What Does An Agentic Workflow Actually Cost?

Nobody EVER enjoys talking about the bill. When people first get excited about AI agents that work while they sleep, the conversation is usually all about what the pipeline can do, and almost never about what it costs to run. Then their first usage statement arrives, and it turns out that “set and forget” has a price tag that’s easy to overlook until it isn’t.

The good news is that agentic workflows are usually far cheaper than the manual work they replace. The not-as-good news is that “usually” carries a lot of responsibility in that sentence, and the only way to know how that affects your own pipeline is to measure it. Here’s a simple way to do that.

Why the cost surprises people

A single chat with an AI assistant feels free, or at least flat-rate. An agentic pipeline is different in two ways. First, it runs many steps, and each step reads context and produces output, all of which burn usage tokens. Second, it runs without you watching, so an inefficient setup will secretly burn through your budget while you’re not paying attention.

A pipeline with four agents, each reading a project folder and passing their results down the pipeline, will process a lot more text than a one-off conversation does, simply because each agent starts off by reading what the last one wrote. Add a long AGENTS.md that every agent must load at the start of every run, and the baseline cost goes up. None of this is wrong, or even wasteful by default. It just needs to be visible and accounted for.

The unit that matters: tokens

Model usage is measured in tokens, which are chunks of text roughly the size of a short word (or part of one). You pay separately for tokens going in (everything the model reads) and tokens coming out (everything it writes), and output is always priced higher. In the budget exercise I wrote for Creative AI, we use example rates of $3.00 per million input tokens and $15.00 per million output tokens. Those numbers will change over time as the market fluctuates, so check your provider’s current pricing before you trust any math. However, the structure stays the same: several tokens go in, several come out, multiplied by two rates.

You don’t need to count those tokens by hand, of course. The tools report their usage so that an agent can estimate it for you. What you need is a place to write it down after each run, so it’s not lost.

Build a budget log

This is the simplest useful suggestion in this whole post, and it takes about fifteen minutes to set up. Add a section to your summary agent’s instructions that makes it produce a budget log at the end of every run. The log has four parts:

• Per-agent token estimates. A small table with a line for each agent in the pipeline, roughly how many tokens it read, and how many it produced.

• Cost estimate. Input tokens times the input rate, output tokens times the output rate, added together. A few lines of arithmetic.

• ROI check. How many hours this work would have taken you manually, what you’d charge for them, and what the pipeline cost as a percentage of that manual cost.

• Notes. Anything unusual: a retry, an unexpectedly long document, a step that took far more tokens than the others.

After a few runs, you stop looking at individual numbers and start seeing patterns. Which agent is the expensive one? Does cost scale with the size of the brief or stay flat? What’s the average cost of a complete run? With a little attention over time, it’s a snap to fine-tune your agent definitions (and usage) to fit your general project budgets.

The ROI check is the actual point

A raw dollar figure doesn’t mean a lot on its own. Four dollars to process a client brief only sounds like a lot or a little based on the alternative option. So it’s not surprising that the next question we should ask is how that compares to doing the same work by hand.

If a pipeline run costs a few dollars and replaces two hours of your time at your usual rate, the math isn’t a question. If it costs a few dollars and produces something that you then spend two hours fixing, the math is different, and the log will reflect it because you’ll know the time you actually spent on review. That’s the honest version of ROI: not what the pipeline cost against the best case, but what it cost against what you really saved.

For me, this ended up saving a ton of work on a processing job for a client that was editing >1000 images individually, based on the colors within each one of them. I created a n8n workflow to visually process these images one by one, decide which of a series of color-range-specific overlays should be added to each, customize the overlay to the specific image, and then save the new image back out to file. Seems straightforward, right?

What we found running the script on a 10-file sample directory was that if the average color in the image fell into a few key ‘blind spot’ ranges, the AI would struggle to decide, then run a second (and sometimes a third) analysis, quickly running up costs. If I hadn’t been there to see the error on that smaller set of images, we would have probably seen the bill for that full 1000-image run come back almost three times its projected cost. That’s amazingly important information to learn before running the actual job.

So I re-sequenced the workflow to do a less invasive scan and ‘loose-separate’ the images into the different color ranges first, before then running the processing. Turned out this not only saved a few tokens, but the errors went away as instead of agents getting stalled on a single image amid a long run, any errors weren’t included in the big run. I then cleaned the errors up by hand v. The 5 minutes of manual tagging after the fact cost far less (in hourly rate terms) than if the model had racked up a huge bill midstream when it hit a questionable image, without finishing the run.

Planning your budgeting into your agentic definitions also helps you price your own work more effectively. If you run a studio or freelance practice, knowing exactly what a pipeline step costs helps you decide which services you should offer, how to best bundle them together, and where your profit margin lies.

Where the money usually leaks

Once you’re up and running with budget tracking your agents, a few common sources of waste become easy to spot.

Oversized context. This is when each agent loads the whole project folder but really only needs one file. Trimming what each agent reads in is the single biggest saving in most pipelines.

Long instructions repeated everywhere. A manual that’s thousands of words long gets read each run. Keep the shared parts of a workflow lean, and put role-specific detail only where it’s needed.

Retries and loops. An agent that fails a step and tries again, then again, can quickly multiply the cost of a run. A rule like “retry once, then stop and log the failure” costs nothing and prevents a terrible night.

The wrong model for the job. Not every step needs the most capable model. Sorting files or reformatting text can often run on something much cheaper than your default model, saving the more powerful frontier models for the work that needs judgment.

Set a ceiling anyway

Even with good logging, it’s always worth maintaining a hard limit for each project up front. Most platforms let you set a spending cap, and it’s worth five minutes to do it before you schedule anything to run unattended. Think of this as a guardrail that covers any failure you didn’t anticipate. A runaway loop or a mis-written schedule can’t spend more than the cap, and you’ll find out about it the next morning, instead of on your credit card statement.

Pick a number you’d be comfortable losing and raise it only as the pipeline earns your trust.

A worked example, in rough numbers

Let’s say your intake-and-proposal pipeline processes a typical client brief. The intake agent reads a two-page email and a short manual and writes a project summary. The strategist reads the summary and the manual and writes an outline. The writer reads both and drafts the proposal. Across those three steps, you could see something like fifty thousand tokens read in, and ten thousand written out. At the example rates noted above, that’s about fifteen cents of input and fifteen cents of output, thirty cents total for the run.

Now compare that to the three or four hours the same work takes by hand. Even at a modest hourly rate, the pipeline costs a fraction of one percent. The point of the example isn’t the specific numbers, which will undoubtedly differ for you. It’s more the habit of writing them down. Once you have an actual figure, you can answer a question most people only guess at. Is this pipeline worth running, and how many times a day can it run before it stops being cheap?

What has your own workflow cost you so far? Even a rough number helps, and I’d love to compare notes in the #general channel on my Discord.

And if you’d like a ping when the Creative AI book by Tom Green and myself goes on sale, sign up at my Creative AI page. The book includes the full budget-aware pipeline exercise, with templates you can copy straight into your own setup.

Hero image: generated by Midjourney, text and editing by human in Affinity