How to identify AI and automation opportunities in your business: a practical method

Ask ten Australian business owners what's stopping them getting more value out of AI and most will point to the tools: which product, which licence tier, which vendor to trust. That's rarely the real gap. Microsoft Copilot, Power Automate and Copilot Studio are all available today, priced per seat, documented, and already sitting inside most Microsoft 365 tenants in some form. The gap is upstream of all of that. Very few businesses can say, with any confidence, exactly where in their operation an hour of automated work would save real money, and where it would just be a novelty nobody uses past week three.

That gap is what a structured discovery process is for. It's the part of the work that happens before a single agent gets built or a single Copilot licence gets assigned, and it's the part most businesses skip, because buying a tool feels like progress and running an audit feels like admin. The businesses getting a genuine return from AI and automation are, almost without exception, the ones who did the discovery work first. This guide covers the method: how to find candidate opportunities, how to tell a real one from a demo-friendly distraction, how to score and rank what's found, and where Microsoft 365's AI capability actually fits once that ranking exists.

Start from the task, not the tool

The single most common mistake in this space is starting with the product. A business hears about Copilot Studio agents, decides an agent sounds useful, and goes looking for something for it to do. That's the wrong direction of travel. It produces solutions in search of a problem, and it's why so many AI pilots quietly stop being used after the initial excitement fades, the underlying task was never painful enough to justify keeping the tool running.

The right starting point is the work itself, specifically the work that's repeated, rule-based, and currently done by a person moving information between systems or documents. That work leaves a trail. It shows up as recurring calendar blocks, as spreadsheets that get rebuilt every week, as the same three questions being answered in the same tone every time a customer emails, and as the quiet workarounds staff invent when a system doesn't do quite what they need. Finding automation opportunities is largely a matter of noticing that trail, not imagining new capability from scratch.

The audit: where do the hours actually go

The most reliable way to surface genuine opportunities is a structured hours audit, a short, deliberate exercise in tracing where time goes on tasks that are repeated rather than novel. It doesn't need to be exhaustive or run for months. A focused two-week window across the roles that touch the most volume, usually finance, operations, customer service and admin, is enough to surface the patterns that matter.

1. 1Pick the roles, not the whole business. Start with the two or three roles handling the highest transaction volume, invoicing, scheduling, order processing, customer enquiries, not the roles doing the most varied, judgement-heavy work.
2. 2Ask each person to log tasks for a fixed period, a week is usually enough, tagging each block of time against a short list: repeatable and rule-based, repeatable but judgement-based, one-off or novel.
3. 3In parallel, interview each role directly. People under-report the workarounds they've built for themselves, because they've stopped noticing them. A direct question, 'what do you copy and paste between two systems every week', surfaces more than a time log alone.
4. 4Look specifically for four signal patterns: copy-paste between systems that don't talk to each other, spreadsheets rebuilt from scratch on a schedule, the same question answered in writing more than a handful of times a week, and approval or reconciliation steps that exist purely to check someone else's manual work.
5. 5Write every candidate down with the same three facts attached: how often it happens, roughly how long it takes each time, and what happens downstream if it's done late or wrong.

• Copy-paste between systems, an order taken in one system and re-keyed into another, a lead captured in a form and manually entered into a CRM
• Spreadsheet reconciliation, a weekly or monthly file rebuilt from scratch by pulling numbers out of two or three other sources
• Repeated written answers, the same policy explanation, order status update, or pricing question typed out fresh each time it's asked
• Manual approvals and checks, a person re-verifying data that a system already validated, purely because nobody trusts the system to have done it correctly
• Scheduling and coordination, chasing availability across people or resources by email or phone when the constraints are actually well defined

By the end of this exercise, most businesses have a list of ten to twenty candidate tasks. That list is the raw material. The next step, and the one most businesses skip, is scoring it properly rather than picking whichever item sounds most impressive in a meeting.

Score every candidate on three axes, not one

A task feeling time-consuming isn't the same as it being a good automation candidate. The businesses that get this wrong tend to score on a single axis, usually 'how annoying is this', and end up building something that's technically impressive but doesn't move any number that matters, or something cheap to build that quietly breaks the moment an edge case appears. A useful score needs three separate axes, assessed independently before they're combined.

Business value

What does this task actually cost the business today, and what would fixing it actually return? Value has three components worth separating: hours saved (frequency multiplied by duration multiplied by the wage cost of the person doing it), error cost (what a mistake in this task costs when it happens, not just the time to redo it), and downstream effect (does this task block or delay something more valuable, like an invoice that can't go out until reconciliation finishes). A task that's only moderately time-consuming but sits upstream of monthly billing can outrank one that eats more hours but has no downstream consequence.

