Last month a founder friend forwarded me a quote he’d gotten from an agency: $85,000 for an “AI transformation program” at his 22-person company. He wanted to know if that was normal. I told him the truth, which is that the number was either a bargain or highway robbery, and the quote gave him no way to tell which. Nobody publishes real AI automation cost numbers, so everyone buying this stuff is guessing in the dark.
That annoyed me enough to write this. If you’re new here, I’m Damian, and I’ve spent six years building automations for small and mid-sized businesses, which means I’ve both written these quotes and received them. Our no-hype playbook for AI automation in SMBs covers the what and the why, but it never fully answered the money question. So here are the ranges I actually see in 2026, with the caveat that this market moves fast and some of these figures will be stale by summer.
Why nobody publishes real AI automation cost numbers
Three reasons, and only one of them is legitimate.
First, scope varies wildly. “Automation” can mean a $200 Zapier connection or a multi-agent setup that reads contracts, drafts replies, and updates your ERP. Any average across that spread is meaningless. Second, agencies protect their pricing on purpose. Publish $8,000 for a lead-routing build and every prospect negotiates down from $8,000 forever. Silence is more profitable. Third, the legit one: costs keep falling. Model API prices dropped something like 90% between 2023 and 2025, so published numbers have the shelf life of yogurt.
“It depends” is still a coward’s answer, though. It depends within a range, and ranges can be published. So that’s what this post is.
The three budget tiers I actually see
Across roughly 80 builds, client projects cluster into three bands. Not two, not five.
Tier 1: Quick wins ($0-5,000 to build, under $200/month to run)
This is you, a Make or Zapier account, an API key, and a couple of focused weekends. Early last year I built lead triage for my own pipeline in Make: inbound form submissions get read by a small OpenAI model (GPT-4o mini at the time), scored, routed to the right follow-up, and posted to Slack. Twelve hours of work, all in. Running costs came to $29/month for Make plus $15-20/month in tokens, and it freed about five hours a week of manual sorting. It’s kind of addictive once the first one works. The full story of how I fixed my inbox chaos is on the site if you want the build details.
Don’t want to DIY it? A good freelancer charges $1,500-4,500 for a single well-built workflow. If your first project quotes above that range, someone’s padding.
Tier 2: Real projects ($10,000-40,000 one-off, or $2,000-5,000/month)
Multiple connected workflows, proper error handling, a dashboard so you can see what’s actually happening, some custom code where the platforms run out of road. Six to ten weeks with a competent partner. Most SMBs that get serious land here, and honestly this is where the money usually is: enough investment to make things reliable, not so much that payback takes years. Retainers make sense when you want continuous improvement instead of a one-time handover. I’ve seen both models work and both get abused. The real difference is whether someone stays on the hook when things break.
Tier 3: Full programs ($50,000-150,000+ per year)
Multi-agent setups, custom infrastructure, legacy integrations, compliance work. If you’re curious how these agentic systems actually work, that’s a separate post, because the architecture is a topic in itself. On cost I’ll say this: if someone pitches you Tier 3 before you’ve run a single Tier 1 experiment, walk away. That’s not a strategy, it’s a shopping spree.
My honest guess is that 80% of SMBs belong in tiers one and two. Could be wrong. Not by much, though.
What moves the number
Data quality, more than anything. Not the AI part. The plumbing.
I learned that the expensive way. In 2024 I quoted a manufacturer $9,000 fixed for an order-to-invoice automation: pull orders from their system, generate invoices, file everything to Drive, sync the ledger in QuickBooks. Turns out their “ERP” was a CSV export that three different people hand-edited, each with their own naming conventions for the same customers. We burned three weeks on cleanup before any automation could make sense. The project landed at $14,500 and I ate part of the difference, because I quoted before auditing the data. My fault, entirely.
Still stings a bit.
I still quote fixed price, by the way, which smarter people than me call dumb. They might be right.
Other drivers, roughly in order of how often they blow up budgets:
- Number of systems connected. Every integration is a mini-project with its own auth quirks and failure modes.
