To find the right processes to automate, look for work that is frequent, rule-based, digital and costly when done by hand, then remove unnecessary steps and simplify the process before automating it. Score candidates on volume, rules, data quality, systems involved and risk, choose the simplest tool that fits, design for exceptions and measure results against a baseline before scaling.
Business process automation (BPA) is one of the most dependable ways to reduce costs, speed up service and free skilled people from repetitive work. It is also easy to get wrong. Organizations automate the wrong processes, automate broken processes, or pick tools before they understand the problem. This guide walks through a practical method for choosing what to automate, selecting the right approach and proving the value.
What is business process automation?
Business process automation is the use of software to carry out steps in a business process that people would otherwise do manually, such as moving data between systems, routing approvals, generating documents or sending notifications. The goal is not to automate everything. It is to let software handle the predictable parts of a process so people can focus on judgment, relationships and exceptions.
BPA is an umbrella term. Under it sit several different techniques, each suited to different situations:
| Approach | How it works | Best suited to | Watch out for |
|---|---|---|---|
| Built-in automation in existing software | Rules, workflows and triggers inside your CRM, ERP, help desk or accounting tools | Processes that live mostly inside one system | Features you already pay for but have not switched on |
| Integration and workflow platforms | Connect applications through APIs and trigger actions across them | Moving data and triggering steps between cloud applications | Sprawl of undocumented automations built by different people |
| Robotic process automation (RPA) | Software “bots” operate application user interfaces as a person would | Legacy or desktop applications that are hard to integrate directly | Fragility when screens or layouts change |
| Custom software | Purpose-built applications or services | Core, differentiating or high-volume processes that off-the-shelf tools cannot handle well | Higher upfront investment and ongoing maintenance |
| AI-assisted automation | Language models read, classify, extract or draft within a workflow | Steps involving unstructured inputs such as emails and documents | Variable accuracy; needs review and monitoring |
Microsoft’s own documentation illustrates the distinction between the integration and RPA approaches. It describes Power Automate cloud flows as automated workflows that “connect your apps and services,” triggered by events such as an email arriving or by a schedule. Its desktop flows, by contrast, bring robotic process automation to rule-based tasks on a workstation, including work with legacy applications, by interacting with user-interface elements, images or screen coordinates.
Eliminate and optimize before you automate
The most common automation mistake is making a bad process run faster. If a process includes approvals nobody needs, duplicate data entry or reports nobody reads, automating it locks that waste in place.
The US General Services Administration’s Federal EOA Playbook, published in June 2026, frames process improvement as three linked strategies:
- Eliminate “legacy activities, processes, requirements, and deliverables that are no longer needed.”
- Optimize by standardizing and streamlining what remains.
- Automate “processes, tasks, and activities that are manual, rule-based, and mundane.”
The playbook was written for federal agencies, but the logic applies to any business. It also warns that “one of the biggest pitfalls is rushing to solutions before understanding the problem.” In our approach, every automation project starts by asking whether a step should exist at all.
Step 1: Build a process inventory
Start broad. Ask each department to list the recurring processes and tasks that consume its time. Keep the descriptions specific and verb-led: “enter supplier invoices into the accounting system,” “create a customer account after a contract is signed,” “compile the weekly sales report,” “chase overdue timesheets.”
For each item, capture a few basic facts:
- How often it happens and in what volume.
- Roughly how long each instance takes and how many people are involved.
- Which systems, spreadsheets, emails or paper forms it touches.
- What goes wrong, how often and what that costs.
- Who owns the process today, if anyone.
Estimates are fine at this stage. The goal is a list wide enough to compare, not a precise business case for every line.
Step 2: Map the processes that matter
Pick the most promising ten or so candidates and map how they actually work, not how the procedure manual says they work. Sit with the people who do the work. Capture every step, handoff, system, decision and waiting period, and note where information is re-keyed or checked twice.
Pay particular attention to three things:
- Exceptions. How often does a case not follow the standard path, and what happens then? Exceptions are where automation projects run into trouble.
- Waiting time. In many processes, most of the elapsed time is spent waiting for someone to pick up the next step, not doing the work. Routing and notifications can shorten this dramatically without automating the work itself.
- Decision rules. Are decisions based on clear rules, or do they require situational judgment? Write down the rules people actually use, including the unwritten ones.
Mapping often reveals quick wins that need no technology at all, such as removing an approval step or agreeing on one standard form.
Step 3: Score and prioritize candidates
With the processes mapped, score each one. The GSA playbook suggests assessing automation suitability through questions about data structure, decisions and rules, available resources, systems and interfaces, and transaction volume. Adding business value and risk gives a practical scoring model:
| Criterion | Question to ask | Favors automation when… |
|---|---|---|
| Volume and frequency | How often does this run, and how much does volume vary through the year? | It runs daily or weekly at meaningful volume |
| Rules | Are the decisions rule-based, or do they need situational human judgment? | Most cases follow clear, stable rules |
| Data structure | Are inputs structured and digital, or unstructured and on paper? | Inputs are digital and consistent, or can be made so |
| Systems | How many systems are involved, and are they connected or controlled by others? | Systems offer APIs or built-in automation |
| Value | What would faster, more accurate handling be worth in time, cost, revenue or customer experience? | The gain is material and measurable |
| Risk | What happens if the automation gets a case wrong? | Errors are detectable and recoverable |
Score each criterion from one to five and plot candidates on a simple value-versus-effort grid. The best first projects are high in value, low in complexity and low in risk. They are often unglamorous: invoice intake, onboarding checklists, report compilation or request routing.