Effort to build and maintain

Effort isn't just build time. It includes integration complexity (does this touch two systems that already talk to each other via a supported connector, or does it need custom API work), change management load (how many people need to change a habit for this to work), and ongoing maintenance (does the underlying process change often enough that the automation needs constant rework). A task that looks like a quick win on the surface can carry a maintenance cost that erodes the return within a year, particularly where the source data or the process itself shifts regularly.

Data readiness

This is the axis businesses skip most often, and it's usually the one that decides whether a project ships on time. Is the data this task depends on structured, accessible, and reasonably clean, or does it live in someone's inbox, a shared drive full of inconsistently named files, or a system with no usable export? A high-value, low-effort task built on data that's scattered and inconsistent will cost far more than expected once the real work of cleaning and connecting that data begins. Data readiness should be scored honestly before a task moves forward, not discovered halfway through a build.

Score each candidate one to five on all three axes, then prioritise for high value, low effort, and high data readiness. Anything scoring well on value but poorly on data readiness isn't a bad idea, it's a data project first and an automation project second. Sequencing that correctly avoids a huge amount of wasted build time.

Telling a genuine opportunity from an AI novelty

Not every candidate that survives the scoring exercise is worth building, and it's worth being explicit about the difference between a genuine automation opportunity and something that's interesting mainly because it involves AI. The distinction isn't about the technology used, it's about whether the task has the underlying shape that makes automation pay off at all.

A genuine opportunity is repeatable at real volume, has clear and consistent inputs and outputs, and has a measurable cost attached to doing it manually today. It's the kind of task a new staff member would be trained to do the same way every time. An AI novelty, by contrast, tends to be a one-off or low-frequency task dressed up as a capability demo, has inputs that vary enough that no two runs look alike, or solves a problem nobody was actually measuring the cost of before the tool arrived. Summarising a single unusual document is a demo. Summarising the same category of document forty times a week, in a consistent format, for a person whose job is currently reading all forty, is an opportunity.

A simple test that holds up in most cases: if removing the manual version of this task tomorrow wouldn't be noticed by anyone within a week, it's a novelty. If removing it tomorrow would generate a support ticket, a missed deadline, or a customer complaint by Friday, it's real.

Where Microsoft 365 fits, once the ranking exists

The ranked list is what determines which Microsoft 365 capability is the right fit, not the other way around. The three tools sit at different points on the effort and structure spectrum, and matching a task to the wrong one is a common cause of projects that stall.

• Microsoft 365 Copilot fits tasks that are already embedded in a high-frequency, high-value workflow inside Word, Excel, Outlook or Teams, drafting a first pass of a recurring document, summarising a long email thread before a meeting, or pulling a first-cut analysis out of a spreadsheet. It's the lowest-effort option because it requires no build, only training and adoption discipline, but it depends entirely on a person choosing to use it well.
• Power Automate fits rule-based, structured tasks with a defined trigger and a defined outcome, moving data between two systems on a schedule or on an event, routing an approval, generating a standard notification. It suits tasks scoring high on data readiness and low on judgement, where the steps genuinely don't vary.
• Copilot Studio agents fit tasks that need to combine several steps, some judgement, and access to more than one data source, answering a structured customer question by checking order status and policy in the same interaction, or triaging an inbound request against a set of defined rules before it reaches a person. This is the highest-effort option of the three and should be reserved for candidates that scored well across all three axes, not the first idea that sounds impressive.

A common failure pattern is picking the most capable tool for every task regardless of fit, building a Copilot Studio agent for something a Power Automate flow would have handled at a fraction of the cost and maintenance burden. Matching the tool to the task, after the task has been properly scored, is what keeps the total cost of an AI programme proportionate to the value it returns.

Why a structured review beats guessing

The alternative to this method is what most businesses actually do: a senior person has a strong opinion about which department needs AI, a licence gets bought, and six months later usage has quietly dropped to near zero. That pattern is expensive in ways that don't show up on the invoice. Licence spend that doesn't get used is the smallest cost. The larger one is the change-management goodwill spent on a rollout that didn't land, which makes the next attempt, even a well-scoped one, harder to get staff behind.

A structured review avoids this by separating discovery from decision. The audit and scoring produce a ranked list before any tool gets chosen, which means the business is deciding what to build based on evidence rather than on whichever task happened to be top of mind when the budget got approved. It also produces a defensible answer to the question every finance lead eventually asks: why this task, and not one of the others.