What does an AI project cost? The parts that set the price

Why there is no honest price list

It is the Friday after Thanksgiving. The office is quiet, and the owner is finally working on next year's budget. One row in the spreadsheet says AI automation, and the cell next to it is empty. On the desk sit two quotes. One is a single number on a single page. The other runs three pages and splits the work into parts. As written, they are hard to compare, and neither one says what the business will actually spend over the year.

Answers online run from a monthly software subscription to a full custom build. Both can be accurate, because they describe different work. The useful question is which parts of your work set the price.

This guide has no dollar figures on purpose. Provider rates change, and a number without your scope attached would mislead more than it helps. Instead, it lists what moves the price, so you can read any quote and know what is in it.

What sets the one-time build price

Scope comes first. One workflow, such as sorting incoming service requests or drafting follow-up on open estimates, has a clear start and finish. A connected department, where intake, quoting and scheduling share information, has more handoffs to design and test. A custom application, a screen your staff or customers use every day, adds design, access control and support. Our guide "One workflow, a connected department or a custom application?" covers how to choose. Each step up adds work.

Next is the number of systems the work has to touch: your phone system, email, scheduling or CRM software, accounting software, perhaps a supplier portal. Ask whether each one offers a way for other software to connect to it, often called an API. Some do. Some only let you export a spreadsheet, and a few only show the data on a screen. Every connection built around a limit like that costs more to build and more to keep working.

Then look at the state of your data and the shape of the work. Customer records split across two systems, a price list that lives in one person's head, or job notes typed three different ways all need sorting out before a system can use them. Exceptions matter just as much. A workflow where most requests look alike is simpler than one full of special cases, because each exception needs a rule, a person to send it to, or both.

• Scope: one workflow, a connected department or a custom application.
• Connections: how many systems, and whether each allows a direct connection or only an export.
• Data: how complete and consistent your records are today.
• Exceptions: how many cases fall outside the usual pattern, and who handles them.

Review steps, testing and training belong in the build

Every point where a person approves the work before it goes out has to be designed: what that person sees, how they approve or correct it, and what happens when they are out that day. These steps are worth paying for. A person on your team should keep the decisions that carry judgement, money or a customer relationship. The system's job is to make those decisions easier to review, not to skip them.

Testing on real examples takes real effort too. A system that handles a tidy demo can stumble on last spring's actual emails, a handwritten work order, or the customer who answers a yes-or-no question with a paragraph. Ask how many of your own past examples the vendor will test against, and who on your team checks the results.

Then budget for training. Your office manager and the people who handle exceptions need to know what the system does, what it hands to them and how to pause it. Their hours are a real cost even when no invoice shows it.

What drives the monthly bill

Some costs arrive every month and move with your volume. AI model providers bill API usage, the kind a custom system makes, by the token, a small chunk of text. Anthropic's pricing page, for example, lists its API rates per million tokens, with separate rates for the text sent in (input) and the text the model writes back (output). Longer documents, longer replies and more requests all add tokens.

Phone and text services bill by the unit as well. Twilio's US pricing pages, as one example, show calls billed per minute and texts billed per message segment, so one long text can count as several. The texting page also lists carrier fees per message, a monthly charge for each phone number and registration fees for business texting. A system that answers calls and sends reminders has a bill that follows your volume. In South Florida that bill can climb when seasonal residents return and the phones get busy, so ask what your busiest month looks like, not your quietest.

Where the model runs changes the bill too. A hosted model from a large provider is billed by usage. A private model on dedicated hardware or a private cloud environment shifts the cost toward hosting and upkeep, which you pay whether the system is busy or quiet. Privacy rules add their own requirements. A medical or dental practice covered by HIPAA generally needs a signed Business Associate Agreement with each vendor that creates, receives, keeps or sends patient information on its behalf. Not every provider offers one on every plan, which can narrow your choices and change the price, so ask your attorney which vendors need one before any patient information moves.

The work after launch is part of the cost

A system that goes live still needs someone watching it. Prices change, a software update moves a field, a new service gets added, and providers retire the models that systems were built on. NIST's AI Risk Management Framework, a voluntary guide, treats this as continuing work: its Manage function includes monitoring plans after deployment and processes for tracking, responding to and recovering from incidents and errors. Our article "What happens after your AI system goes live?" walks through what that work involves.

This is where a low quote can mislead. A one-time price with no monthly line may leave out monitoring, corrections, changes as your business changes, and the upkeep of every connection. Those costs do not go away. They come back later as a surprise invoice, or as a system nobody maintains that slowly stops being useful.

How to read a quote before you budget

A useful written quote separates one-time costs from monthly costs. One-time costs cover discovery, design, connections, data cleanup, testing and training. Monthly costs cover management and monitoring, hosting, software subscriptions and usage. The quote should also say who pays each usage bill. Some vendors pass AI, phone and text charges through on their own invoice. Others have you open accounts in your own name and pay the providers directly, which makes those accounts easier to keep if you ever change vendors. Either can work, as long as you know which.

For next year's budget, fill that empty cell with several lines instead of one: the one-time build, monthly management and hosting, usage estimated from your own busiest month, and your staff's time for review and training. ProcessRoot AI's pricing page shows one way to structure an engagement: a diagnostic, a scoped build and one monthly management charge, with the number agreed in writing before work starts. Our guide "Hiring an AI company? 12 questions a business owner should ask first" covers data, ownership and the off switch. The questions below are only about price.

• Which costs are one-time, and which are monthly?
• Which usage bills are included, and who pays the providers?
• What will the monthly bill look like in our busiest month?
• What does the price assume about our data and our current software?
• How many of our own examples will you test on before launch?
• What counts as a change, and how is a change priced?
• Who monitors the system after launch, and is that in the monthly price?
• What is not included?