Every few weeks someone asks a version of the same question: we've got a task that runs 50 times a month — does it need an AI agent, or is a single prompt enough? The usual frame is about complexity. How many steps, how much reasoning, does it need tool access. But complexity isn't what kills these builds. The short answer: before you decide, ask whether you can write a verification check — a rule that catches a bad output without a human reading it. If yes, automation is viable. If not, you're building a draft factory.
The question most people ask first
The standard checklist sounds sensible. Is the task repetitive? Check. Does it involve reading something and extracting structured data? Check. Do multiple steps depend on each other? Sometimes. Tick enough boxes, build the agent.
That checklist is fine for scoping. It does nothing to tell you whether automation will hold up once real data starts flowing through it.
Here's a scenario I've seen more than once: an invoice extraction agent reads PDFs, pulls totals, and writes rows to a spreadsheet. It works on 95 of 100 test invoices. It ships. Six weeks later the bookkeeper notices a discrepancy. The agent parsed "$1,234.00" as the integer 1234 on invoices with a particular layout. Every row passed validation — the field exists, it's a number, it's not null. The agent ran clean. The data was wrong, and nobody caught it for a month and a half.
Can you write the check?
Before you decide whether to build an agent, answer this: can you write a rule that catches a bad output?
Not a perfect rule. Not one that catches every failure mode. Just something definitive — a total outside a plausible range, an invoice number that doesn't match the filename format, a required field that came back empty, a response that fails a schema. If you can write that check, you have what you need to build safely. The check doesn't catch everything. It catches enough that you'll know when things are going wrong, rather than finding out from a client.
If you can't write any check, you're not automating a process. You're generating volume that someone still needs to read. Which brings up the cost question people get backwards.
If nobody's going to read every output, someone needs to write the rule that catches every bad one.
The cost isn't the tokens
People worry about API costs when they're evaluating whether to automate. That concern is almost always the wrong thing to optimize on.
Claude Haiku 4.5 runs at $0.25 per million input tokens — we've already done the math on three real workflows and most small business agents land between $20 and $60 a month in API fees. That's not the risk.
The risk is review labor. Say your agent runs 200 times a month and each output takes five minutes to eyeball. That's 17 hours of monthly review work. At a $30/hour admin rate, you're over $500/month in hidden cost — for a workflow you built to save time. Human review time scales linearly with volume. Tokens don't. The thing that kills agent ROI isn't the model bill; it's what happens when no one writes the verification check.
This applies to content automation too. An agent that writes 150 social posts a month works if "good" has an enforced definition — a character limit, a required keyword, a brand voice checklist you can run programmatically. If "good" means "I'll know it when I see it," you've built a creative tool, not automation. Nothing wrong with that, but they're different things with different cost profiles.
The token bill almost never kills these. Human review time on unverifiable output does — and it scales with the thing you were trying to scale.
When a single prompt is the right call
A one-shot prompt with you in the loop is the right tool when you're reading the output directly anyway — a draft email, a summary of a call, a first cut at a scope document. If the task happens rarely, or output quality is inherently subjective and you're the judge regardless, there's no reason to add agent infrastructure.
High-frequency tasks can also stay at the prompt level when you're sampling the outputs yourself. Ten subject line options in your email client — you pick the one you want. No verification check needed because you are the verification check.
When an agent earns its setup cost
Build the agent when the task repeats at volume (50+ times a month is a reasonable starting threshold), when output touches another system — a CRM record, a calendar event, a database row, an outbound email — and when you can write the check.
That last condition matters more than the first two. An agent that writes to a real system without a verification step is just moving the review problem downstream, closer to a client or a live transaction. Approval gates solve this for high-stakes steps; for lower-stakes ones, the question is whether first failures are recoverable. Start agents on tasks where undoing a mistake takes five minutes, not five hours.
When clients come to us wanting to automate something that doesn't have a clear success test yet, the first conversation is always about what "wrong" looks like before we write a single node. That's not a delay. It's the design.
If you're working through this for a specific workflow and want a second opinion, drop us a line. We'll tell you honestly whether the verification step you're thinking of is enough, or whether you're about to build something that needs a lot of babysitting.
— Cole
Sources
- AI Agent vs. Single Prompt: When Automation Actually Pays Off — dev.to community discussion (August 2026)
- What an AI Agent Actually Costs to Run (With Real Numbers) — Genesis Web Digital (June 2026)