Cost to Build an Internal AI Tool

What Actually Drives the Cost

"How much does an AI tool cost" is really four separate questions wearing one coat. The technology itself — an API call to a language model — is cheap, often fractions of a cent per request. What you're actually paying for is everything around that call: figuring out exactly what the tool needs to do, connecting it to your existing systems, making sure it works on your real data, and building in enough review that a mistake doesn't reach a customer or a ledger unnoticed.

Four variables move the number more than anything else:

• Scope — one clear task with one input and one output costs a fraction of an open-ended tool meant to handle "whatever comes in"
• Integrations — reading from a spreadsheet is trivial; reading and writing to a CRM, an ERP, or a legacy system with no clean API adds real engineering time
• Data infrastructure and quality — a tool is only as good as what it can see; if your data is scattered across formats, half-structured, or simply missing, cleanup and pipeline work often outweighs the AI work itself
• Compliance requirements — audit trails, access controls, and data residency requirements (Swiss-hosted infrastructure, for instance) add scoping and testing time, not just a checkbox

Cost by Project Type

The ranges below are what we typically see when scoping projects for Swiss SMEs. They're a starting point for a conversation, not a price list — the real number for your project depends on the four variables above.

What Naive Estimates Usually Miss

Most back-of-envelope estimates account for the build — connect this API, write this prompt, ship it. What they miss is everything that separates a demo that works once from a tool your team actually trusts to run unattended. This is where scoped budgets tend to blow past a first quote, not because anyone padded the number, but because the real work wasn't visible until the first version was already in front of real users and real data.

• Prompt engineering and iteration — a prompt that works on the five examples you tested rarely survives contact with the full range of real inputs; getting it reliable takes rounds of testing against edge cases, not one write-and-ship pass
• A human review interface — anything consequential needs a place for a person to see what the tool proposed and approve, edit, or reject it before it ships; building that interface is often underestimated because it isn't "AI work," but it's real engineering time
• Data cleanup — inconsistent formats, duplicate records, and missing fields have to be sorted out before a tool can reliably read them, and this step is routinely left out of first estimates entirely
• Ongoing monitoring and maintenance — someone needs to watch for drift, failed calls, and edge cases the tool starts hitting once it's live, which is a recurring cost, not a one-time build item

What It Costs to Run, Not Just Build

The build is a one-time cost. Running the tool is ongoing, and it's the number most estimates leave out entirely — partly because it's genuinely hard to predict before the tool is live and you can see real usage volume.

Two things make up the monthly bill: the LLM or API usage itself (billed per request or per token, scaling with how much the tool is used), and hosting or infrastructure for anything that needs to run continuously — a monitored inbox, a scheduled job, a database. For a simple tool handling a modest volume of requests, that's typically CHF 50–400 a month. For a more complex tool processing high volumes, calling multiple models, or running several agents in parallel, CHF 300–2,000+ a month is more realistic.

• Simple tool, low volume (e.g., a weekly report generator, a single inbox monitor): roughly CHF 50–400/month
• Moderate tool, regular use (e.g., a daily lead-qualification agent): roughly CHF 200–800/month
• Complex or high-volume system (multi-agent, several integrations, heavy usage): roughly CHF 300–2,000+/month

Build In-House, Buy SaaS, or Hire a Studio?

There are three real paths to an internal AI tool, and each makes sense for a different situation. Buying an off-the-shelf SaaS product is fastest and cheapest up front if a generic tool actually fits your process — but you're renting someone else's roadmap, and if your workflow doesn't match the product's assumptions, you'll spend more time working around its limits than the customization would have cost.

Building in-house makes sense if you already have engineering capacity and the tool is core enough to your business to justify owning it long-term — but factor in that your team is learning agentic AI development on your dime, and the maintenance burden doesn't go away once it ships. Hiring a studio sits in between: faster than building from scratch because the process is already understood, more tailored than SaaS because it's scoped to your actual workflow and data, and — done honestly — handed off in a way that doesn't lock you into permanent dependency on the studio for every small change.

The Honest Answer

These ranges are patterns from scoping conversations, not a rate card — the real number for your project depends on your data, your systems, and what "done" needs to look like for your team to trust it. The fastest way to get a number that means something is to walk through one real process and let us map where the actual cost sits, rather than guessing from a category label like "AI tool."