Do not index
Which of your agency workflows are actually worth handing to an AI agent now? That question changed on June 30, and most agency owners have not re-run the math. Here is the short answer. The blocker was never whether agents work. It was cost per run and reliability, and both just moved. Anthropic shipped Claude Sonnet 5 at roughly 40% below flagship pricing for near-flagship agent capability. "It is now $2 per million input tokens and $10 per million output tokens," according to Asanify's AI news deep dive. If you priced agent automation six months ago and walked away, your numbers are stale.
The reliability side moved too. A Zapier tester ran a two-part Salesforce task the model completed end to end and summed it up in five words. "That used to stall halfway," according to the same Asanify report. Stalling halfway is exactly the failure mode that kept serious operators from trusting agents with real client work. Half-finished automation is worse than none, because someone has to notice it, diagnose it, and redo it.
This is for agency owners between $200k and $2M in revenue, ghostwriters charging $5k to $30k per month, and 3 person teams where the founder still touches every deliverable. At that size the constraint is not demand. It is the founder's hours, and deciding which hours to buy back is a real financial decision, not a tooling hobby.
This is not for teams that have never documented a workflow, because you cannot automate a process that exists only in your head. Skip this if you are hoping an agent will handle client judgment, voice, or strategy, because handing those to a model is how retainers die. And if you are a solo operator with two clients and spare capacity, the arithmetic will not favor you yet. Automation pays when repetition is expensive, and yours is not expensive yet.
How to run the Cost-Per-Run Sort
I use what I call the Cost-Per-Run Sort to decide what gets automated in my own content operation. List every workflow your team repeats weekly. For each one, write down two numbers: the hours it eats per month and how tolerant it is of an imperfect first pass. Then sort. Workflows that are high-repetition and high-tolerance go first. Think client reporting drafts, research and news monitoring, onboarding checklists, transcript cleanup, and content repurposing. Workflows that are low-tolerance stay human no matter how cheap the tokens get. Strategy calls, client voice, pricing conversations, and anything a client would fire you over belong to people.
The sort matters because the new pricing changes which rows clear the bar, not whether the bar exists. At $2 and $10 per million tokens, a research feed that runs daily costs less per month than one billable hour. A reporting workflow that eats six hours a month at a $100 effective rate is a $1,800 per quarter problem you can now solve for the price of lunch. Six months ago the same automation was a line item you had to justify. Now the justification runs the other direction. You explain why a high-repetition, high-tolerance workflow is still manual.
One rule from the Asanify piece mirrors how I run my own systems: write the guardrail before the automation. Every agent workflow I operate stops at an approval gate before anything touches a client. The agent does the mechanics. I do the judgment. That split is also why automation strengthens the quality side of a retainer instead of threatening it, the same logic behind the quality control system that prevents client churn before a retainer ever gets shaky.
What the adoption gap means for a small agency
The same report cites SHRM data showing 54% of organizations still have no AI in their HR function and no plans to add it. Read that as a market signal rather than an HR statistic. Adoption is uneven everywhere, which means operational advantage is still on the table for whoever moves while others deliberate. Your competitors are having the same internal debate you are, and most of them will still be having it in December.
The strategic implication is about margin structure, not saved hours. A 3 person agency that automates its top three high-tolerance workflows starts operating with the overhead profile of a smaller team while delivering the output of a larger one. Those margins fund better writers, longer runway, and pickier client selection, and each of those compounds. Six months from now the gap between agencies that re-ran the automation math this summer and agencies that did not will look like a difference in strategy. It will actually be a difference in who noticed a pricing update and acted on it.
