Automate first the processes that happen often, take real time, follow rules a new hire could learn, and touch systems software can reach, ideally where today's errors already cost you something. To identify processes to automate with AI, score each candidate on volume, time per instance, error cost, how much judgment it needs, and how accessible its data is, then plot value against feasibility. The winner is usually unglamorous: invoice entry, enquiry triage, status updates, document chasing.

This post gives you a two-hour mapping workshop you can run with your team leads, a scoring matrix, a method for spotting the hidden exception handling that wrecks automation projects, and a prioritization grid. There's a worked example and a template table you can copy.

Why does the first process you automate with AI matter so much?

Your first AI workflow sets the tone for everything after it. A good first choice produces a measurable win in weeks, builds trust with your team, and creates reusable pieces (integrations, logging, review habits) for the next one. A bad choice, usually the most exciting or most visible idea rather than the most suitable, stalls in testing and teaches everyone that AI "doesn't work here".

The selection mistakes are predictable:

  • Picking the process with the loudest champion instead of the best fit
  • Picking a process that runs a few times a month (not enough volume to pay back)
  • Picking a process that is mostly judgment (AI assists, but can't carry it)
  • Picking a process whose data lives in someone's head or a system with no way in

A structured scoring exercise avoids all four.

How do you run a two-hour process mapping workshop?

Invite the people who actually do the work, not just their managers. Three to six people is ideal.

Part 1: List (30 minutes). Each person writes down the repetitive tasks that fill their week, one per sticky note or spreadsheet row. Prompt them with questions:

  • What do you do every day that follows the same pattern?
  • What do you copy from one system into another?
  • What do you chase people for?
  • What do you dread at month-end?
  • What would you hand to a new assistant first?

Aim for 15-30 tasks. Merge duplicates.

Part 2: Describe (40 minutes). For the top ten or so, capture in a sentence each: what triggers it, what the inputs are, which systems are involved, what the output is, and roughly how long it takes.

Part 3: Score (30 minutes). Score each one on the matrix below. Do it as a group; disagreements are useful because they usually reveal hidden complexity.

Part 4: Choose (20 minutes). Plot the scores on the prioritization grid, pick a first candidate and a backup, and agree who will measure the baseline.

How do you score each process?

Score each factor from 1 to 5. The first three measure value; the last three measure feasibility.

Factor1 means5 means
VolumeA few times a monthMany times a day
Time per instanceUnder a minute20+ minutes of effort
Error painMistakes are rare and cheapMistakes are common or costly today
Rule clarity (inverse of judgment)Mostly expert judgment, every case differentA new hire could follow written rules
Data accessInputs on paper or in people's heads; systems with no APIInputs digital; systems have APIs or integrations
Exception rateMost cases are unusualMost cases follow the normal path

Value score = Volume + Time + Error pain (3-15) Feasibility score = Rule clarity + Data access + Exception rate (3-15)

You can weight the factors if one matters more to you (for example, double Volume if you're growing fast), but keep the scoring simple enough that the team trusts it.

How do you spot hidden exception handling?

This is where most automation estimates go wrong. A process that sounds like "read the order, enter it into the ERP" often contains a dozen unwritten rules the person applies without thinking. Those rules are the cases the automation will get wrong.

Listen for these phrases during the workshop:

  • "Usually..." ("Usually the PO number is at the top.") What happens when it isn't?
  • "Except when..." ("Except when it's from the hospital group, they need a different form.")
  • "I just know." That's an undocumented rule. Ask them to explain how they know.
  • "I check with..." Every "ask someone" moment is a decision point the AI needs a rule or an escalation for.
  • Side spreadsheets and sticky notes. If someone keeps a personal list of customer quirks, that list is part of the process.

Then test it. Pull 20 recent real instances of the candidate process and count how many followed the normal path without any special handling. If 18 did, your exception rate is low and the process is a strong candidate. If 8 did, either the process needs standardizing first or the automation needs to be designed mainly as a triage-and-escalate system.

What does the prioritization grid look like?

Plot each process by its two scores:

Lower feasibility (3-9)Higher feasibility (10-15)
Higher value (10-15)Plan it. Worth doing, but fix access or standardize the process first.Start here. Your first AI workflow.
Lower value (3-9)Park it. Revisit next year.Quick win. Use an off-the-shelf tool or a simple automation; don't custom-build.

