Here's the direct answer, and then the part that makes it useful.
Across the small-business market, a single well-scoped automation is generally a four-figure project, a connected set of workflows or a custom internal tool generally runs into five figures, and there's a monthly running cost underneath both that most first-time buyers never budget for. Those are broad market bands, not a quote, and the spread inside each one is enormous. The same task can be worth ten times more to one business than another, which is exactly why nobody honest will name your number before seeing your work.
What follows is what actually moves the price, so you can build the estimate yourself instead of waiting for someone to hand you one.
The three ways to buy, and what each really costs
Off-the-shelf tools. A subscription to something that already does roughly what you need. Cheapest by a wide margin, usually tens of dollars per user per month. The real cost isn't the subscription, it's fit: you adapt your process to the tool, and if the tool covers 70% of your workflow, someone is still doing the other 30% by hand forever. Try this first, always. A surprising number of "we need custom AI" conversations end with a tool that already exists.
A contractor or small practice building something custom. You get a system shaped around how your operation already runs rather than the other way around. This is where most owner-operated businesses land when the off-the-shelf option doesn't fit. Cost scales with the complexity below, not with how impressive the thing sounds.
A larger agency or enterprise firm. Higher rates, more process, more people, and a floor that often sits above what a small business needs. Sometimes correct, particularly when compliance or scale demands it. Frequently overkill for a company trying to stop retyping invoice data.
There's a fourth option people forget: doing nothing, deliberately, for now. That has a cost too, and it's the number the whole exercise is really measured against.
What actually moves the price
Six things, roughly in order of how much they matter.
1. How many steps and systems it touches. One trigger, one action, one system is a small build. One trigger fanning out into five actions across five systems is a different animal, not because any single step is hard, but because every connection point is a place things can break and therefore a place that needs handling.
2. Whether the systems you use have a real API. This is the single biggest swing factor and almost nobody asks about it. Software with a clean, documented API is straightforward to connect. Software with a partial API, or a vendor portal with no programmatic access at all, turns a two-day integration into a two-week one. If your core system is a legacy platform or a desktop application, expect that to dominate the estimate.
3. The condition of your data. If the information lives in a spreadsheet where the same vendor is spelled four ways and dates are entered in three formats, cleaning that up is real work that has to happen before anything can be automated on top of it. This is the most commonly underestimated line item in the whole category.
4. How wrong the output is allowed to be. Drafting an internal summary and touching a client invoice are not the same risk. Anything near money, contracts, or client relationships needs a staged approval step where a person sees what the system did before it commits. That's a design requirement, it costs real engineering time, and it's the right call every time.
5. How cheaply you can verify the result. There's a hidden cost inside every AI productivity claim: someone has to check the work. Some tasks have a built-in answer key, like a total that must reconcile, and checking is nearly free. Others don't, and reading the output carefully can cost as much as producing it did. Two businesses can automate the identical task and get wildly different value from it for this reason alone.
6. Who maintains it. If your team can maintain what gets built, the engagement is a build plus a clean handoff. If nobody can, it's a retainer, and it should be priced and named as one up front rather than discovered later.
The cost line almost everyone forgets
Ask any prospective vendor for the monthly number, not just the build number. It typically includes:
- Model API usage. Usually modest for a small business, and genuinely variable. It scales with volume and with how much text moves through each run.
- The automation platform. Whatever runs the workflow, on a monthly plan tied to run volume or task count.
- Any SaaS the integration touches, occasionally including a higher tier of software you already pay for, because API access is frequently gated behind a more expensive plan. Check this early. It has embarrassed better estimates than mine.
- Human attention. Somebody glances at the approval queue and notices when something looks off. This is small but it isn't zero, and pretending it's zero is how automations quietly rot.
None of these is usually large. Together they're the difference between a system that runs for three years and one that breaks in month four and nobody notices.
How to calculate what it's worth to you
You can do this yourself, right now, before talking to anyone. Every automation has the same underlying arithmetic.
Weekly hours = how often the task happens per week, multiplied by how long it takes each time.
Weekly cost = weekly hours, multiplied by what that person's time is actually worth per hour. Their real rate, whoever does the task today, whether that's you, an employee, or a contractor. Not a market-rate guess.
Yearly value = weekly cost multiplied by roughly 50, leaving a couple of realistic weeks out rather than assuming perfect uptime.
A task that eats six hours a week at $40 an hour is worth around $12,000 a year the moment it's handled correctly. That number is real precisely because it came from your business rather than an industry benchmark that has never seen it.
Do that for your three worst tasks and you'll have something most buyers never bring to a vendor conversation: a defensible ceiling. If the yearly value of a task is $3,000, you now know what you shouldn't spend, and you know it before anyone has a chance to anchor you somewhere else.
Why nobody publishes a flat number
It would be easier to put a price on a page. "$2,000 a month" fits neatly and requires nobody to think.
It would also be misleading, because the value of automating a given task varies by an order of magnitude depending on who does it today and how often. The same automation is worth very different amounts to a solo founder doing everything herself and to a company where three junior staff split the work. A flat price either overcharges one or undercharges the other, and there's no way to know which without asking.
That's the honest reason, and it's worth distinguishing from the dishonest version of the same behavior. A consultant who won't explain how they price is hiding something. A consultant who won't quote before understanding the work, but will walk you through exactly how the number gets built once they do, is being careful. That distinction is one of the more useful filters available when you're evaluating people to hire.
The question that outranks the price
Getting a number is the easy half. The harder half is deciding, in advance, what the recovered hours are for.
Saved time doesn't stay saved by default. It gets absorbed by whatever was already pressing, and six months later nobody can point to what changed. So alongside "what does this cost," it's worth answering: when this frees up four hours a week, where specifically does that time go? A named answer, before the build, is what turns an automation into something you can actually feel.
If you want a real number for your own situation rather than the market bands at the top of this post, the 7-minute Automation Assessment runs the arithmetic above against your actual answers and returns price bands ranked by how fast each one pays for itself. It costs nothing and there's no account to create. Come out the other side with your own numbers, then go talk to whoever you like.