What Does AI Cost? The Cost Structure Explained for SMEs

Why There Is No "AI Price" – Only a Cost Structure

Asking "What does AI cost?" is as precise as asking "What does a vehicle cost?". The answer depends entirely on what you want to solve, how much you contribute yourself and how long you run it. So AI cannot honestly be answered with a price tag, only as a cost structure: a manageable set of cost blocks that appear in every project – with varying weight.

For an SME that is precisely the good news: once you know the cost blocks, you can assess any quote, spot hidden items and build a defensible budget – whether the outcome is a CHF 30 subscription or a project spanning several months. The sections below break AI costs into their components and show what to watch when budgeting.

One-Time vs Recurring Costs

The most important line runs between one-time investments and ongoing costs. One-time costs make AI usable in the first place; recurring costs arise for as long as you use it. Many SMEs see only the first block – and are surprised by the running costs. Always calculate over three years (total cost of ownership), not just the entry price.

• One-time: use-case analysis, concept and tool selection
• One-time: integration into existing systems (CRM, ERP, website)
• One-time: preparing and cleaning your own data
• One-time: staff training and rollout support
• Recurring: licence or subscription fees per user or team
• Recurring: usage-based costs (tokens, API calls, compute)
• Recurring: maintenance, updates, monitoring and support

Licence or Custom Build – the Central Decision

The biggest cost lever is the question: buy a finished product or build your own? Both paths are legitimate, but their cost profiles are opposite. A standard SaaS tool (an AI assistant, a transcription service, a chatbot platform) has almost no start-up cost but predictable monthly fees – you rent the capability. A custom build or deep integration shifts the weight forward: high one-time cost, in return for control, fit and generally lower marginal cost per use.

• Off-the-shelf licence makes sense when: standard task, fast start, small team, unclear value (test first)
• Custom build makes sense when: unique process, sensitive data, high volume, long-term core function
• Often cheapest: start with standard tools, customise only once value is proven
• Watch lock-in: check how expensive a later switch or data export becomes

Hidden Costs SMEs Regularly Underestimate

The licence fee is in the quote – the most expensive items often are not. The bulk of the effort rarely sits in the tool itself, but in embedding it cleanly into daily work and keeping it there. Budget these items from the start rather than meeting them later as a surprise.

• Data work: collecting, structuring, cleaning and maintaining the data the AI works with
• Change management: staff time to learn, resistance, an early productivity dip
• Quality assurance: human review of AI outputs, corrections, feedback loops
• Integration and interfaces: connecting to existing software is rarely free
• Compliance and legal: revFADP review, data protection impact assessment, vendor contracts
• Scaling: costs rise with usage – successful pilots get more expensive at rollout
• Exit: effort for data export, migration and onboarding to a successor system

Calculating ROI Realistically, Not Hoping

Costs only make sense against value. The return on investment of an AI application usually comes from time saved, errors avoided, faster response or added capacity without new hires. Calculate conservatively: estimate the measurable monthly value, subtract all running costs and check after how many months the one-time investment is recovered. If payback lies beyond twelve to eighteen months, a critical look is warranted.

• Quantify value: hours per week times internal hourly rate, plus avoided error costs
• Start small: one clearly measurable use case as a pilot before investing broadly
• Name non-monetary value but keep it separate: quality, speed, employee satisfaction
• Add a buffer: first estimates are usually too optimistic – plan for a learning curve

A Realistic AI Budget in Seven Steps

• 1. Problem first: which concrete process or bottleneck should AI improve?
• 2. Quantify today's cost of the status quo (time, errors, lost revenue)
• 3. Assess off-the-shelf vs custom build – for the chosen use case
• 4. Estimate both cost blocks: one-time and recurring, over three years
• 5. Add hidden items: data, training, integration, QA, compliance
• 6. Check payback: weigh monthly value against total cost
• 7. Budget as a pilot, measure, then decide on expansion

Swiss Perspective: revFADP, Data Location and CHF

For Swiss SMEs, AI has its own cost dimension: data protection and data location. Anyone processing personal data through AI falls under the revised Data Protection Act (revFADP/nFADP), supervised by the FDPIC. This creates real costs – for instance to check whether a provider processes data outside Switzerland or the EU, for data processing agreements and, in sensitive cases, for a data protection impact assessment. These items are not a luxury but part of the base calculation.

Also Switzerland-specific: prices from foreign providers are usually billed in US dollars or euros. Exchange rate, foreign-currency fees and VAT on services from abroad belong in the CHF calculation. Accounting for these details early avoids budget deviations and leads to a decision that also holds up to the regulator.

• Clarify data location: where is your data stored and processed?
• Check the provider's data processing agreement and revFADP compliance
• Factor foreign currency, exchange risk and import VAT into CHF
• Handle sensitive data (health, job applications) with extra care – impact assessment if needed