AI integration costs: what your quote should actually include

An AI integration budget should cover discovery, data preparation, connections to existing systems, evaluation, deployment and the people who will keep it working. There is no honest universal price before those decisions are clear. A chatbot that drafts an internal note and a system that changes customer records are different purchases, even when both have a friendly little sparkle icon.

The question I would start with is: what needs to happen after the model produces an answer? If the answer still has to be copied, checked, approved and entered elsewhere, those steps belong in the proposal. Otherwise you are pricing a demonstration and hoping an operating process arrives with the invoice.

Start with the unit of work you are buying

Write one sentence that starts with a trigger and ends with an accepted result. For example: a supplier sends a delivery email; the system extracts the relevant fields; a person reviews exceptions; the approved record reaches the existing tool. This is an illustrative workflow, not a client case study.

Then name the exceptions. What if the attachment is unreadable, a customer appears twice, the destination system is unavailable, or nobody owns the approval? Each answer changes the amount of work. The word “integration” becomes useful only when both ends of the connection have names.

A worked budget, with the assumptions visible

The following is a fictional planning example for one bounded internal workflow with one source and one destination system. The hourly rate and effort are assumptions for arithmetic, not my price list, a market average or a quotation. Taxes, separately purchased licences and unusual migration work are excluded.

The planned work plus that reserve is €12,960. A reserve is not permission to spend it automatically. Agree what would trigger its use and who approves a scope change. A missing access decision can consume more calendar time than the original coding estimate without creating any useful feature.

The monthly cost does not stop at the model

Imagine another explicitly illustrative month: €300 for tools and infrastructure, 12 hours of human review at an assumed €35 loaded hourly cost, and four hours of maintenance at the example €90 rate. That is €300 + €420 + €360 = €1,080 per month. With the project reserve included, twelve months of operation would bring this example to €25,920. Actual demand, support terms and staffing can change every input.

AWS’s Generative AI Lens highlights model choice, resource limits and workflow stopping conditions as cost controls. My buyer-side question is simpler: what prevents one difficult item from quietly consuming a large amount of time or usage?

The FinOps Foundation’s unit-economics guidance connects technology spending to the outcome it supports. For this fictional budget, €1,080 divided by 3,000 accepted items is €0.36 per item. At 1,500 accepted items it is €0.72, if the same monthly total still applies. Count accepted outcomes consistently; raw model calls are not completed customer work.

Five questions that expose a vague proposal

1. What proves that it works? Ask for examples of acceptable output and a representative test set, including failures.
2. What am I expected to provide? Access, usable source material and a decision maker are dependencies, not details the supplier can conjure.
3. What happens when it is unsure? A manual queue with an owner is a design choice. Silent guessing is an expensive surprise.
4. What is included after launch? Separate fixes, monitoring, new requests and changes caused by outside systems.
5. What can we stop after a pilot? Agree a decision point before committing to a wider rollout.

Cheap and expensive are incomplete descriptions

A small scope can be excellent value if it removes a painful, repeated step and your team can operate it. A large proposal can also be justified when several systems, access boundaries and teams are involved. The useful comparison is the same outcome, under the same assumptions, with the same responsibilities included.

Do not turn every hour of theoretical time saved into a cash saving. People may use the recovered capacity to improve quality, respond faster or catch up with existing demand. That can be valuable without reducing a salary bill. The distinction matters when someone asks the project to pay for itself.

Bring a workflow, not a wish list

A useful first conversation needs the current process, approximate volume, systems involved, common exceptions and the person who can accept the result. No sensitive customer files are needed for an initial note. An anonymised example and an honest description are enough to begin scoping.

My experience connects product decisions and hands-on delivery, including integrations at RobosizeME and AI automation at Lexicon Labs. You can explore my AI integration work and read the professional background. If the proposed project cannot yet be described clearly, clarification may be the first useful piece of work. It is usually cheaper than building the ambiguity.

Questions people ask

How much does AI integration cost for a business?

There is no single reliable price without scope. Estimate discovery, source-data work, system connections, testing, deployment and operating ownership separately. The example in this article uses deliberately fictional inputs so you can replace them with your own.

What should an AI implementation quote include?

Look for a defined workflow, assumptions, acceptance criteria, access dependencies, evaluation, exception handling, handover, support boundaries and a clear change process. Ask which licences and recurring costs are separate.

Can I begin with a small AI integration pilot?

Yes, when one workflow can be tested without committing the whole organisation. Agree the representative inputs, acceptable output, owner and stop-or-expand decision before beginning.