# AI Costs: Why One in Three Companies Underestimates the Bill

> Author: Chris Jon Graf (AI Strategist & CEO)
> Updated: 2026-08-18
> URL: https://ai-outsourcing.ch/insights/ai-costs-why-one-in-three-companies-underestimates-the-bill

## Summary

41% of companies now use AI in production — double last year's share. Yet 33% report higher costs than planned. The reason: budgets cover licences, not the operating costs, data preparation and integration that follow. Swiss SMEs should calculate total cost of ownership upfront, not correct it afterwards.

## The Bitkom Numbers in Plain Terms

Bitkom's 2026 survey put a hard number on what many executives already sensed: 41% of German companies now run AI in production — double the 17% share recorded a year earlier. Another 48% are planning or actively discussing deployment. Adoption is clearly accelerating. But alongside that growth sits a less comfortable pattern: 33% of companies report that actual costs came in higher than originally budgeted.

**41%** — of companies now use AI in production, according to Bitkom 2026 — double last year's 17%

**33%** — report AI costs higher than originally planned

The upside is real: 77% of companies say AI has improved their competitive position, most visibly in text work (71%), marketing (53%), customer service (42%) and data analysis (31%). The technology delivers measurable value. At the same time, 19% of companies report job cuts linked to AI — a sign that automation and cost pressure are more closely linked than most budget plans account for.

## Why the Cost Gap Happens

The reason for the budget surprise is rarely poor judgment — it's a structural blind spot in planning. Most organisations calculate model and licence costs, the line items that appear on a vendor invoice. What's systematically missing are the operating costs that only appear once a system runs in production and scales.

- Inference at scale — per-request compute costs that can run 3 to 5 times higher than initial estimates once usage grows
- Data preparation — cleaning, structuring and maintaining data, which often accounts for the largest share of total effort
- Integration into existing systems — connecting AI to ERP, CRM and line-of-business applications, rarely a trivial task
- Human-in-the-loop review — ongoing oversight and approval processes that remain necessary for sensitive use cases
- Rework — pilots that need to be rebuilt because guardrails or governance were skipped the first time
- Change management — training, process redesign and internal communication

This gap is not confined to any one market. A WitnessAI survey of 300 executives found that 68% experience AI cost overruns, with a third saying it happens 'mostly' or 'always.' Only 9% report measurable ROI on more than three-quarters of their AI projects. Mavvrik's State of AI Cost Governance study paints a similar picture: 62% of organisations had an unexpected AI cost materially alter a business decision, and only 11% can forecast their AI spend within ±10%.

## What AI Projects Actually Cost in Switzerland

For Swiss SMEs, several market analyses point to concrete cost ranges. They vary widely by ambition level, but the pattern is consistent: the initial investment is the beginning of the cost curve, not the end.

- Custom AI agent: CHF 100,000–400,000 in year one, then CHF 30,000–150,000 per year in operation
- Simple chatbot: CHF 3,000–6,000 one-off plus CHF 100–200 per month
- Comprehensive process automation: CHF 15,000–40,000 one-off plus CHF 500–1,000 per month
- First project in three phases (analysis, proof of concept, production rollout): CHF 15,000–50,000 total
- Comprehensive AI project for a 20-50 employee SME over three years: CHF 50,000–150,000 total cost of ownership

> **A Rule of Thumb for Budgeting**
>
> Plan for actual total costs of 3 to 5 times the pure licence price from the outset, and set aside 15–25% of the project budget separately for change management. This ratio holds up consistently across industries — ignore it, and a supplementary funding request becomes almost inevitable.

## The Blind Spots That Blow Up Budgets

Visibility into AI spend is surprisingly low across many organisations. According to KPMG, only 26% of companies have full visibility into their AI costs — the rest operate on assumptions rather than numbers. Flexera puts wasted AI spend at 14%, often driven by unpredictable, usage-based pricing models. And DataRobot's survey of 413 practitioners found that 94% struggle with so-called Day 2 operational issues after launch. 72% have already exceeded their operating budgets, and 71% say plainly: running an AI agent costs more than building it.

Security and compliance are another frequently overlooked cost driver. An EMA survey of 152 IT leaders found that security or compliance requirements delayed AI projects for 82.9% of respondents, and more than 80% had to scale back already-deployed AI agents over security concerns. Address these requirements only at the end of a project, and you pay twice — once for the original build, once for the fix.

