TL;DR: simple two-app flow $500-1,500. Multi-step business process $1,000-3,000. Custom pipeline with AI and error handling $3,000-8,000. Full custom system replacing a manual back office $8,000-25,000. Running costs $50-200/month for most small businesses. The number that surprises people is not the build, it is the second year.
Every article with this title is useless in one of two ways. Either it hands you a range so wide it means nothing, or it is a soft pitch that stops right before the number and asks you to book a call. I have been on the buying side of that and it is infuriating, so here is the version I wish someone had written for me.
These are the bands we quote in, why a project lands in one band rather than another, and the running costs that never make it into the proposal. I run 2pizza.team, we have shipped automation systems for 80+ clients, and the numbers below are what we charge and what we see competitors charge. Where I am uncertain I say so.
Why the range is genuinely that wide
AI automation projects run from $500 to $50,000 and the spread is not vagueness. Connecting two apps with an if-this-then-that rule and building a system that runs a back office are both called automation, and they are not the same kind of work. One is configuration. The other is software engineering with a different label on it.
A useful test: if the whole thing breaks and nobody notices for a week, you are buying configuration. If the whole thing breaks and your operations stop by lunchtime, you are buying software, and software carries the cost of error handling, monitoring, and someone who can fix it. Most of the price difference between the bands is that second category.
The four things that actually move the number
Scope conversations go in circles until you name the variables. There are four, and in our experience they explain most of the variance between two quotes for what sounds like the same project.
1. How many systems have to talk to each other
Each integration is its own small project: read the API docs, handle authentication, map the fields, deal with the one endpoint that behaves differently from the rest, then test it against real data. Two systems is a day. Five systems is not two and a half days, because now you also own the question of what happens when system three is down and systems one and two have already committed. Integrations compound.
2. Whether the logic is rules or judgment
A rule is cheap: if the invoice total is over 5,000, route it to the finance lead. Judgment is expensive: read this supplier email and decide whether it is a delivery delay, a price change, or a complaint. Judgment means an AI layer, and an AI layer means prompt work, a test set, a fallback for when the model is unsure, and a way for a human to correct it. Budget roughly 30-50% on top of the equivalent rules-only build.
3. Volume and what failure costs
Fifty records a month and fifty thousand records a month are different systems. At low volume you can retry by hand. At high volume you need queuing, rate-limit handling, idempotency so a retry does not double-charge someone, and alerting that tells you before the client tells you. This is the single most common reason a quote comes back higher than expected.
4. How clean your data is
This is the one nobody budgets for. If your product names are inconsistent, your customer records are duplicated across two tools, or half your orders arrive as free text in a chat window, someone has to reconcile that before any automation can be trusted. We have seen data cleanup take longer than the automation itself. It is worth asking any agency directly how they price it, because plenty of them discover it mid-project and come back for more money.
Secondary drivers, smaller but real:
A dashboard or admin UI instead of no interface at all: add $1,000-3,000
Proper error handling and alerting: cheap to build in, expensive to retrofit
Compliance requirements, GDPR or the Australian Privacy Principles: add scoping time before any code
Number of edge cases that must be handled rather than escalated to a human
Whether you need the system documented well enough for someone else to maintain it
Price bands by project type
What a project costs to build, and more usefully, what you actually get for the money at each level.
$500-1,500: a single flow
Two or three steps between standard apps. Form submission creates a CRM record and sends a Slack message. New order triggers a supplier email. This is Make or Zapier configuration. It works, it is worth doing, and it will not survive a change in either app without someone looking at it. No error handling beyond the platform default.
$1,000-3,000: a business process
Five to ten steps, two or three integrations, real branching logic. Lead comes in, gets enriched, gets scored, routes to the right person, follows up if nobody touches it in 48 hours. This is the band most small businesses should start in, because it automates something that currently occupies a person rather than something that currently occupies a click.
$3,000-8,000: a pipeline you can rely on
Custom n8n or code, an AI layer doing classification or extraction, error handling, retries, alerting, and a place to see what happened. Invoice processing, document extraction, support triage, outbound systems. The difference from the band below is not features, it is that this one is built to be trusted with money or with customer-facing decisions.
$8,000-25,000: a system that replaces a back office
A custom application with a database, an interface people log into, and integrations out to everything else. Production ERPs, client portals, scheduling systems. You buy this when the off-the-shelf tools genuinely do not fit your process and you have checked, properly, that they do not.
Typical builds and their bands:
Simple two-app flow, standard connectors: $500-1,500
Multi-step process with logic and 2-3 integrations: $1,000-3,000
AI customer support with retrieval, classification and escalation: $2,500-6,000
Invoice processing with vision extraction, matching and ERP write-back: $4,000-8,000
Outbound sales system with enrichment, sequencing and qualification: $5,000-12,000
Custom production ERP or client portal: $8,000-25,000
Three builds and where they landed
Bands are abstract, so here are three systems we shipped and the shape of each. I am giving the band rather than the invoice, because what a specific client paid is theirs, not mine to publish.
An online store run by one warehouse worker
Simbago sells pet supplies. Orders flow in from WooCommerce, get picked and confirmed automatically, supplier reorders fire when stock hits a threshold, and a chatbot handles the bulk of support. One employee now runs what would normally take three or four people. Multiple integrations, an AI layer on support, real volume: this is a mid band build, the $3,000-8,000 shape.