- AI judgment vs. deterministic steps. A zap that moves data is cheap. A workflow that makes judgment calls needs evals and guardrails, and those cost real hours.
- Error tolerance. A demo that works 90% of the time takes days. A production system that works 99.5% of the time takes weeks. That last stretch is where the money lives.
- Who builds it. The same lead-routing workflow can cost $2,000 from a solo freelancer or $40,000 from a consultancy whose sales team outnumbers its engineers.
Platform fees, compared honestly
Ballpark 2026 numbers, with the caveat that every one of these companies reprices more often than I change my mind about Zapier’s task model. Verify before you budget.
- Zapier: starts around $20-30/month, but task-based pricing gets expensive fast. Busy SMBs often end up paying $100-300/month.
- Make: starts around $9-10/month. Most of my clients sit at $29-79/month. Cheaper per operation than Zapier, steeper learning curve.
- n8n: self-hosted is free, plus $10-20/month for a VPS and some of your evenings. Cloud plans start around $20-30/month. Maximum freedom, maximum rope.
- AI API spend: genuinely small now. Classifying a few thousand leads or documents with a small model costs dollars, not hundreds. Realistic SMB spend runs $10-150/month unless you’re processing serious volume.
We keep a detailed comparison of automation platforms if you’re still choosing, and the Zapier vs. Make breakdown covers the two most SMBs actually pick. My opinion on fees generally: people obsess over $30 versus $80 a month when the labor to build and maintain the thing is pretty much 10x that. The platform fee is a rounding error. Pick one in an afternoon and move on.
The hidden costs that actually get you
Your build quote is never the whole number. I’ve stopped presenting builds without these, because pretending they don’t exist is how projects sour. Add:
- Failure handling and monitoring. Tokens expire, APIs change, workflows fail silently. Budget $500-2,000 for real alerting and retry logic. This is the discipline of workflow reliability engineering, and skipping it is how automations quietly rot.
- Maintenance. Plan on 10-15% of build cost per year. Anyone promising zero maintenance is lying or hasn’t waited long enough.
- Security review. The moment an AI agent holds credentials and inbox access, you’ve added attack surface. Agent security isn’t optional at tiers 2 and 3.
- Testing and iteration. Prompts drift and edge cases surface. Treating prompt work as a one-time setup cost fails in predictable ways, which is why loop engineering beats prompt tinkering. Budget for continuous testing.
- Team adoption. People need to trust the system before they’ll use it. Count $1,000-3,000 of internal time for training and hand-holding.
ROI math I actually trust
Simple version: monthly benefit = (hours saved × fully loaded hourly cost) + rework avoided + any revenue effect, like faster lead response. Payback in months = build cost ÷ (monthly benefit minus monthly running cost).
Worked example from a real tier-2 project: invoice processing that saved about 40 hours a month across two roles, loaded cost around $38/hour. That’s $1,520 a month. Build was $12,000, running costs about $120. Payback lands near 8.5 months, and everything after that is margin.
Two honesty checks, though. Hours saved rarely get banked; the time gets redeployed into other work, which is still valuable but isn’t a payroll cut, so don’t model it as one. And speed-to-lead revenue claims are the most inflated numbers in this industry, mine included when I’m not paying attention. The only ROI I fully believe is one measured against 90 days of live data, not a spreadsheet built before launch.
One more thing worth pricing: the cost of doing nothing. If your team spends 15 hours a week on work a machine could handle, that isn’t free either. It’s just invoiced in salary instead of software.
If a vendor’s model assumes 100% adoption and zero maintenance, you’re being sold to, not consulted.
Where I’d start
With a $2,000 experiment on your most painful recurring task, measured honestly for a quarter, then scaled if it holds up. Cheap failure is the entire point of tier 1, and it’s the cheapest education you’ll buy in this space.
If a quote is sitting in your inbox right now and you want a gut check on whether the numbers make sense, that’s literally my job. Send it over.
Cheers,
Damian