Step 4: Choose the simplest tool that fits
Once you know what to automate, choose how. A useful order of preference is:
- Use what you already own. Check whether your existing CRM, ERP, help desk, HR or accounting system can already do it through built-in workflows.
- Integrate through APIs. Where a process spans cloud applications with good APIs, an integration or workflow platform is usually more robust than screen-based automation.
- Use RPA for systems you cannot integrate. Screen-based automation is valuable for legacy or desktop applications without suitable interfaces, but plan for maintenance when screens change.
- Build custom software for core processes. When a process is central to how you compete, runs at high volume or needs logic that off-the-shelf tools handle poorly, purpose-built software can be the better long-term investment. Our software and product engineering team covers when that makes sense.
- Add AI where inputs are unstructured. Use language models for specific steps such as reading emails, extracting data from documents or classifying requests, inside a workflow that remains predictable.
Whatever tool you choose, consider who will maintain it. An automation that only one person understands becomes a risk the day that person leaves.
Work with GeekOzy
Put AI to work where it pays off.
Tell us what you're working on. We'll come back with honest thinking on a practical next step.
Start a projectExplore AI & AutomationWhere AI fits in process automation
Traditional automation needs structured data and clear rules. Generative AI extends automation to work that involves reading, writing and interpreting language. In a June 2023 analysis, McKinsey estimated that current generative AI and other technologies “have the potential to automate work activities that absorb 60 to 70 percent of employees’ time today.” That is an estimate of technical potential, not a forecast of what any particular business will achieve.
How value is captured matters more than the technology. McKinsey’s State of AI in 2026 survey found that nearly three-quarters of AI high performers report fundamentally redesigning workflows because of their AI use, compared with one-quarter of other respondents. In other words, the organizations getting the most from AI change the process, not just the tools.
For most businesses, the practical pattern is a predictable workflow with AI handling specific steps. Anthropic’s engineering guidance, Building effective agents, distinguishes these workflows, where models and tools follow “predefined code paths,” from agents that direct their own steps, and recommends starting with the simplest solution that works. Our guide to AI agents for business covers when a more autonomous approach is justified.
Step 5: Design for exceptions and people
Automations fail at the edges. Before building, decide what happens when:
- An input is missing, incomplete or in an unexpected format.
- A connected system is unavailable or returns an error.
- A case falls outside the rules or needs judgment.
- AI-extracted data has low confidence or fails validation.
The answer is usually an exception queue: cases the automation cannot complete go to a named person with the context already assembled, and the volume and type of exceptions are tracked. Monitoring exceptions also shows where rules need refining or where upstream data quality needs fixing.
Involve the people who do the work from the start. They know the exceptions, they will operate the new process and their confidence determines whether the automation is used or worked around. Be clear about how roles will change, and plan training for the new way of working.
Step 6: Pilot, measure and scale
Before switching anything on, record a baseline for the current process. Then run the automation on a subset of volume, a single team or a single location, and compare. The GSA playbook recommends tracking outcome, output and process goals and maintaining an accessible performance dashboard. For a business, useful measures include:
- Cycle time from start to finish of each case.
- Hours of manual work removed or redirected to higher-value tasks.
- Error and rework rates before and after.
- Exception rate and how it changes as rules improve.
- Customer or employee experience, such as response times or satisfaction.
- Running cost, including licenses, usage fees and maintenance time.
Scale only when the pilot meets its targets. Then document the automation, assign an owner, and add it to an inventory so that you always know which automations exist, what they touch and who is responsible for them.
Good first candidates by department
Every business is different, but the following processes are often worth assessing early because they tend to be frequent, rule-based and digital:
- Finance: invoice capture and matching, expense report checks, payment reminders and month-end report compilation.
- Sales: lead capture and assignment, CRM data updates, quote generation from templates and contract-to-onboarding handoffs.
- Customer service: ticket categorization and routing, order-status responses and follow-up surveys.
- HR and operations: employee onboarding and offboarding checklists, access requests, timesheet reminders and document collection.
- IT: account provisioning, password resets and routine monitoring alerts.
Treat this as a list to evaluate, not a list to automate automatically. The scoring step will show which of these matter most for you.
Common mistakes to avoid
- Automating before simplifying. Remove unnecessary steps first, or you will automate waste.
- Choosing the tool first. Starting with a platform and looking for uses leads to automations that are clever but not valuable.
- Ignoring exceptions. Designing only for the standard path leaves people firefighting the cases the automation drops.
- No baseline. Without measuring the current process, you cannot prove the improvement.
- No owner. Automations break when systems change. Each one needs someone responsible for keeping it working.
- Untracked sprawl. Many small, undocumented automations built by different teams can become hard to maintain and a security concern, especially when they rely on personal accounts or shared credentials.
The bottom line
The right processes to automate are frequent, rule-based, digital and valuable to improve, and they have been simplified before any tool is chosen. Build an inventory, map how work really happens, score candidates on volume, rules, data, systems, value and risk, pick the simplest tool that fits, design for exceptions and prove the results with a baseline and a pilot.
If you want help identifying and delivering your highest-value automation opportunities, explore our business automation solutions and AI & Automation services, or get in touch to discuss your processes.
Sources
- Overview of cloud flows — Microsoft Learn (Power Automate)
- Introduction to desktop flows — Microsoft Learn (Power Automate)
- Federal EOA Playbook: Federal Process Improvement through Elimination, Optimization, & Automation — US General Services Administration (June 2026)
- The economic potential of generative AI: The next productivity frontier — McKinsey (June 2023)
- The state of AI in 2026: On the road to ROI — McKinsey (August 2026)
- Building effective agents — Anthropic



