Which Workflows to Automate First With AI Agents

Pick the first workflow by score, not by excitement. The best first workflow is high volume, mostly rule-driven, easy to undo, low on sensitive data, and owned by someone who will show up.

You can score your candidates in one afternoon with a spreadsheet. This post gives you the sheet.

Your first choice matters more than your model choice. A good workflow makes a mediocre agent look useful. A bad workflow makes a great agent look broken.

Why the first choice matters

Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027. Its stated reasons are rising costs, unclear business value, and weak risk controls. None of those is a model problem. All three trace back to what you pointed the agent at.

The MIT NANDA report, as covered by Fortune in August 2025, found the same thing from another angle. It put the biggest returns in back-office automation. Yet more than half of generative AI budgets went to sales and marketing tools. Teams chose what sounded exciting, not what paid back. Treat the 95% headline number with care: the authors describe it as preliminary research, not a universal failure rate.

The pattern is common. Teams pick the flashy workflow, hit exceptions on day three, and lose trust before they learn anything. A boring first workflow gives you a clean win and real data. If your pilots keep stalling, why AI pilots fail covers the wider pattern.

Five criteria and how to score them

Score each candidate workflow from 1 to 5 on each criterion. Five means “good for a first agent.” One means “bad for a first agent.” The total is out of 25.

Criterion	Question	1 (bad)	5 (good)
Volume	How many cases per week?	Under 10	500 or more
Exception rate	What share of cases need human judgment?	Over 50%	Under 10%
Reversibility	If the agent is wrong, can you undo it?	Money or records leave your control	Easy to undo, nothing leaves
Data sensitivity	What data does the agent touch?	Regulated or customer-private	Internal, low risk
Owner availability	Will one named person review output weekly?	Nobody owns it	Owner commits hours each week

Here is why each one is on the sheet.

Volume. Agents pay back on repeated work. The best-known field result is the Brynjolfsson, Li, and Raymond study of 5,179 customer support agents. It found a 14% average productivity gain, and 34% for novice workers. Support is high-volume, repeating work. Low-volume work cannot repay the build cost, however clever the agent is.

Exception rate. This is the share of cases that do not follow the normal path. The lower it is, the more cases the agent can finish alone. High exception rates are where pilots drown. I cover how to count them below.

Reversibility. OpenAI’s agent guide tells teams to rate each tool as low, medium, or high risk. Its factors include read-only versus write access, reversibility, permissions, and financial impact. I’d apply the same test to the whole workflow. Drafting a reply that a person sends is reversible. Wiring money is not.

Data sensitivity. Start where a mistake does not create a legal event. Regulated data can come later, with the controls in place. The enterprise AI security checklist lists those controls.

Owner availability. This is the one teams skip. An agent needs someone who knows what “correct” means and who reviews its output every week. Without that person, nobody can tell you whether the pilot worked.

Two rules before you add up the score

Any 1 on reversibility or data sensitivity disqualifies the workflow for a first pilot. The total does not matter. Defer it.
Any 1 on owner availability means stop. Find an owner first, or pick something else.

What score makes a workflow a good candidate?

After the two rules, use these bands. They are my rule of thumb, not an industry standard.

20 to 25: Strong first candidate. Brief it this week.
15 to 19: Workable, but fix the weakest criterion first.
Under 15: Wait. Use it as a later project.

A worked example with made-up numbers

Say a 60-person operations team has four candidate workflows. These numbers are invented to show the method. Yours will differ.

Workflow	Volume	Exceptions	Reversibility	Sensitivity	Owner	Total
Triage and route the shared support inbox	5	4	5	4	4	22
Code incoming invoices to GL accounts	4	3	3	4	4	18
Review vendor contracts for risk	2	2	3	2	3	12
Approve customer refunds over $500	3	2	1	3	3	12

The inbox triage wins at 22. Volume is high, a wrong route is easy to fix, and the data is low risk. Invoice coding scores 18 and is a good second project once someone fixes the owner and exception gaps. Contract review is low volume and uses sensitive text, so it waits. Refund approval hits a 1 on reversibility, so the rule knocks it out whatever its total says.

Notice what happened. The two “impressive” workflows lost. That is normal.

One caution on the inbox example. Triage is a good first job because the agent routes and drafts. A person still sends and approves. If you want to design that handoff, read the human-in-the-loop glossary entry.

Workflows that should wait

Some work is a poor first pilot even when someone on the team badly wants it. I’d hold off on these:

Anything that moves money or changes a system of record without review. Payments, refunds, and payroll changes fail the reversibility test.
Regulated decisions. Credit, hiring, claims denials, and clinical calls need governance you probably have not built yet.
Low-volume strategic work. Board memos, pricing strategy, and negotiations do not repeat enough to measure.
Workflows with no written process. If three people do it three ways, you do not have a workflow. You have habits. Write down the process first.
Anything with a fuzzy success test. Anthropic’s guidance points to support and coding as natural fits because success is measurable. Support has resolutions. Code has tests. If you cannot say how you would check the output, pick something else.
Work a simple script already handles. OpenAI’s guide says that if a use case does not show complex decisions, hard-to-maintain rules, or heavy unstructured data, a deterministic solution may suffice. Anthropic says to find the simplest solution possible. Not every task needs an agent.

