns and warranty intake | Collects the details, checks eligibility rules and creates the return, handing exceptions to a person | Level 1 |
| 21 | Knowledge gap detection | Finds frequent questions with no good knowledge article and drafts one for review | Level 1 |
| 22 | Complaint escalation watch | Flags cases showing signs of a formal complaint or regulator escalation, so a senior person sees them early | Level 2 |
Which IT and service management use cases suit AI agents?
IT teams are often the easiest place to prove value, because the work already runs through a ticketing system with clean data. On Atlassian, Rovo comes with out-of-the-box agents that cover several of these, including a Request Router for triage, a Service Request Helper for drafting responses and Ops Expert for incidents.
| # | Use case | What the agent does | Start at |
|---|---|---|---|
| 23 | Service request triage | Categorises requests, fills in missing fields and assigns them to the right team | Level 2 |
| 24 | Password and access requests | Handles standard access requests against role rules, routing anything privileged to a person | Level 1 |
| 25 | Incident summaries | Writes the running incident summary and the first draft of the post-incident review from the ticket, chat and alert history | Level 2 |
| 26 | Change risk checks | Reviews change requests against recent incidents and the change calendar, and flags clashes or missing rollback plans | Level 1 |
| 27 | Licence and SaaS clean-up | Finds inactive users and duplicate tools across SaaS platforms and proposes removals | Level 1 |
| 28 | Integration monitoring | Watches integration error queues, retries known transient failures and opens tickets with diagnosis for the rest | Level 2 |
Use case 28 is where agents and integration meet. An agent can only act on systems it can reach, which is why our connected enterprise guide argues that the integration backlog has become the AI backlog.
Which HR and people use cases suit AI agents?
HR is where the hard stops bite hardest. Anything that ranks, screens or decides about a person is in Privacy Act territory, and in hiring it also carries anti-discrimination risk. The safe agent work in HR is administrative and employee-facing.
| # | Use case | What the agent does | Start at |
|---|---|---|---|
| 29 | Policy questions | Answers employee questions about leave, expenses and policies from the approved handbook, with links to the source | Level 2 |
| 30 | Onboarding coordination | Opens the IT, facilities and payroll tasks for a new starter and chases them until they're done | Level 2 |
| 31 | Interview scheduling | Finds times across interview panels and candidates and sends invitations | Level 2 |
| 32 | Timesheet reminders | Prompts staff to complete timesheets and flags entries that look wrong to their manager | Level 2 |
Leave candidate screening and performance ratings out of the first wave. If you do use AI there later, treat it as a high-risk use under the National AI Centre's guidance, described below.
Which operations and supply chain use cases suit AI agents?
Operations work is full of email-driven exceptions between the ERP and outside parties. Microsoft's Supplier Communications agent for Dynamics 365 is built for exactly this: it chases missing purchase order confirmations and handles delayed shipments by email.
| # | Use case | What the agent does | Start at |
|---|---|---|---|
| 33 | Supplier confirmation chasing | Reads supplier replies, works out which purchase order lines changed, updates dates and drafts responses | Level 1 |
| 34 | Sales order intake | Turns emailed or PDF orders into draft sales orders and checks prices, stock and credit before a person confirms them | Level 1 |
| 35 | Stock exception alerts | Spots likely stock-outs or overstock from sales and supply signals and proposes transfers or orders | Level 1 |
| 36 | Freight and delivery updates | Monitors carrier updates and tells customers and internal teams about delays, with a revised date | Level 2 |
| 37 | Product data enrichment | Fills missing product attributes in the PIM from supplier data sheets for review | Level 1 |
Our article on agentic AI in PIM integration covers use case 37 in more depth. For agents across an ERP program, see our agentic AI for ERP implementations solution.
Which project and PMO use cases suit AI agents?
Project offices spend much of their week collecting status and chasing updates. Agents can do the collecting, so people can spend the time on the decisions.
| # | Use case | What the agent does | Start at |
|---|---|---|---|
| 38 | Status report drafting | Builds the weekly status report from Jira, timesheets and the RAID log, and asks owners to confirm | Level 2 |
| 39 | Risk and dependency watch | Flags slipping milestones and cross-project dependencies before the steering meeting | Level 2 |
| 40 | Requirements and meeting notes | Turns workshop notes into draft requirements, decisions and actions in Confluence and Jira | Level 1 |
What does it cost to run AI agents in AUD?
Agent platforms now charge mostly by usage, so the running cost depends on how many actions your agents take. Build and integration effort is usually the larger cost in year one. List prices at the time of writing:
| Platform | How it's charged | List price |
|---|---|---|
| Salesforce Agentforce | Flex Credits: a standard action uses 20 credits, a voice action 30 | AU$700 per 100,000 credits, which is about AU$0.14 per standard action. The Agentforce User Licence, which gives every employee access to agents, is AU$7 per user per month and also draws on Flex Credits |
| Microsoft 365 Copilot and Copilot Studio | Per-user licence for Copilot; Copilot Credits for autonomous and external agents, pre-purchased or pay-as-you-go through Azure | Microsoft 365 Copilot is AU$44.90 per user per month, paid yearly, excluding GST |
| Atlassian Rovo | Monthly Rovo credit allowance by plan, with billing for extra usage from December 2026 | See our breakdown of Atlassian Cloud pricing in AUD |
| Custom agents (for example on n8n with a model API) | Model tokens plus hosting | Varies with the model and volume; self-hosting keeps data in your own environment |
A worked example. Say a retailer's service agent handles 3,000 "where is my order?" enquiries a month, and each takes two Agentforce actions (authenticate, then retrieve the order). That's 6,000 actions, or 120,000 Flex Credits, which is about AU$840 a month at list price. Compare that with the staff time those 3,000 contacts take today, and add the build cost and the time spent reviewing the agent's work while it's at level 1 or 2. That comparison, not the licence line, is the business case.