A confectionery studio running on WhatsApp and a notebook
Cupcake Studio was taking custom orders through Instagram DMs, calculating ingredient quantities by hand per recipe, and tracking production capacity from memory. Past 30-40 orders a week that stops working: double bookings, miscalculated ingredients, inconsistent client communication. We built a purpose-built production ERP covering orders, CRM, a recipe database with auto-scaling, scheduling, procurement and delivery. Admin went from around 8 hours a day to about 1. Custom application with a database and an interface: top band.
A B2B supplier with a good product and no sales system
B2B Pallets, in Australia, had one salesperson doing manual outreach and burning hours on qualification calls that went nowhere. We built list building and enrichment, personalised sequences, and an AI voice agent that runs the first qualification call, scores the lead, and only passes warm prospects to the human. Outreach volume went up roughly fivefold and coverage went from one city to nationwide. Several integrations plus a voice AI layer: upper mid band.
The running costs nobody puts in the proposal
The build is a one-off. The running cost is forever, and it is where budgets get embarrassing in month seven. Three separate lines, and they behave differently.
Platform fees, predictable
Make sits at roughly $9-16/month for most small business automations. n8n cloud starts around $20/month, or $30-50/month in server costs if you self-host, which also buys you data residency if that matters to you. Zapier is the expensive one at volume and that is the main argument against it. These numbers are stable and easy to plan around.
AI API usage, the one that surprises people
For typical small business volumes, $20-100/month on Claude or OpenAI. The surprise is that this scales with usage, not with subscribers, so a successful automation gets more expensive as it works harder. Ask any agency what your per-record model cost is and what happens to it at ten times the volume. If they cannot answer, they have not measured it. A well-built pipeline usually has a cheap model doing the routine cases and an expensive model only on the hard ones, and that design choice can be a five to ten times difference in the bill.
Maintenance, the honest version
Most well-built automations run for months without anyone touching them. What does break them: an app you integrate with changes its API, a supplier changes their invoice format, or your own process changes and nobody told the system. Realistically budget a few hours a quarter, or a retainer if you have no technical person at all. A $500/month maintenance retainer on a simple Make flow is not maintenance, it is a subscription to someone answering the phone. Ask exactly what is included and what response time you are buying.
Total for most small business automations: $50-200/month all in. High volume or multiple AI-heavy systems: up to $500/month. If someone quotes you a build with no running cost at all, they have either not thought about it or they are hiding it in a retainer.
Fixed price or hourly, and why it matters to you
Hourly billing on automation work is a signal worth reading. An agency that has built the same shape of system before knows how long it takes. If they will not commit to a number, either they have not built it before, or the incentives are pointed the wrong way, because under hourly billing the slower option pays better.
We quote fixed price on fixed scope. You know the number before anything starts, projects run two to six weeks, and if we mis-scoped something that is our cost to absorb, not a change order. The trade for you is that scoping has to be done properly up front, which takes a day or two and some of your attention. That trade is worth making.
Things in a quote that should make you ask another question:
An hourly rate with no estimated total
A maintenance retainer with no list of what it covers
No mention of what happens when an integration fails at 2am
No mention of who owns the accounts, the code and the data at the end
A discovery phase that costs real money and produces only a document
How to make your own project cheaper
Most of the levers are on your side of the table, not the agency’s. These are the ones that actually reduce a quote rather than just squeezing a margin.
Automate one process end to end instead of three processes halfway. Half-automated processes still need a person watching them, so you pay and keep the headcount.
Clean the data first, or agree explicitly who is cleaning it. This is the largest single hidden cost.
Decide the edge cases before the build. Every ‘what should happen if’ answered during the build is a change of direction; answered before, it is a line in the spec.
Let the system escalate to a human on the rare cases. Insisting on 100% automation of a case that happens twice a month can add thousands.
Use the tools you already pay for where they fit. Replacing a working CRM for elegance is a cost with no return.
When the honest answer is do not automate this
We turn down work over this and I would rather say it here than on a call. If the process runs fewer than a handful of times a month, if it is about to change anyway, or if it is broken as a process and automating it just makes it break faster, the return will not be there. Automating a bad process gives you a bad process running at speed and with less visibility.
The rough arithmetic: take the hours the process eats each month, multiply by a realistic loaded hourly cost, and compare that to the build spread over 12 months plus the running cost. If it does not clear comfortably inside a year, the project is not the problem, the priority is.
What to bring to get a real number
A scoped quote takes us one to two days once we have four things. Without them, any number anyone gives you is a guess dressed as a quote.
The process written out step by step, including the steps you think are obvious
The systems involved: every tool that touches it today, including the spreadsheet
Rough volume per month, and what you expect it to be in a year
What the output is and where it needs to end up
If you are not sure the process is even worth automating, that is a fine place to start from. Half the useful part of a scoping conversation is finding out that the expensive thing is not the step you thought it was.
Four questions, two minutes, no call required: take the audit at 2pizza.team/audit and get an estimate for your specific process. If the honest answer is that you do not need us yet, that is what it will say.
Originally published at 2pizza.team. We build AI and automation systems for small teams – fixed price, two to six weeks. See the work.