Do not index
"How much will it cost to automate this?" Every agency owner evaluating agentic AI right now is asking that question, and almost all of them are answering it with chatbot-era pricing. That answer is wrong by orders of magnitude. BCC Research reported that agentic AI applications can consume 100 to 1,000 times more tokens per request than conventional chatbot interactions, which means the automation you priced from a demo is not the automation that shows up on the invoice at volume. Every agent decision is a unit-economics decision before it is a workflow decision.
The variance is the part that should stop you. As the reporting on that research framed it, three orders of magnitude of variance in a cost input means the vendor demo tells a marketing leader almost nothing about what the deployment costs at volume. A tool that costs pennies to demonstrate on one client can cost real money running on twelve, and nobody selling you the tool is incentivized to model that curve for you. The demo is a single request. Your business is thousands.
The cost curve nobody puts in the deck
Here is the practical shape of the problem. A chatbot request is one call and one answer. An agent request is a loop. It reads context, plans, calls a tool, reads the result, revises, calls again, and every one of those steps is billed. The loop is exactly what makes agents useful and exactly what makes them expensive, and the two cannot be separated by choosing a cheaper vendor.
What I run before automating anything now is what I call the Unit Cost Sort. Take the workflow and answer three things in order. How many times does this run per month at your actual client count, not your demo client count. What does one run cost when the agent loops five times instead of once, which it will on messy real inputs. And what would a competent person be paid to do the same task at that frequency. If the third number is lower than the second, the workflow is not an automation candidate, it is a hiring decision you were about to make badly.
That sort has changed my answers. Some things I automated stayed automated, mostly high-frequency mechanical work where the input is clean and the loop is short. Other things I built, ran for a month, and abandoned, because the token bill on messy inputs made a virtual assistant at $1,200 a month the cheaper and more reliable option. There is no shame in that outcome. The mistake is discovering it on a September invoice instead of in a spreadsheet in August.
Why agencies get this wrong specifically
An agency between $200k and $2M in revenue runs on gross margin per retainer, and the retainer price is usually locked for six or twelve months. That means a variable cost input sitting inside a fixed-price deliverable, which is the exact structure that quietly destroys margin without ever showing up as a lost client. You keep the account, you keep the revenue, and the profit leaks out the bottom while everyone congratulates the team on the new workflow.
Solo creators and small teams get hit differently. They budget from the monthly subscription price, which covers the seat rather than the consumption, and they only meet the real cost once usage crosses a threshold. The gap between the seat price and the usage cost is where the surprise lives, and agentic tools widen that gap by design.
This is not for you if you are running one or two agents on low-frequency work. If you generate a handful of drafts a week and the whole thing costs less than a client lunch, model nothing and get on with it. Skip this if your billing model is hourly and you pass tooling through at cost, because then the client absorbs the variance and your incentive structure is already aligned. The Unit Cost Sort matters when you have fixed-price retainers and rising volume, which is most of the agencies I talk to.
What makes this harder is that the cost is easy to measure and the benefit usually is not. Token spend arrives as a clean number on a bill. The value of the automation shows up as time nobody logged and quality nobody scored. That asymmetry pushes teams toward defending the spend rather than evaluating it, which is the same trap that shows up when agencies measure LinkedIn success by whatever the dashboard happens to count. If you are going to spend on agents, decide in advance what result would justify it and what number would make you shut it off.
None of this is an argument against automation. It is an argument against automating without arithmetic. The agencies that come out of the next two years with intact margins will be the ones that treated every agent as a line item with a unit cost and a kill threshold, rather than as a capability they were told they needed. The ones that struggle will not have made one large mistake. They will have made forty small ones, each defensible in isolation, and discovered the total only when they tried to raise prices to cover a cost structure they never chose. Knowing which of your workflows are genuinely cheaper as software, and being willing to say the rest are cheaper as people, is going to be a real competitive position while everyone else is still guessing from a demo.
