The honest answer is that AI automation costs anywhere from about $3,500 for a two-week strategy engagement to $75,000 or more for a custom build that touches your core systems. Most single-workflow automation projects land between $10,000 and $30,000. That range is wide because “AI automation” covers everything from a chatbot that answers billing questions to a document pipeline that replaces a 14-person review queue.
This guide breaks down what you actually pay for, which pricing models are worth trusting, and what pushes a quote up or down. The numbers here are our published prices at PXT AI, and the project figures come from builds we have shipped, not market averages pulled from thin air.
The short answer, in one table
Here is what different kinds of AI work typically cost in 2026. These are our own entry prices, and they are close to the market for engineering-led shops that build production systems rather than demos.
| Type of work | What you get | Typical price | Timeline |
|---|---|---|---|
| AI strategy sprint | An audit, an opportunity map, and a 90-day roadmap | $3,500 fixed, up to $6,000 | 2 weeks |
| Workflow automation build | One or two automated workflows live in your stack | $10,000 to $30,000 (capped) | 4 to 10 weeks |
| Custom AI or LLM build | A chatbot, RAG system, or embedded AI feature | $15,000 to $75,000 (capped) | 6 to 12 weeks |
| AI product development | A full web or mobile product with AI at the core | From $25,000 (capped) | 10 to 17 weeks |
| Embedded AI team | A senior pod working inside your company | From $15,000 per month | 3-month minimum |
On the sprint range specifically, since the spread invites suspicion: $3,500 is the price, not a teaser. It buys the standard two-week audit, opportunity map and 90-day roadmap for a single business. It moves toward $6,000 when the audit has to cover multiple business units or systems, when there are existing pilots to assess rather than a blank sheet, or when a data audit is needed alongside the workflow review. The fintech engagement further down was $6,000 for exactly that reason: three stalled pilots to evaluate and a data audit on top.
If you only remember one number, remember this: a working automation for a real business process usually starts around $10,000. Anything advertised well below that is either a template with your logo on it or a proof of concept that will never survive contact with your actual data.
Why “it depends” is a real answer, not a dodge
Buyers hate hearing “it depends,” and they are right to be suspicious of it. But AI automation genuinely varies more than, say, building a marketing website, because most of the cost lives in parts you cannot see from the outside.
Two projects can look identical in a sales call and differ by 3x in price. A support-ticket classifier for a company with clean, well-labeled historical tickets is a two-week job. The same classifier for a company whose tickets live in three systems, half of them free text and none of them tagged, is a six-week job with a data-cleanup phase in front of it. Same demo. Very different invoice.
So when a vendor gives you a firm price in the first fifteen minutes, that is not confidence. It usually means they have padded the number to cover their own unknowns, and you are paying for their uncertainty.
The five ways AI work gets priced
Almost every AI automation quote you receive will use one of these five models. Knowing which one you are looking at tells you where the risk sits.
Fixed price works when the vendor has shipped the exact thing many times. A productized strategy sprint or a standard chatbot setup can be fixed-price because the scope is genuinely known. We fix-price our strategy sprint at $3,500 for that reason: we have run it dozens of times and we know what it takes.
Capped budget is the right model for custom builds. You agree on a ceiling before work starts, then get billed against actual hours up to that cap. You never pay more than the cap without signing off on new scope, and if the work comes in under, you pay the lower actual number. This gives you the protection of a fixed price without the 30 to 40 percent padding that vendors add to fixed quotes for novel work.
Monthly retainer suits ongoing work, like an embedded team that flexes across several projects. Our embedded squad starts at $15,000 per month with a three-month minimum. A retainer is still a cap, just a recurring one.
Pure hourly, with no ceiling, is the model to avoid. Time and materials with no cap punishes you when the work turns out to be hard and rewards the vendor for slow going. If someone quotes you an hourly rate and nothing else, ask for a cap.
Per-outcome or per-seat pricing shows up mostly with SaaS-style AI products rather than custom builds. It can be fair, but read the fine print on what counts as a billable “outcome.”
What actually drives the cost
When we scope a project, four things move the number more than anything else. None of them is the AI model.
Data readiness is usually the biggest lever. If your data is clean, accessible through an API, and reasonably consistent, you skip an entire phase. If it lives in PDFs, screenshots, and a legacy database nobody has touched since 2015, expect a data-preparation phase before any AI work begins. On our logistics document project, the hard 30 percent that off-the-shelf tools could not handle was the whole reason the client needed a custom pipeline. That messy long tail is where the hours go. Our readiness checklist scores this dimension in about an hour, and it is the cheapest way to find out which side of that line you are on before you ask for a quote.
Integrations come next. Connecting to one modern system with a documented API is cheap. Connecting to a 12-year-old ERP, a mortgage loan origination system, or a hospital scheduling tool with no public API is not. Every integration is a small project of its own, with its own edge cases and its own way of breaking.
Compliance and data handling raise the floor. HIPAA, financial data, or personally identifiable information all require extra engineering: anonymization steps, retention controls, deploying inside your own cloud, signing a data processing agreement. This is work that never shows up in a demo but always shows up in a real production system.
How much of the work is new is the last big factor. A build the vendor has done five times before is cheaper and less risky than genuinely novel work. This is also why the same vendor might fix-price one project and cap another: the fixed one is repeatable, the capped one is not yet. It is also why the buy-versus-build decision moves the number so much, which we work through step by step in our build versus buy guide.
