The quote you get for an AI tool or build is the visible part of the cost. Data preparation, integration, review time, training and monthly running costs usually decide what you actually spend.
If you run a firm of a few dozen people and someone has handed you a number for an AI project, this is for you. The number may be honest. It just tends to cover the part the vendor controls, and leaves out the parts that land on your team.
Data preparation
AI works on the information you give it. If your customer records are split across a CRM, a shared drive and someone's inbox, somebody has to gather, clean and label them before any model can use them well. This work is slow, it needs people who understand the business, and it is easy to leave out of a quote because the vendor cannot do it without you.
Ask early: which data does this need, where does it live today, and who on my team will get it ready?
Integration with the systems you already run
A tool that sits on its own gets used for a few weeks and then forgotten. To stay useful it usually has to read from and write back to your accounting system, ticketing tool or document store. Each connection is real engineering work, and it has to keep working when either side changes.
Human review
Most business uses of AI still need a person to check the output, at least at the start. That checking time is a cost. If a tool drafts replies that each take two minutes to review instead of five to write, you have a saving. If staff end up rewriting most drafts, you have added a step. Measure review time in a short trial before assuming the saving.
Training and change management
People need to learn when to trust the tool, when to override it, and what to do when it gets something wrong. Someone has to write that down, show the team, and answer questions for the first few weeks. Skip this and adoption stalls, which makes every other cost wasted.
Running costs
Many AI services charge by usage, so the monthly bill grows with volume. On top of that there is hosting, monitoring, and the occasional fix when results drift or a provider changes its model. Ask for an estimate at your expected volume, and at double that volume, so you know how the bill behaves as you grow.
Questions to put to any vendor
- What data do you need from us, in what shape, and who prepares it?
- Which of our systems does this connect to, and is that work in the quote?
- How will we measure accuracy, and what review do you expect our staff to do?
- What will it cost per month at our current volume, and at twice that?
- Who owns the configuration, prompts and data if we stop working together?
A vendor who answers these plainly is giving you a real number. One who cannot is giving you a starting price.
Keep the first version small
The surest way to control the cost of AI implementation is to start with one narrow job where the benefit is easy to measure: sorting inbound requests, extracting fields from a single document type, drafting one kind of reply. Run it on real work for a fixed period, count the hours saved after review, and compare that to the full cost above. If the numbers work at small scale, expand. If they do not, you have lost little.
This is also where you learn whether your data and processes are ready. Many projects that look like AI problems turn out to be data or workflow problems, and fixing those first is often cheaper and more useful than any model.
If you have an AI quote on your desk and want a second view on what it leaves out, I can walk through it with you. Book a conversation with me at daks.me.
