What It Costs to Automate Agency Operations: Hourly, Per Project, Per Month

Because there is no defensible number to publish, and the tidy bands you do find were written by somebody with a reason to want them there.

I went looking for a market rate worth quoting and came back with a mess. Search for an agency automation consultant and the cost guides in the top five will tell you $150 to $350 an hour with total confidence. Not one of the ones I read all the way through cites where that came from. The most detailed of them runs past 4,000 words, carries no external sources at all, and has no tables or charts either. That is a remarkable thing to publish about money.

The wage data is real, and it disagrees with itself. Salary.com put the average US base salary for an automation consultant at $71,570 as of 1 July 2026, an hourly equivalent of $34, with the middle 80 percent of the market, tenth percentile to ninetieth, falling between $57,294 and $92,565.

Indeed put the same job title at $109,017, updated 31 July 2026, drawn from 54 salaries taken from job postings over three years. Same month, same title, a $37,447 gap. Neither is wrong. They are counting different people, and neither is counting the thing you are actually trying to buy, which is an outside specialist rather than an employee on your payroll.

So treat every published range for this service, this one included, as an opening position rather than a price. The figures that survive being checked are the ones with a price list behind them: what the platforms charge, what an employee earns, and what a given shop is willing to put on its own website.

Hourly is for the part of the work where nobody knows yet what is being built, and that is the only part it is good for.

The bands you will be quoted sit somewhere between $150 and $350 an hour for an independent in the US, and higher through a firm. I cannot source those, for the reasons above. What I can source is the pair of numbers that bracket them. The Bureau of Labor Statistics puts the median US software developer at $65.38 an hour, which is roughly what the skill costs at cost, before anybody's overhead, tooling, sales time or risk. Salary.com's $34 hourly equivalent for an automation consultant employee says the same thing from the other end. When a rate card says $250, most of the distance between that and $65 is not extra skill. It is somebody pricing the risk of a project nobody has scoped.

Here is the real problem with hourly, said plainly. The meter rewards the wrong behavior on both sides of the table. Every hour spent understanding your mess is billable, so nothing about the arrangement is in a hurry to stop understanding it. And you have no idea what the total is until it lands. The failure mode always has the same shape: a two-week discovery, still running in month four, still producing documents.

Buy hours for exactly one thing. A scoped audit, with a fixed end date and a written deliverable, priced before it starts. If a shop wants to bill you hourly for the build itself, that is a shop telling you it cannot size the build, which is worth knowing and is one of the things that should stop the meeting.

Fixed price is the right default for anything with an end state, because it is the only shape where the person quoting has to do the thinking before you pay for it.

The figures published for a single-workflow build tend to run $5,000 to $15,000, with multi-workflow programs quoted from $15,000 up. Same caveat as before: nobody sources them. What a fixed price genuinely buys is a transfer of risk. If an integration turns out to be nastier than anybody thought, that is now their problem, and the reason a shop can take that bet is that it has built the shape before.

What decides where a quote lands is rarely how clever the automation is. It is four things: how many systems it has to touch, whether those systems have real APIs, whether a human has to approve something in the middle, and whether the process is written down at all. The last one moves the number most. A workflow that three people run three different ways is not an engineering project yet, and quoting it as one is exactly how a fixed price turns into a change order in week three.

Our numbers, since I am asking everybody else to publish theirs. A one-off starter build, meaning one scoped automation shipped and handed over, is $1,500 to $2,500 fixed. A two-week production sprint that puts a full workflow live is $5,000 fixed. Both sit on our pricing page and neither has moved this year.

A retainer is for work that keeps arriving and for having somebody on the hook when a step fails at 2am. It is a different purchase from a project, bought for different reasons.

Ours start at $5,000 a month. That is a floor for ongoing build-and-run work, not a minimum spend to talk to us, and the difference matters more than it sounds. Plenty of agencies should buy a $1,500 starter build, get one Thursday back forever, and never speak to us again. Any shop presenting its retainer floor as the entry price for everything is quietly telling you it would rather not sell you the small thing.

The honest test is whether the queue is real. If you can name the next three workflows and each one is already annoying somebody by name, a retainer is cheaper than three separate scoping exercises with three separate ramp-ups. If you cannot name them, you are paying a monthly fee for the option to think of something later, and that option is worth a lot less than the fee.

The half of a retainer nobody prices properly is ownership when it breaks. Every automation breaks eventually. The question is whether it fails loudly and gets fixed on a Tuesday, or fails quietly for three weeks while your account team keeps sending clients numbers that stopped updating in April. That failure rate is the least discussed and most expensive variable in this whole purchase, which is why we published what actually breaks and how often instead of a success story.

