What does generative AI really cost? Licences, tokens and hidden costs explained

Generative AI costs a fashion company far more than the licence price. The total includes per-user subscriptions, usage-based token fees for automated workloads, integration and development, data preparation, and ongoing governance, training and review time. Licences are the most visible line, but integration, data work and human review frequently cost more over the life of a use case.

What are the main cost components of generative AI?

How much do enterprise AI assistant licences cost?

Per-seat pricing is the easiest part to compare, though it changes often and should always be checked on the vendor's page. Microsoft lists Microsoft 365 Copilot at 30 US dollars per user per month paid yearly, or 31.50 dollars paid monthly with an annual commitment, and it requires a separate qualifying Microsoft 365 licence. Google bundles Gemini features into most Google Workspace business plans, with full Gemini in Docs, Sheets, Slides and other apps listed for the Business Standard and Plus tiers and more limited access in Business Starter.

Two points matter more than the headline price. First, which employees genuinely need a paid seat: rolling licences out to everyone often leaves many seats idle. Second, what the bundle includes, such as connectors to company data, agent features and admin controls, which may otherwise need to be built or bought separately.

Read also

AI for fashion CFOs: where the money is and how to govern it

How does token-based pricing work?

When AI runs inside a company's own systems, for example generating product descriptions for thousands of SKUs or powering a customer service chatbot, it is usually billed through an API per million tokens (a token is a fragment of a word). OpenAI's published API price list, checked in October 2026, shows the pattern: its flagship model costs 5 dollars per million input tokens and 30 dollars per million output tokens, while its small model costs 0.20 and 1.20 dollars respectively. Cached input is cheaper, and the Batch API offers a 50 percent discount for jobs that can wait up to 24 hours.

The practical implications for fashion teams:

• Output is the expensive part, so long generated texts cost more than long prompts.
• Bulk tasks such as translating a season's catalogue overnight suit batch pricing.
• Smaller models are often good enough for structured tasks such as attribute extraction, at a fraction of the cost.
• Agents that call models many times per task can multiply token consumption, so test costs on realistic volumes before launch.

Image generation follows similar logic but is usually priced per image or per credit rather than per token. For campaign visuals or product imagery, the more relevant comparison is with photography, retouching and studio costs, which means the business case depends on how many generated images are actually usable without rework, and on the legal checks needed before publication.

What are the hidden costs of generative AI?

Gartner warned in July 2024 that generative AI deployment approaches come with significant costs ranging from 5 million to 20 million US dollars, and that these costs are less predictable than for other technologies, depending on use cases, spending levels and deployment approach. It also predicted that at least 30 percent of generative AI projects would be abandoned after proof of concept, citing escalating costs among the reasons. In June 2025 it again named escalating costs as a reason why over 40 percent of agentic AI projects would be cancelled by the end of 2027.

In fashion the hidden costs tend to cluster in three places. Product data: AI-generated descriptions are only as accurate as the materials, fit and care attributes in the PIM, and cleaning these is labour-intensive. Integration: connecting outputs to e-commerce, ERP or B2B ordering systems often costs more than the AI itself. Review time: someone must check outputs for accuracy, brand voice and legal claims, and that time should be counted against claimed savings.

People costs are the most underestimated item of all. Training sessions, champion time, legal and security reviews of new tools, and the hours managers spend redesigning processes rarely appear in a vendor quote. For a mid-sized brand these internal costs can easily match the licence budget in the first year, and they should be estimated openly rather than treated as free because staff are already on the payroll.

How can fashion companies control generative AI costs?

1. Assign licences by role and review usage quarterly, reallocating idle seats.
2. Match model size to task: use smaller, cheaper models for classification and extraction, larger ones for complex reasoning.
3. Use batch processing for catalogue-scale content work.
4. Set spending limits and alerts on API accounts and agent workflows.
5. Fix product data at source rather than paying repeatedly for AI to work around gaps.
6. Check whether existing suites already include AI features before buying overlapping tools.

Read also

Why do AI pilots stall in fashion, and how do you scale them to production?

How should the cost be weighed against value?

A credible business case compares the full cost (licences, usage, integration, data work, governance and review) with measured improvements against a baseline. The most dependable candidates in fashion are usually high-volume tasks whose outputs can be checked, such as content production, document processing and service handling. For more speculative uses, start small, measure, and only then commit to multi-year licences.