Two cost traps to plan for. First, metered usage grows quietly once agents are popular, so set budget alerts and caps from day one. Second, an agent that loops or retries can burn credits fast. Test for this before go-live, and monitor actions per completed case afterwards.
What do Australian rules require when we deploy agents?
Three sets of rules matter for most mid-market organisations. None of them bans agents, but each shapes the design.
- The Australian Government's AI guidance. The National AI Centre's essential AI practices set out six practices: decide who is accountable, understand impacts and plan accordingly, measure and manage risks, share essential information, test and monitor, and maintain human control. They are voluntary, but they are the baseline your board, auditors and customers will expect. There is a foundations version for early use, and an implementation version for complex or higher-risk use, which asks for an AI register and clear accountability for each system.
- The Privacy Act. From 10 December 2026, privacy policies must explain where personal information is used in automated decisions that could significantly affect individuals. Our Privacy Act guide covers what counts and what to disclose. The Australian Privacy Principles also apply to any personal information an agent sends to a model hosted offshore.
- APRA CPS 230, for banks, insurers and superannuation trustees. If an agent platform supports a critical operation, the provider is likely to be a material service provider. That brings due diligence, contract and register requirements, and your tolerance levels must account for the agent failing.
In practice, that means four controls for every agent: a named business owner, a log of every action it takes, a tested way to switch it off, and a human approval step for anything on the hard-stop list. Our December guide to governing AI agents will cover board-level frameworks in more detail.
How do we get from a long list to a working pilot?
A focused pilot takes about 90 days from scoring to a measured result. Here is the sequence we recommend:
- Weeks 1–2: score and pick. Run the scorecard over your long list with business owners. Pick two or three use cases that score 18 or more out of 25 and don't trip a hard stop.
- Weeks 2–4: write the process down. Document the steps, rules, exceptions and who approves what. This is the step most often skipped, and the one agents depend on most.
- Weeks 3–6: connect the data. Give the agent read access first, then narrowly scoped write access through APIs, with a service account per agent.
- Weeks 5–10: run at level 1. The agent drafts and people approve. Measure accuracy, time saved per case and the cost per completed case.
- Weeks 10–13: decide. Move to level 2 where accuracy is proven, fix or drop what isn't working, and choose the next use cases.
The documentation step is where groundwork pays off. In our Microsoft Dynamics 365 program for a manufacturer with more than 1,000 employees, we ran more than 40 process discovery and design workshops, simplified and standardised over 120 business processes, and built a library of workflows, business rules and exception handling. That program was an ERP replacement, not an agent project, but a library like that is exactly what an agent needs to work from. Organisations that have one can move straight to connecting data.
Similarly, our Salesforce implementation for a law firm automated document generation, approvals and task assignment, cutting document processing time by 40%. Rules-based automation like that is often the right first step, and it leaves clean, structured data for agents to build on later.
Frequently asked questions
Do we need a data platform before we can use AI agents?
No. Most first-wave use cases work against the systems you already run, such as the ERP, CRM and service desk, through their APIs. What you need is reachable, reasonably clean data for the specific use case, not an enterprise-wide data program. A data platform becomes worthwhile once several agents need the same joined-up view of customers or products.
Should we use the agents built into our platforms or build our own?
Start with the agents built into the platforms where the work already happens: Agentforce for work inside Salesforce, Dynamics 365 and Copilot agents for Microsoft, and Rovo for Jira and Confluence. Build custom agents when a use case spans several platforms, or when you need data to stay in your own environment. Platform agents are quicker to start but tie you to each vendor's pricing meter.
How many use cases should a mid-market organisation run at once?
Two or three in the first 90 days. More than that spreads the scarce people, such as process owners, integration developers and reviewers, too thin to measure anything properly.
Will AI agents replace roles?
In the first wave, agents take on the repetitive parts of roles, such as matching, chasing, triaging and drafting, while people approve and handle exceptions. Plan for that change openly with the teams affected. The NAIC data shows trust is the biggest barrier to adoption, and staff trust matters as much as customer trust.
Where should we start?
Pick one department where volume is high and the rules are written down. In most organisations that's finance, IT service management or customer service. Score five candidate use cases with the scorecard above, and run the top two at level 1 for a quarter. Measure accuracy and cost per case before you widen anything.
Great Minds designs and builds agentic AI and enterprise automation across Salesforce, Microsoft, Atlassian and custom platforms, including Rovo. If you'd like help scoring your shortlist and scoping a 90-day pilot, talk to our team.