Two line items keep billing after handover: the platform the workflow runs on, and the model calls it makes. Both are published prices, and both are smaller than people brace for.

n8n charges EUR 20 a month billed annually for 2,500 workflow executions, and EUR 50 for 10,000. Zapier Team is $69 a month billed annually at 2,000 tasks, or $103.50 if you pay monthly. Make Core is $12 a month for 10,000 credits, where one credit is one module action. A single agency workflow lives comfortably inside the cheapest paid tier of any of them. The same workflow running for forty clients does not, and that is where n8n and Zapier stop being interchangeable.

Model calls are the item people expect to hurt and then find anticlimactic at agency volume. Anthropic's published API pricing puts Claude Haiku 4.5 at $1 per million input tokens and $5 per million output, and the worked example on that same page prices 10,000 conversations averaging about 3,700 tokens each at roughly $37. Ten thousand. If your agency is putting 10,000 of anything a month through one workflow, you are running a bigger business than the one this post is written for.

The reason that lands as a surprise is that the loudest cost warning in this field is aimed at somebody else entirely. Gartner expects inference to account for at least 70% of a model's lifetime cost, and predicts that through 2028 at least half of GenAI projects will overrun their budgets on poor architecture and missing operational know-how, as reported by THE Journal. Both things are true, and both are statements about enterprises running models at scale in production. At the volume one agency pushes through one reporting workflow, the build dominates and the run cost rounds to noise. Do not let an enterprise number talk you out of a $2,000 build.

Multiply before you negotiate. The payback math here fits on a napkin, and almost nobody does it before taking the sales call.

Count the minutes first, per client, per month, for the process that actually happens rather than the one on the wiki. Then price the time, and note that the full three-leg version of this count adds the rework and the billing you gave up. The Bureau of Labor Statistics puts the median market research analyst and marketing specialist at $37.87 an hour, which is base wage before employer costs, so an agency should be working from something closer to $47 to $53 fully loaded. Hours saved per month, times that number, times twelve, is the annual value. Quote divided by monthly value is the payback period.

A worked version. Client reporting takes six hours a client a month across ten clients, so 60 hours. Get it to one hour a client and you have bought back 50 hours a month. At a loaded $47 that is $2,350 a month, or $28,200 a year, and a $5,000 fixed sprint pays back in about nine weeks. That is why reporting is almost always the first thing anybody automates: it repeats per client per month, so it scales with the exact thing you are trying to grow.

The number people leave out is review time in month one. Nobody trusts a new automation, and nobody should. Budget your team checking its output against what they would have produced by hand for the first four weeks, and treat that as part of the price rather than as a nasty surprise in the second invoice. Any quote that has no line for it was written by somebody who has not run one of these.

Five things, roughly in the order of how hard each one pushes.

Systems without a real API. If a step in your workflow lives inside a tool that only gives up its data when a human clicks Export, that step costs more than the four around it put together. Find out which of your tools have APIs before you ask anybody for a quote, because the answer changes the quote.

Approvals in the middle. A workflow that runs end to end is cheap. A workflow that has to stop, show something to a person, wait, and resume is a different piece of software: state to keep, notifications to send, timeouts to handle, and a decision about what happens when nobody replies for four days.

Judgment on messy input. This is the line where the cheap tools stop and a build starts. A filter can route an email by sender. It cannot read forty client emails and work out which three are really about a scope change. If your process has a step like that, you are buying something categorically different from a Zap, and we have written up exactly where that line falls.

Anything a client sees. Client-facing output carries a quality bar your internal tooling does not, plus a rollback plan, plus somebody's name on it. Price it accordingly.

And the one entirely within your control: whether the process exists. Three people running it three ways means the first week of the engagement is a management decision wearing an engineering hat, and you are paying engineering rates for it. Write it down first and you will be quoted a smaller number.

Four lines. A quote with all four is comparable to the next quote, which is the real problem when nobody publishes anything, and the same reason we had to anchor the cost of hiring an AI automation agency on wage data instead of rate cards.

What is fixed and what is hourly, listed separately. A single number covering "discovery and build" is a number where discovery can quietly eat the build. Ask for discovery priced on its own with its own end date.

Who pays for the platform and the model calls. Trivial money, going by the published prices above, but it belongs on your card in your account rather than billed through somebody else at a markup. The second version means you cannot see it, and what you cannot see you cannot cancel.

What handover includes, spelled out: credentials in your accounts, code in your repository, written documentation, and one of your people who has changed a step while somebody watched. Missing any of those and you are renting the automation rather than owning it.

What happens when it breaks, and for how long after the final invoice. A support window with a number of days attached is a real answer. "We stand behind our work" is not an answer, it is a mood.