Map internal processes that are suitable for automation using AI

The challenge is that most organizations don’t know which ones to prioritize. They have hunches—invoice processing, the shared email inbox, preparing the monthly “ reporting ”—but no hard data, so the decision ultimately comes down to whoever speaks up the loudest in meetings.

The Four Criteria That Make a Process Automateable

A process is a good candidate when it meets all four criteria. If even one is missing, the project will cost more than it brings in.

• It's repetitive. The same sequence, several times a week. A one-time monthly treatment never justifies automating it.
• It is documented, or can be documented on a single page. If no one knows how to write the rules, no machine will be able to apply them. A process that exists only in one person’s mind is a documentation project before it is an AI project.
• Its entries are structured—or are in the process of becoming so. A form, a file, or a message with fixed fields. A handwritten attachment that’s been scanned crookedly is a whole different story.
• The result holds true. There’s a simple way to verify that the output is correct. Without it, you’ll never know if the automation is working, and you’ll never let it run on its own.

A fifth criterion determines the timeline, not eligibility: someone must want the change. A technically flawless process that the team doesn't want yields no benefits.

Moving from Intuition to Data Collection

The most costly mistake is asking teams how much time they spend on a task. The answers are honest but inaccurate: no one measures their own time accurately.

In Power Automate , Microsoft provides a process mining capability that tackles the problem from the other end. It extracts event data from your management systems and reconstructs the process as it actually unfolds—not as it is described in a procedure. It visualizes workflows, compares variations, and identifies the root cause of bottlenecks.

From a management perspective, this changes the nature of the discussion: we have the number of runs, the actual duration, and the bottlenecks, rather than an estimate that everyone can interpret differently.

For processes that leave no trace in a system—such as manual processing, email, or a spreadsheet—the tally is done differently: a two-week count, kept by those performing the work, with a single column for each instance. The method is rudimentary, but it is more than sufficient to reach a conclusion.

Three categories of candidates, and what they entail

Sorting and routing. A shared inbox, a request form, a ticket queue. The agent reads, categorizes, routes, and suggests a response. This is the most cost-effective approach, because the volume is high and the rules are simple. It assumes that the routing categories already exist and are stable.

Data extraction and entry. Transferring information from a document to a system. Orders, invoices, registration forms. The benefit is proportional to the volume and consistency of the format. It requires output verification: an incorrect extraction entered into an accounting system costs more than the manual data entry it replaces.

Document preparation. Meeting minutes, summary notes, first drafts of sales proposals. In this category, the machine produces a starting point that someone then revises. The benefit is real but diffuse, and it cannot be measured in terms of jobs saved.

Note: The first two categories are handled by an agent built around a specific process. The third category is handled by a wizard built into the applications, which is already available in Microsoft 365. The costs and timeframes are not the same.

What Mapping Should Produce

One page per selected process, and no more:

1. Volume. How many visits per week, recorded rather than estimated.
2. Unit time. The duration of a cycle, from input to output.
3. The rule. What determines the outcome, written in short sentences.
4. Verification. How do we verify that the output is correct, and who does it?
5. The owner. The person responsible for the process who will make the final decision in borderline cases.

The first three lines indicate the potential profit. The last two determine whether the project is financially viable, because they specify the amount of human labor that remains—and that labor never disappears entirely.

Our Reading

Agents' arguments generally overlook one point that nevertheless has a significant impact on the calculation: human labor does not disappear; it simply shifts.

An agent who sorts a shared inbox does not delete the item. Instead, they remove the sorting and create a new task: verifying the agent’s decisions, processing what they set aside, and updating the rule when a new case arises. This task is shorter and more interesting than the one it replaces, but it takes time, and a company that builds its workload plan by treating it as zero will be off by half.

Our recommendation: Don’t choose the process based on what comes up in executive meetings. The process that gets brought up in that setting is almost always the one that causes the most frustration, and that frustration isn’t related to volume. The right candidate is often invisible from the top down: a 15-minute data entry task that someone does 40 times a week without ever complaining. That’s what the log is for, and that’s precisely why you shouldn’t skip it.

When not to proceed. If the process in question is undergoing a redesign, or if its rules change every quarter, leave it alone. You’d be automating a moving target, and the cost of maintaining the automation would end up being higher than the cost of manual work. Wait until the process stabilizes—or stabilize it first; that’s often the real project.

Finally, a note on tools. You can easily compile the data using a spreadsheet and two weeks of disciplined work. Process mining provides accuracy and speed when the data already exists in a system; it does not replace decision-making. Start with the method; purchase the tool only if the method reveals that you’re missing something.

What to Do

1. Choose up to three candidate processes, based on volume, not on how annoying they are.
2. Track them for two weeks before deciding on a tool. The tracking process takes a few hours and prevents months of errors.
3. Without hesitation, rule out any process for which no one can write the rules on a single page.
4. Assign an owner for each process before assigning a service provider.
5. Include the cost of the final inspection in the budget. This is the item that estimates often overlook.
6. Do not launch the second process until the first one is in production and has been evaluated.

The technical implementation of these automation systems is the least risky part of the project; what determines the outcome is the quality of the assessment and the clarity of the guidelines. Lambert Consulting conducts this assessment with the relevant teams before proposing anything to be built, Copilot Studio as well as Power Apps. And while the assistant built into Microsoft 365 is part of the scope, reorganizing access rights remains the prerequisite that must be addressed first.

Microsoft Sources

• An Overview of Process Mining in Power Automate
• Microsoft Documentation Copilot Studio
• Licenses and Billing Copilot Studio
• Documentation Power Automate