The most common question on a discovery call is some version of "could an agent do this?" The honest answer is usually yes. The useful answer is usually "it could, and it shouldn't." Across the automations and agents we have in production, roughly four out of five processes never use an AI model at all, and most of the rest use one for a single step.
This is a field guide to choosing between the three tools on the spectrum: rules, workflows, and agents. It is the framework we use in scoping.
Three tools, one spectrum
- Rules. "When X happens, do Y." A trigger, a few conditions, one or two actions. Rule-based, cheap, and boring in the best way. A new deal in HubSpot creates a folder, a Slack channel, and a kickoff task.
- Workflows. Several systems, branches, exceptions, and usually a person somewhere in the middle. Still rule-based. The logic is just bigger. Customer onboarding with two approvals and three main systems is a workflow.
- Agents. An AI model reads something, decides what to do, and acts through tools. It is the only tool of the three that can handle input nobody structured and decisions nobody wrote down.
The mistake is treating this as a ladder with agents at the top. It is a spectrum of how much judgment you hand over. Handed-over judgment has a cost: money, delay, inconsistent results, and tests that are harder to write.
The four questions we ask about every process
The answers usually settle the choice before anyone opens a design tool.
| Question | Points to rules or a workflow | Points to an agent |
|---|---|---|
| Is the input structured? | Form fields, data from other systems, spreadsheet rows | Emails, PDFs, call notes, screenshots |
| Is the decision written down? | A policy, a rate card, an if-then everyone agrees on | "It depends," answered differently by two senior people |
| What does a wrong answer cost? | Anything irreversible: payments, deletions, customer commitments | Drafts, triage, first passes a person will check |
| How often does the process change? | Quarterly or slower | Weekly, or every time a new kind of case shows up |
Score it honestly. If the first two answers point to rules, we build rules, however interesting the agent version would be. If they point to an agent and the third answer says a wrong answer is expensive, the agent gets an approval gate and never acts on its own.
Where agents earn their keep
Agents pay off when the input is free-form and there are too many different cases to list.
Halcyon's support queue is the clearest case we have. Around 1,400 tickets a week, no two phrased the same way, with answers scattered across a help center, billing records, and three years of past tickets. There is no rule that maps "my invoice looks wrong again" to the right response. An agent that looks things up in your own records now resolves 61% of those tickets without a person, cites the source for every answer, and hands the rest to a human with a summary attached. First response time went from six hours to under three minutes.
Notice what the agent is not allowed to do there. Refunds and account changes wait for a person. Reading, looking things up, and drafting are the agent's job. The irreversible parts are gated.
Where agents quietly lose
Three processes we are regularly asked to build as agents, and what we build instead.
- Recall reminders. Meridian Dental Group wanted fewer no-shows across twelve locations. The tempting version was a conversational agent that talks to patients. What cut no-shows by 31% in 90 days was a nightly update from the practice system, a messaging flow with reschedule links, and a morning list of open slots for each front desk. No AI model involved. Eight days from kickoff to live.
- Lead routing. "Route inbound leads to the right rep" sounds like judgment. It is a lookup table: territory, segment, round-robin. An agent adds delay and a small error rate to something a rule does perfectly. A model helps one step earlier, sorting a free-text "what do you need?" answer into the segment the table expects.
- Invoice matching. A three-way match between purchase order, receipt, and invoice is arithmetic. The model earns its place only on the exceptions: reading an invoice PDF whose line items do not map cleanly, then handing a structured guess to a person.
If you can write the decision down as an if statement, write the if statement. It never has an off day, it costs nothing to run, and you can test it in an afternoon.
A worked split: the quote pipeline
The clearest way to see the split is inside one workflow. Lumen Logistics' quote pipeline takes an emailed request and produces a draft quote for a dispatcher to approve. It has eleven steps. One of them is a model.
- Gmail trigger on the shared quotes inbox (rule)
- Skip threads already quoted (rule)
- Extract origin, destination, equipment, weight, and dates from the email body and any attached PDF (model)
- Check the extraction against the expected format. Send to a person if a required field is missing (rule)
- Look up locations and compute mileage (rule)
- Pull 180 days of route history from Postgres (rule)
- Apply the rate card and margin policy (rule)
- Check carrier availability in the transport system (TMS) (rule)
- Write the draft quote into the TMS (rule)
- Post an approval card to the dispatcher (gate)
- On approval, reply on the customer's thread (rule)
Step three is the only step where the input is unstructured. Everything after it is a rate card and a margin policy the dispatch team already had on paper. Putting a model in step seven would have made the quote slower, more expensive, and impossible to explain to a customer asking why the price moved. The result was a median time to quote of nine minutes, down from four hours, and 14 hours a week back per dispatcher.
The rule of thumb
Start with the rule-based version. Add a model to exactly one step when all three of these are true:
- The step's input is text, documents, or images that no upstream system will structure for you.
- A rule can check the step's output before anything acts on it.
- You have at least 100 past examples to test against before launch, and a way to keep collecting them.
Escalate from a single model step to a full agent only when the process has too many branches to draw. If it fits on one page, it is a workflow. If the page has more than a dozen "it depends" boxes and they keep multiplying, it may be time for an agent, with gates around every action that costs money or touches a customer.
What this means for the budget
Cost follows the same spectrum. A rule-based run costs fractions of a cent and finishes in under a second. A workflow with one model step, like the quote pipeline above, costs a few cents per run and finishes in seconds to minutes. A full agent, which may make eight to twenty model calls per task, costs anywhere from a few cents to a few dollars each, takes longer, and needs an evaluation set (past cases with known answers to test against), step-by-step logs, and someone watching the approval rate.
Spend agent money on agent problems. The teams getting the most out of AI right now have already finished their rule-based plumbing, so the model only does the part nobody else could.
If you are unsure which side of the line a process falls on, that is a good discovery call. We will tell you if the answer is a rule, and we will be happy about it.