## How to Budget for AI Realistically

1. Calculate over three years, not one quarter — total cost of ownership, not purchase price
2. Separate pilot and production budgets, since scaling generates its own cost curve
3. Budget data preparation as its own line item, not as a footnote to implementation
4. Set aside 15–25% of the total budget for change management and training
5. Factor in Swiss hosting and data protection compliance from the outset, not as a later adjustment
6. Build in a monitoring and observability budget to keep cost development visible over time

This kind of reality check matters given current investment levels: Swiss SMEs invest, on average, just 1.2% of annual revenue in digitalisation and AI. That leaves little buffer for cost surprises — one reason many organisations remain cautious in actual usage despite high ambitions.

## A Pilot Is Not Production

A common mistake is projecting the cost of a successful pilot straight onto full production. For custom-built AI agents, the break-even point against a platform solution typically falls between month 22 and 26 — before that, a custom build is usually more expensive, not cheaper. Skip this timeline in your planning, and the make-or-buy decision gets made on the wrong numbers.

> **When the Budget Spirals**
>
> Mavvrik's research shows how sharply unexpected AI costs escalate inside organisations: 40% triggered board-level escalation, 33% led to an emergency spending freeze, and 25% caused an AI initiative to be delayed or cancelled entirely. Realistic cost planning from the start is considerably cheaper than an emergency stop mid-project.

## What This Means for Swiss SMEs

The Bitkom numbers are not a reason to slow down on AI — a 77% improvement in competitive position speaks for itself. They are a reason to take budget planning more seriously than before. Factor in operating costs, data preparation, integration and change management from day one, and you avoid exactly the surprise that a third of companies have already experienced. For a broader look at positioning AI strategy without losing focus in the global race, see [this analysis of AI strategy for Swiss SMEs](https://www.ki-podcast.ch/ki-standort-schweiz-kmu-strategie-und-globales-rennen).

In the end, a realistic cost calculation isn't a brake on AI adoption — it's the precondition for sustainable growth with it. Companies that know their total cost of ownership negotiate better terms, scale more cleanly, and avoid the emergency stops that a good third of companies are experiencing today.

## FAQ

### What does an AI project actually cost for a Swiss SME?

Depending on ambition level, costs range from CHF 3,000–6,000 for a simple chatbot to CHF 100,000–400,000 in the first year for a custom AI agent. A comprehensive AI project for a 20-50 employee SME typically runs CHF 50,000–150,000 in total cost of ownership over three years.

### Why are AI costs so often underestimated?

Because budgets usually cover only model and licence costs while leaving out ongoing operating costs such as inference at scale, data preparation, integration, security measures and change management. According to Bitkom 2026, 33% of companies report higher costs than planned.

### What is total cost of ownership (TCO) for AI projects?

TCO covers all costs beyond acquisition across the full usage period: operations, scaling, data maintenance, integration, security, training and rework. Studies show actual total costs often run 3 to 5 times the pure licence price.

### How much do Swiss companies currently budget for AI and digitalisation?

Swiss SMEs invest, on average, around 1.2% of annual revenue in digitalisation and AI — a budget range that leaves little buffer for unexpected costs.

### How can I avoid cost surprises in AI projects?

Calculate over three years rather than one quarter, separate pilot and production budgets, set aside 15–25% for change management, budget data preparation as its own line item, and factor in compliance requirements from the outset.

## Sources

- [Bitkom Digitalisierung der Wirtschaft 2026](https://www.software-journal.de/2026/08/14/ki-einsatz-in-deutschen-unternehmen-verdoppelt-sich-binnen-eines-jahres-auf-41-prozent-jedes-dritte-meldet-zugleich-hhere-kosten-als-erwartet/)
- [WitnessAI 2026 AI Cost Overruns Report](https://www.cfodive.com/news/7-10-firms-report-ai-cost-overruns/825961/)
- [Mavvrik 2026 State of AI Cost Governance](https://www.cfodive.com/news/1-in-4-companies-delay-cancel-ai-projects-over-cost/827524/)
- [KPMG AI Quarterly Q2 2026](https://kpmg.com/kpmg-us/content/dam/kpmg/pdf/2026/aipulsesurvey-q2.pdf)
- [Flexera 2026 AI Pulse Report](https://info.flexera.com/FLX1-REPORT-AI-Pulse)
- [DataRobot Unmet AI Needs Survey 2026](https://www.datarobot.com/resources/unmet-ai-needs-survey-2026/)
- [EMA State of AI Friction 2026](https://www.enterprisemanagement.com/product/the-state-of-ai-friction-why-enterprise-ai-deployment-is-slower-costlier-and-more-limited-than-expected/)