Real numbers from real projects
Averages hide more than they reveal, so here are actual figures from builds we have delivered, including what they cost and what they returned.
A Canadian mortgage brokerage was spending eight or more hours of doc prep per loan file. We rebuilt their intake to pull documents from email threads, parse them into the loan origination system, and score discovery calls. File prep dropped from about eight hours to roughly 25 minutes. Eight weeks against a $48,000 cap, came in at $44,200, about 8 percent under.
A mid-market logistics SaaS had a 14-person team manually reviewing incoming shipping documents. We built an LLM pipeline that cleared the backlog in the first week and pushed straight-through processing from zero to 89 percent. Average processing time went from 14 hours to 1.8 hours. Seven weeks against a $77,000 cap, landed at $72,000.
A DTC fashion brand launching in six new markets could not write product copy fast enough. We built a custom content engine with brand-voice tuning, multi-language generation, and a human review queue. Time per product listing fell from 47 minutes to 9, and the content team stopped writing and started reviewing. A larger product build: 14 weeks, an $85,000 cap, $79,500 actual.
A Series A fintech had board pressure to “do AI” and three stalled pilots. A two-week strategy sprint at $6,000 identified roughly $250,000 of misdirected spend to stop and a $400,000 priority backlog worth building. The sprint cost less than one month of a single senior engineer’s salary.
The pattern across all of these: the cap held, the actual came in under, and the return was measured in hours and headcount, not vague “efficiency.”
What a team actually costs per hour
If you would rather reason from rates than packages, the senior roles that do the work on an AI build run $75 an hour for a senior AI specialist, $60 for a delivery lead and $55 for a senior engineer. We publish the same rates for everyone, with no separate enterprise price ladder. The full card is on our pricing page, including designer, mid and junior rates.
Two things about that card matter more than the numbers. Mid and junior engineers are used only on product-development work and only on parts that genuinely do not need a senior, like a dashboard or a UI edit, and they never work an engagement alone; a senior is on every meaningful decision. That is the constraint that makes a blended rate honest rather than a way of quietly staffing your project cheaply.
A typical six-month product engagement blends these: one senior engineer, one senior AI specialist, a part-time senior designer, a delivery lead, a couple dozen hours of founder advisory, and some mid and junior engineers on the parts that do not need a senior. The point of a blended team is that you are not paying a specialist rate to build a dashboard.
Red flags in AI automation pricing
A few pricing patterns should make you walk away, or at least ask hard questions.
Open-ended time and materials with no cap is the most common trap. It sounds flexible. In practice it means the vendor carries none of the risk and you carry all of it. Ask for a ceiling.
Padded fixed price on novel work is the opposite failure. If a vendor firmly fixes the price on something genuinely new, they have priced in every worst case, and you are paying for problems that will probably never happen. A cap gives you the same certainty without the tax.
Change-order theater is when a vendor underbids to win, then bills every clarification as new scope. A fair vendor treats anything they missed in scoping as their problem, not a change order.
Vague deliverables are the quiet one. If the proposal does not say what you will own at the end, code, models, prompts, and documentation, assume you will own less than you think.
Pricing is only one of the things worth interrogating before you sign. Our ten-point vetting checklist covers the other nine, including who actually does the work and what happens to the system after hand-off.
How to budget without writing a blank check
The model to ask any vendor for is a hard cap set before work starts: a ceiling in the statement of work before anyone signs, billing against actual hours at published rates, burn-rate updates as you approach it, work stopping at the cap rather than quietly passing it, and the right to pause at a phase boundary paying only for what was delivered. The five rules we run spell out how that works in practice.
The part worth testing on any vendor is the proof rather than the promise. Of the four projects above, three ran against a cap and all three came in under it: $44,200 against $48,000, $72,000 against $77,000, $79,500 against $85,000. Ask a vendor how often their projects land under cap, and ask for the numbers. A cap that has never been beaten is a fixed price with better marketing.
Frequently asked questions
How much does it cost to automate a single business process with AI? Usually between $10,000 and $30,000 for a build that goes live in your real stack, depending on data quality and how many systems it touches. If you are not sure which process to start with, a $3,500 strategy sprint pays for itself by killing the bad-fit ideas before you spend on them.
Is it cheaper to build in-house? Only if you already have senior AI engineers with spare capacity, which most companies under 500 people do not. A senior AI engineer’s fully loaded cost, salary plus benefits, equity and overhead, runs into six figures a year before they have shipped anything and before you have paid a recruiter or spent the months to hire. That is more than most of the builds above, for one person, and real systems rarely need just one skill set. In-house becomes the cheaper option once you have enough continuous AI work to keep a team busy every week, not before.
Why are some AI automation quotes so much lower? Low quotes usually buy a template or a proof of concept, not a production system. The gap between a demo that works on stage and a system that survives your messiest 30 percent of real data is where most of the real cost and most of the failed projects live.
Do we own what you build? On our engagements, yes: code, models, prompts, and documentation are all yours. Always confirm this in writing with any vendor before you sign.
Before you ask anyone for a quote, do two things. Score your data readiness, because it is the single biggest lever on the number and you can assess it yourself with the readiness checklist. And write down which systems the automation has to write into, because integrations are the second biggest lever and the one vendors underestimate.
With those two answers, any competent vendor can give you a real range instead of a shrug. Our pricing page has the full card and the cap rules if you want to check a quote against published numbers. Sometimes the honest answer is that a workflow is not worth automating yet, and we would rather tell you that than sell you a build.
Last updated