What Actually Makes
an AI Project Expensive?

Start with how well the work is understood.

If the steps are known, the inputs are consistent and someone can explain what a good result looks like, the implementation has a clear target.

If the process lives mostly in people’s heads, differs from person to person or changes week to week, part of the project becomes defining the work before improving it. That’s valuable, but it’s effort.

A well-understood process is easier—and cheaper—to improve than a vague one.

Work that happens inside one application is simpler than work that moves between several.

Each additional system adds a connection to build, permissions to manage and another place where something can change unexpectedly. Older systems, or systems without a straightforward way to connect, add more.

The AI may only touch one step. The systems determine how much has to be built around it.

AI can only work with the information it can reach.

If that information is accessible, reasonably clean and consistently organized, things move quickly.

If it’s scattered across inboxes, spreadsheets and shared drives—or if different sources disagree—someone has to decide which version is right and how the system should find it.

Data work is rarely glamorous. It’s often where a project finds its real scope.

Not everyone should see everything, and an AI system shouldn’t either.

If the work involves client records, financial details, personnel information or anything regulated, the system needs to respect the same access rules people do—and keep respecting them as roles change.

The more sensitive the information, the more care goes into deciding what the system can read, what it can keep and who can see what it produces.

The normal path is usually the easy part.

Effort tends to build up in the cases that don’t follow it: the document in an unexpected format, the request that fits two categories, the record missing a field.

A dependable system needs a plan for each of those—either handling them, or recognizing them and handing them to a person.

The more varied the work, the more of the effort goes into exceptions.

A system that drafts something for a person to review is very different from one that acts on its own.

As autonomy increases, so does the work required to make it safe: clearer limits, more testing, better monitoring and a reliable way to catch mistakes before they matter.

We look at that spectrum in AI Agents vs. AI Automation: What Does Your Business Actually Need?

More autonomy isn’t wrong. It just needs to earn its cost.

A system nobody uses doesn’t create value.

Implementation includes helping the team understand what changed, adjusting how work is handed off and deciding who owns the process once it’s running—who notices when it drifts and who updates it when the business changes.

Without a clear owner, even a well-built system slowly falls out of step with the work it was built for.

Most of these costs become far more manageable once the work is clearly understood. That’s why the most useful first question is:

Not:

“What AI should we build?”

Once the work is understood, choose the least complicated technology capable of improving it.

Sometimes the answer will be much simpler than an AI agent or a custom application: a well-configured platform the team already has, a straightforward automation, or a single AI step added to an existing process. We explore that decision in Build vs. Buy: When Does Your Business Need a Custom AI System?

Each of these areas can move an implementation up or down. Work with known systems, accessible data, limited sensitivity and predictable steps tends to sit toward the lower end of a range. Work that stacks several complicating factors sits higher.

For the planning ranges FXNL uses at each level of implementation, see How Much Does It Cost to Implement AI in a Business? →

For a rough figure for your own situation, the AI Investment Estimator asks about the work rather than the technology and returns a planning estimate—not a proposal or a fixed scope.

The model is rarely the hard part. The work around it is.