That last point is a real filter. If your exception rate is near zero and the inputs are clean, use a plain automation. Agents earn their cost where judgment, messy documents, or sprawling rules make scripts break.

How to measure exception rate fast

Do not guess. Count. You can do this in two or three hours.

Pull a sample. Take 50 to 100 recent cases from the real workflow. Use a spread across weeks, not just the easy ones. Fifty is enough for a rough number. I’d call it directional, not precise.
Sort each case into three bins.
- Clean: followed the normal path, no human thinking needed.
- Known exception: off-path, but the fix follows a rule someone can state.
- Novel exception: needed real judgment or missing information.
Calculate two numbers. Exception rate is (known + novel) divided by total. Novel rate is novel divided by total.
Read the novel pile. Ask the owner what they did in each case and why. Write down the reasoning. This is the best requirements document you will get.

How to read the result: a high known-exception rate is fine, because those are rules the agent can follow. A high novel rate is the warning sign. If more than about a quarter of cases are novel, the workflow is too unpredictable for a first pilot. That 25% line is my judgment call, not a researched threshold.

Do this sample before you build anything. It also gives you the baseline you will need to prove the pilot worked. My back-office ROI post shows how to turn that baseline into a savings number.

Who must be in the room?

Keep the group small. You need four people:

The process owner. They know what correct looks like and approve the scope.
A hands-on operator. Someone who does the work daily and can explain the odd cases.
A data or systems contact. They confirm what the agent can read and write.
A risk or security reviewer. They sign off on the data and the reversibility call.

Add the budget holder for the final decision. More people than that slows you down.

Pick, then write the one-page brief

Take the top scorer and write a one-page brief before any build starts. Use this template:

Workflow: <name>
Owner: <name, hours per week committed>
Baseline: <cases per week, minutes per case, exception rate, error rate>
What the agent does: <reads X, drafts Y, routes Z>
What a human still does: <approves, sends, handles novel cases>
Out of scope: <list>
Hero metric: <one number, with target>
Stop rule: <what result ends the pilot>

Keep it to one page. If you cannot fill in a line, you found a gap that would have hurt you in week three.

The hero metric is one number, such as minutes per case or percent resolved without a human. Pick it now, not after the results come in. For how this fits with a short pilot, see my one-week AI pilot methodology. To understand where this stage sits, read AI POC vs pilot vs production.

What to do with the runner-ups

Do not throw the other scores away. Keep the sheet.

Rank them. Your second and third picks are your roadmap.
Name the blocker for each. Maybe invoice coding needs an owner. Maybe contract review needs a clean data source. Write the blocker next to the score.
Re-score after the pilot. You will know more about your exception rates, your data, and your reviewers. Scores move.
Reuse the plumbing. The second agent is cheaper when it shares access, logging, and review steps with the first. Gartner expects 40% of enterprise apps to include task-specific agents by the end of 2026, up from under 5% in 2025. Narrow, specific agents are where the market is heading, so a sequence of small wins fits.

A final check on vendors. Gartner also estimates that only about 130 of the thousands of agentic AI vendors are real, and calls the rest “agent washing.” When a vendor says their product fits any workflow, ask them to run your top-scoring workflow with your sample data.

Bottom Line

Score your candidates on volume, exceptions, reversibility, data sensitivity, and owner availability.

Disqualify anything irreversible or sensitive for the first pilot. Count exceptions from a real sample of 50 to 100 cases.

Write the one-page brief for the winner. Keep the runner-ups on the sheet for next.

Next step: block three hours this week, pull the sample, and fill in the table.

FAQ

How do I pick the first workflow? Score every candidate from 1 to 5 on volume, exception rate, reversibility, data sensitivity, and owner availability. Drop any workflow that scores 1 on reversibility, sensitivity, or owner. Pick the highest total from what is left.

What score makes a workflow a good candidate? I treat 20 or more out of 25 as a strong first candidate and 15 to 19 as workable with one fix. Below 15, wait. These bands are my rule of thumb, so adjust them to your risk appetite.

Which workflows should wait? Wait on anything that moves money without review, regulated decisions, low-volume strategic work, undocumented processes, and tasks with no clear success test. Also wait on work a plain script already handles well.

How do I count exceptions before building? Pull 50 to 100 recent real cases and sort each as clean, known exception, or novel exception. Divide to get the exception rate and the novel rate. A high novel rate means the workflow is too unpredictable for a first pilot.

Who must be in the room? You need the process owner, a daily operator, a systems or data contact, and a risk reviewer. Add the budget holder for the final call. Without a named owner who reviews output every week, do not start.

Should my first AI agent be customer-facing? Usually not. Internal, high-volume, reversible work gives you a safer first test. A person should review anything that reaches a customer until you have measured accuracy.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect