Guide · Decide
When to use an AI agent: sort every step into four buckets
Before you give a workflow to an AI agent, sort every step: delete it, run it as plain code, give it to an agent, or keep a person on it. A decision tree, real data and a worked example.
The short answer
Before you give any step of a workflow to an AI agent, sort it into one of four buckets. Delete it if it shouldn’t exist after AI. Run it as plain code if it follows rules. Give it to an agent if it needs judgment you can check. Keep a person on it if a mistake would be costly or hard to undo. In most workflows, only a few steps end up with an agent.
Key takeaways
- Most steps don’t need an agent. Many should be deleted, and many are simple rules.
- An agent earns a step when the step needs judgment and you have past cases to check its work.
- Keep people on the riskiest decisions, like approvals, payments and signatures, and make each approval specific.
What are the four buckets?
An agent is the most flexible tool you have, and the least predictable. So before you hand any workflow to one, sort each step into one of four buckets:
Waits, re-keying, chasing and redoing work. Remove it before you automate anything.
- Copy totals into slides
- Re-export with the right filter
- Wait for a reply
If this, then that. Cheaper, faster and more predictable than a model.
- Pull last week’s deals
- Send the deck at 11:00
- Create the record
Messy inputs, many variations, and enough past cases to test against.
- Read account emails
- Reconcile conflicting figures
- Classify an invoice line
Approvals, negotiation, signatures and payments. An agent can prepare the work.
- Approve the discount
- Sign the contract
- Submit the payment
Forward-deployed engineering teams such as Varick Agents work this way, and its CEO, Vasuman Moza, walked through the method with Greg Isenberg on The Startup Ideas Podcast. Applied to real workflows, it tends to surprise people: far fewer steps need an agent than they expect.
How do you decide which bucket a step belongs in?
Ask four questions, in order. The first answer that fits decides the bucket.
Does this step still need to exist once the work is redesigned?
If no: DeleteOtherwise, next question
Does it follow fixed rules, with no judgment?
If yes: Plain codeOtherwise, next question
Would a mistake be costly, public or hard to undo?
If yes: HumanAn agent can prepare the work for the person.Otherwise, next question
Do you have past cases to check an agent’s work against?
If yes: AgentIf not, keep a person on it while you collect cases.
Two details matter. First, ask the questions about the step as it should run, not as it runs today: a step that only exists because of a broken handoff gets deleted. Second, sort with more than one person in the room. Someone can describe a workflow confidently and still be wrong.
Why delete comes first
The biggest gains usually come from removing work, not automating it. In 1990, Michael Hammer argued in the Harvard Business Review that using technology to speed up an old process rarely pays off, because the time sits in handoffs, waiting and rework, not in the steps themselves. His advice fit in three words: “Don’t automate, obliterate.”
It still holds. Re-exporting a report, chasing a missing update, waiting for a reply and merging six versions of a deck aren’t work an agent should do faster. They are symptoms of a process that should be redesigned.
Share of a client’s invoices that went straight through, without manual handling, after Varick Agents redesigned its accounts-payable process. Cost per invoice fell from $31 to $6.
Steps in the same accounts-payable process after the redesign. Cycle time fell from 24 days to 6, and exception loops from six to one.
These are one firm’s reported results for one client, not a benchmark. But the pattern is common: most of the improvement came from fixing the process, and only part of it from agents.
When is plain code better than an agent?
Whenever a step follows rules. Plain code is faster and cheaper, and it does the same thing every time, which makes it easy to test and trust. Most business automation is exactly this kind of work.
Show the data
| What the automation does | Share of patterns |
|---|---|
| Capturing a lead | 29.9% |
| Sending a notification | 19.2% |
| Creating a record | 17.6% |
| Updating a record | 14.4% |
| Everything else | 18.9% |
In Zapier’s data, most automations wait for something to happen, like data changing or a form being submitted, and the median one touches three apps. What gets automated first is the handoff between tools. AI earns a step only where judgment is needed, such as sorting a messy inbound email.
Production systems built this way look almost boring. Daniel Kornum of Varick Agents writes that the ones it ships end up as five to ten deterministic steps with one or two model calls, roughly “85% code and 15% LLM.”
When does a step need an agent?
A step belongs with an agent when two things are true: it needs judgment, and you can check the result.
- Judgment: the inputs are messy or vary a lot, like free-text emails, scanned invoices or spreadsheets that disagree.
- Checkable: you have past cases with known right answers, so you can test the agent before trusting it, and keep testing it after.
Building a sales quote is a good example. The agent has to weigh who the customer is, what the deal is worth and what resources to commit, which rules alone can’t capture. But the result can be checked against past deals.
Where you can, run the agent inside the systems people already use, such as the CRM, the ERP or Slack, instead of asking them to adopt a new tool. It is faster to adopt and easier to govern.
When should a person stay in the loop?
Keep a person on any step where a mistake would be costly, public or hard to undo: approving a discount, negotiating, signing a contract, submitting a payment. An agent can still prepare the work. Without that person, an agent that approves a convincing fake invoice pays a fraudster.
Make each approval specific. A person clicking “approve” on everything isn’t a control. The reviewer should see what has already been checked, what is still uncertain, and exactly which decision they are making, and every action should land in an audit trail.
Over time, people can step back. As your evals show an agent is reliable on a step, you can move from approving every case to reviewing a sample.
Worked example: a weekly revenue review
Here is the method applied to one workflow: the weekly revenue review from Nolli’s product demo. The team’s process document lists five steps. Watching four real Mondays turns up 20, including waits, re-exports and four redone steps.
- Ask reps for missing updates
- Re-export with the right filter
- Wait on Finance
- Fix a broken formula
- Reconcile again after edits
- Copy totals into slides by hand
- Ask everyone to update their slides
- Merge six versions of the deck
- Rebuild slides after late changes
- Export closed-won deals
- Update the forecast workbook
- Flag renewals past their date
- Build the deck from the template
- Send to leadership at 11:00
- Collect account updates from email and Slack
- Reconcile conflicting figures
- Flag mismatches with Salesforce
- Draft notes for each account
- Decide renewal exceptions
- Write and sign off the recommendation
The new workflow:Gather→Reconcile→Build deck→Review→Send
Show the data
| Bucket | Steps |
|---|---|
| Delete | 9 |
| Plain code | 5 |
| Agent | 4 |
| Human | 2 |
Nine of the 20 steps disappear once the work is redesigned. Five become plain code. Only four need an agent, and two decisions stay with the revenue ops lead. The result is a five-step workflow, and the lead’s Monday drops from two and a half hours to about 40 minutes.
Questions people ask
What is the difference between automation and an AI agent?
Automation follows rules someone wrote in advance: the same input always gives the same output. An AI agent uses a model to make judgment calls, like reading a messy email or deciding which category an invoice line belongs to. Use automation wherever rules work, and save agents for judgment.
How many steps of a workflow should be agentic?
Usually fewer than people expect. Varick Agents reports that the production systems it ships end up as five to ten deterministic steps with one or two model calls.
Can a step move between buckets?
Yes. As evals show an agent is reliable, a person can move from approving every case to reviewing samples. As a task becomes standard, an agent step can often become plain code.
Who should sort the steps?
The people who do the work, plus someone who knows the systems. Talk to more than one person, because one person can tell you the wrong story.
Is human approval enough to make an agent safe?
Only if the approval is specific. The reviewer should know what has already been checked, what is still uncertain, and exactly which decision they are being asked to make.
Sources
- Greg Isenberg, with Vasuman Moza of Varick Agents. “Become a $1M/yr FDE (Full Course).” The Startup Ideas Podcast, October 1, 2026.
- Daniel Kornum. “If AI is so great, why isn’t it working?” Varick Agents.
- Michael Hammer. “Reengineering Work: Don’t Automate, Obliterate.” Harvard Business Review, July–August 1990.
- Tim Hey for Zapier. “The most common business automation use cases in 2026, according to 20 million live workflows.” Stacker, via AOL, September 21, 2026.