What AI implementation costs, and how to compare quotes

“How much does it cost?” is the first question and the hardest to answer honestly before an assessment. Not because the price is hidden, but because it is driven by things that cannot be seen from a conversation. This article explains what drives it, which costs proposals usually leave out, and how to compare two proposals when their figures are similar.

A sense of scale helps before you read any proposal. A World Economic Forum piece cites Gartner projections putting the AI services market at $609 billion by 2028 and the AI technology market at $512 billion by 2027. Those numbers tell you nothing about your own invoice, but they explain why proposals vary so widely. The market is growing faster than pricing norms are settling. So compare the contents, not the figure.

What drives the price?

The five factors that change the price most. None of them is technology.
Factor	Why it changes the price	What you can do yourself
Scope of the process	One step with a clear input and output is several times cheaper than a process that crosses three departments.	Break the process into steps and start with one.
Number of integrations	Every system to connect to is separate work. Systems without an API cost the most.	Find out whether your programs have APIs before asking for a proposal.
State of the data	Data in tables is cheap. Data in photos, scanned PDFs and free text is expensive.	Collect data samples and show them as they are.
Cost of an error	The more expensive one mistake is, the more checks, approvals and testing are needed.	Decide where a person must approve a decision and where not.
Your team's involvement	If your people do the assessment and the pilot, there is less outside work.	Assign one person who will give time to the pilot.

Which costs do proposals leave out?

Maintenance. A system that connects to other programs stops working when those programs change. Ask who will fix that and at what price.
Model and tool fees. AI models charge per request. A proposal should include a rough monthly figure at your volume.
Changes. Six months in, you will want the system to do one more thing. If the code and documentation are not with you, every change is a new proposal.

One line belongs next to maintenance that did not exist a couple of years ago: observability. Harvard Business Review describes a Salesforce employee whose whole day runs through agent dashboards, scorecards and monitoring. Once a system does more than execute rules and starts making decisions, someone has to see what it decided. That is hours every week, not a one-off task, and it should be named in the proposal.

How do you compare two proposals with a similar figure?

The figures are similar; the contents rarely are. Give both vendors the same six questions and compare the answers, not the numbers.

Will there be a pilot on our real data before the system is built, and can we stop after it?
Where will the system run: in our accounts or yours?
Who will own the code and documentation after the project?
Who on our team will be trained to maintain the system, and is that included?
What will one change cost a year from now?
What will models and tools cost per month at our volume?

A proposal that answers all six in writing is almost always cheaper over two years, even if the first figure is higher.

When does the price not matter at all?

When the benefit has not been counted. If you do not know how many hours a month the process takes now and what one error costs, you cannot say whether a proposal is cheap or expensive. That arithmetic takes one afternoon and is the first thing I do in the assessment stage. Without it, any figure in a proposal is just a figure.

My own prices and pricing model are stated on the service pages.

A note on the data: the market projections and examples cited are international, from 2025 and 2026, and do not reflect Lithuanian price levels.