A consultancy owner I spoke to recently signed up for an AI-powered helpdesk tool. The sales page showed a £49-per-seat monthly charge, which she read as the full cost. Two months in, her invoice included a second line: AI usage charges of £310. She had assumed the seat fee covered everything, but it covered only access to the platform. The AI processing, every query her team ran through it, was billed separately on a per-use basis.
She wasn’t caught out by anything unusual. This is how the pricing works for many of the AI tools available today, and plenty of owner-managed businesses find out about the variable charge when the first invoice arrives.
What is AI consumption pricing?
AI consumption pricing means you pay in proportion to how much the AI processes, not a flat monthly fee. The common unit is tokens, small chunks of text roughly three to four characters each, used by providers like OpenAI and Anthropic. Other services measure by API calls or items processed. The unit varies, but the underlying logic holds: the meter runs when the AI does.
The two most widespread models you’ll encounter are direct API access and SaaS tools with AI features built in. If your systems connect directly to OpenAI, Anthropic, or Azure OpenAI through an API, you’re billed per token by the model provider. If you’re using a CRM, helpdesk, HR platform, or accounting tool with AI features, the product typically bundles a seat or platform fee with a separate variable charge for AI usage. The two lines can appear on the same invoice, with the variable charge listed as “AI usage”, “credits”, or “processing fees”.
At the individual model level, prices can look very small. GPT-4o mini, one of the more capable affordable models, is currently priced at $0.15 per million input tokens by OpenAI. A typical email is a few hundred tokens. For occasional, low-volume use, the per-use charges are often negligible. The issue arises when AI runs through high-volume, repetitive workflows.
Why does AI consumption pricing catch businesses off guard?
A standard software subscription is predictable: you know the cost in January and it stays the same in December. Consumption-based AI pricing moves more like a utility bill. Your usage can change significantly from month to month as the business adopts new workflows, brings more staff to the tool, or runs a high-volume process through an AI feature for the first time.
Spendesk analysed card spend from 4,300 UK and European SMBs and found that per-company AI tool spending in the UK grew 5.9 times between 2022 and 2024, from an average of around €195 per year to over €1,100. That growth is driven by usage, as firms move AI into more workflows, not by seat-price increases.
The second reason businesses are caught off guard is that implementation costs are rarely quoted alongside consumption charges. Softrobo’s analysis of UK SME AI deployments found that platform licences account for only 30 to 50 per cent of total AI implementation spend. The rest goes on integration, configuration, training, and change management. A consumption charge that looks manageable on a pricing page can sit inside a project that costs considerably more to run properly.
Where will you actually run into AI consumption pricing?
You’ll encounter it in three main places. Developer APIs, where your systems connect directly to OpenAI, Anthropic, or Azure, are billed per token with no cap unless you set one. SaaS tools with AI features charge a seat fee plus usage on top. Usage-tiered plans step the price up once you pass a volume threshold.
Any workflow that could scale is the one to watch. Customer queries handled by a chatbot, documents processed for data extraction, first-draft proposals generated at volume, job applications screened automatically. In each case, the underlying AI processing is almost certainly metered, and the cost scales with the number of items.
All-inclusive licences do exist. ChatGPT Team and Microsoft 365 Copilot both bundle AI usage within fair-use limits, which removes the per-token exposure at the cost of a higher seat price. Whether that’s better value depends on your usage volume. At lighter use, a flat-fee licence often costs more per query. At high volume, it becomes cheaper and simpler to manage. The pricing page rarely tells you which side of that threshold your business sits on, which is why asking for a sample costing at your expected volume matters before you sign.
Brightmine case studies of UK SMEs show what adoption at scale looks like: one UK e-commerce business used an AI chatbot to handle 70 per cent of customer queries and saved over £50,000 annually in staffing costs. The value was real, and so was the per-conversation meter running behind it.
When does consumption pricing matter, and when can you set it aside?
Consumption pricing deserves attention when a workflow could scale quickly. A document-processing tool handling ten invoices a day is a different proposition from one handling 10,000. For steady, low-volume use, the charges are typically modest enough not to require detailed management. The workflows to watch are those connected to a process that could spike: customer queries, proposal generation, or data extraction at volume.
A useful cost benchmark: Wingenious costed processing 100,000 documents using AI at approximately £170, compared with £8,000 manually. The saving is significant. So is the exposure if that volume runs unexpectedly through a consumption-priced API that wasn’t in the budget.
The regulatory dimension is separate from the cost question. The ICO’s guidance on AI and data protection is clear that if personal data flows through an AI API, you’re acting as a data controller and the AI vendor is your processor. That requires a written agreement, clarity on whether data is used for model training, and in some cases a Data Protection Impact Assessment. These obligations apply regardless of how the tool is priced or how small the monthly bill appears.
For FCA-regulated firms, the obligations go further. The FCA’s 2022 discussion paper on AI and machine learning treats many AI services as third-party arrangements subject to outsourcing and operational resilience rules. A low per-token rate doesn’t remove the due diligence requirement.
What else connects to how you pay for AI?
Consumption pricing sits alongside three questions worth thinking through separately. Whether your own pricing model still works once AI shortens your delivery time is one. Whether switching costs will trap you with a single provider as workflows bed in is another. The third is how you’ll track AI spend as a distinct cost category, separate from general software, for both budgeting and client billing purposes.
The pricing model question is live for professional services firms. Analysis by SCOPE Better, reported by Consultancy.uk, found that none of the consulting firms surveyed had a clear plan for how AI would affect their revenue model, and the analysis suggested at least 20 per cent of professional services firms would need to fundamentally change how they price work within five years. If you’re billing by the hour and AI halves the delivery time, the consumption charges you pay become a visible line item in a calculation that previously didn’t need to exist.
The NCSC’s guidance on AI cybersecurity adds a third dimension. Your AI provider is part of your digital supply chain, and the NCSC expects you to assess them for security, data handling, and resilience standards. The question of which AI tools your firm uses, and what your contractual rights are over your data, is a governance question alongside a cost one.
The practical starting point is straightforward. Set a monthly consumption budget per tool, request usage dashboards from vendors before signing, and ask for a sample costing at your expected volume. Many platforms, including OpenAI, Azure, and Anthropic, allow you to set hard usage limits or billing alerts. That ten-minute conversation with a vendor prevents a three-figure line appearing unexpectedly on next month’s invoice.