Your first project comes from the top-right box. If that box is empty, your first project is groundwork: opening system access or documenting a high-value process so it moves right. Our 20-point AI readiness test helps diagnose what's holding feasibility down.

What does a worked example look like?

For illustration, here's how a property management company with around 40 staff might score six candidates. The numbers are invented to show the method.

ProcessVolTimeErrorValueRulesAccessExceptionsFeasibilityBox
Tenant maintenance request intake and triage5341244311Start here
Contractor invoice entry and matching to work orders4441243411Start here
Rent arrears reminders323844412Quick win
Monthly owner statements commentary253103238Plan it
Lease renewal negotiations244101326Plan it (assist only)
Ad hoc owner queries33282226Park it

Two candidates tie at the top. The tiebreaker should be which one has an owner with time to review it, and which one creates reusable pieces. Maintenance intake touches the tenant channel and the work order system, both of which later workflows will need, so it might go first, with contractor invoices next. Lease negotiations score high on value but low on rule clarity: AI can prepare comparables and draft letters, but a person runs the negotiation. Our guide to AI for property management goes deeper on that industry.

What should you capture for the chosen process?

Before anyone builds anything, document the winner on one page:

FieldWhat to write
TriggerWhat starts it (email, form, schedule, message)
InputsFormats and sources, with examples of each
StepsNumbered, including lookups and decisions
RulesWritten-down business rules, including the unwritten ones you found
ExceptionsThe common deviations and how each should be handled
SystemsEvery system read from or written to, and how access works
OutputWhat "done" looks like, and where it lands
CheckpointWhere a person reviews, and what triggers a flag
BaselineVolume, handling time, error rate, cycle time today
OwnerThe person accountable for the workflow's results

That page becomes the brief for the build. Our guide to turning a manual process into an AI-run workflow picks up from here, and if the first candidate is high-value but uncertain, scoping an AI proof of concept shows how to test it cheaply before committing.

Which processes are usually good first candidates for AI automation?

Across industries, the same shapes keep scoring well:

  • Document intake to system entry: invoices, purchase orders, applications, claims. Our guide to AI data entry automation covers this category.
  • Inbound triage: sorting and routing emails, tickets or enquiries, with drafted first responses.
  • Status questions: "where's my order / job / application?" answered from your systems.
  • Chasing: collecting missing documents, confirmations or payments with polite, persistent follow-up.
  • Recurring summaries: daily or weekly reports compiled from several systems.

And the same shapes usually score poorly as first projects: negotiations, anything involving a decision about a person's job, credit or health, low-volume expert work, and processes nobody can describe consistently. For context on all of this, see our no-hype guide to AI for business.

Frequently Asked Questions

How do I identify which processes to automate with AI?

List repetitive tasks with the people who do them, then score each on volume, time per instance, current error pain, rule clarity, data access and exception rate. Start with the process that scores high on both value and feasibility, not the one with the most enthusiasm behind it.

What makes a process a bad candidate for AI automation?

Low volume, heavy reliance on expert judgment, inputs that aren't digital, systems with no way to connect, and a high proportion of unusual cases. Processes with legal, financial or employment consequences for individuals also need a person making the final decision.

Should we fix a process before automating it?

If different people handle it differently and nobody can write down the rules, yes. Automating an inconsistent process makes the inconsistency faster. Standardize it first, then automate the standard version.

How long does process mapping take?

A first pass takes about two hours with the right people in the room. Documenting the chosen process in detail and sampling real instances to check the exception rate usually takes another few days, spread across a week or two alongside the baseline measurement.

Can AI help with the process mapping itself?

Somewhat. AI can summarize workshop notes, draft process documents from recorded walkthroughs and analyze exported logs for patterns. But the unwritten rules come from talking to the people who do the work, and that part stays human.

Want help choosing your first process?

If you'd like someone to run the workshop with your team or sanity-check your scores, tell us about the processes on your list. We'll tell you which one we'd start with and why, even if that means a simple tool rather than a build. Here's how we approach AI automation.