§ 01 · What’s actually happening

The bill is bigger because the tools got bigger.

There are three structural reasons the small-business AI bill has gone up faster than the published per-token price increases suggest, and they compound on each other in a way that is not obvious month-to-month.

First, the default model is more expensive than it used to be. When I set up an API integration in 2023, the obvious default was GPT-3.5 or an early Claude model. They cost almost nothing per call. In 2026, the realistic default for anything non-trivial is one of the frontier models — Claude Opus 4, GPT-5, or Gemini 2.5 Pro. The per-million-token pricing on these is between five and twenty times higher than the cheap-and-cheerful models everyone started with. Most software tools that integrate AI quietly defaulted to the more expensive model in the last 12 months because accuracy is better. The user does not see this happening. The bill does.

Second, the use cases have expanded without anyone deciding they should. The Berlin consultancy in my opening started with two narrow use cases. By the time she emailed me, the team had added: meeting transcript summarisation, client-email drafting, contract analysis, a code-helper for the one developer in the team, automated tagging on their internal documents, and a research assistant for sales prospecting. Each addition was a sensible decision in isolation. Cumulatively, they multiplied the inference volume by an order of magnitude. This is the typical pattern. AI use does not stay where you put it.

Third, the prompts got longer. Most modern AI workflows now include substantial context — system prompts of several thousand tokens, retrieved document chunks, conversation history, structured output schemas. The actual user question might be a single sentence; the full payload sent to the API is fifty times that. Token-based billing charges for the whole payload. Workflows that worked fine on short prompts in 2023 have become extremely expensive in 2026 not because they got worse, but because the underlying patterns for getting good outputs require feeding the model more context.

When you combine these three effects, a workflow that cost a few cents per call in 2024 reasonably costs a euro or more per call today. Multiplied by the volume a small business now generates, the annual figure stops being a footnote in the software budget and starts being a real line item that the accountant flags.

§ 02 · The enterprise context

The big-company AI cost crisis is real, and it tells you something useful.

If you have not been following the enterprise side of this story, the short version is that very large companies are having a worse version of the problem your small business is having, and they are talking about it openly. ServiceNow’s CIO publicly said in early 2026 that she is working with the CFO to figure out whether the company can continue paying for Claude Enterprise through the rest of the year. Salesforce has committed to roughly $300 million on Anthropic tokens in 2026. Zillow is on track to spend somewhere between $7 million and $10 million on AI services this year — close to half of last year’s net income. Uber and ServiceNow have both burnt through annual token budgets in the first few months of the year and are now in active cost-containment mode.

This matters for the small business reader for two reasons. The first is that it tells you the cost pressure is structural rather than something specific to your setup. If billion-dollar enterprises with dedicated AI cost-engineering teams are blowing through their budgets, your modest operation almost certainly will too if you treat the API line as a fixed cost rather than a variable one. The second reason it matters is that the enterprise market is now actively reorganising around cost discipline, which means that the practical responses are getting better documented and easier to copy. The cost-engineering playbook that was a niche skill 18 months ago is becoming standard practice. Small businesses can borrow most of it.

The four responses I see working at small-business scale are not novel. They are the same four things Stripe’s technical staff and ServiceNow’s engineering org are now doing, scaled down. None of them require giving up the AI tools that have actually become useful.

§ 03 · The four responses

What I have actually seen work.

Response one: stop defaulting to the frontier model. Most workflows do not need the top-tier model. They got migrated to it because the AI vendor’s SDK updated its default, or because the tool you use upgraded its underlying model, or because someone on your team read a blog post and changed a config. For document summarisation, email drafting, tagging, basic question-answering, the cheap model is usually fine. The discipline is to test — run a sample of your typical workflow on the cheaper model and check whether the output quality is genuinely worse, or just marginally worse in a way that does not actually matter for your use case. For most non-coding workflows that the average freelancer runs, the answer is that the cheaper model is fine and the cost reduction is somewhere between 70 and 90 percent. This is the single biggest lever and the one most small businesses are not pulling.

Response two: cache the calls that repeat. A surprising share of the calls a small-business AI workflow makes are near-duplicates. The same kind of email summary, the same kind of meeting note, the same kind of document extraction. Modern AI providers offer prompt caching that costs a fraction of the regular per-token rate when the input is similar enough to a recent call. Most small-business setups do not have caching configured at all because the tool they use to call the API does not enable it by default. Enabling caching where it applies typically cuts 30 to 50 percent off the bill for repeat workflows. This is plumbing work, not magic, and it is worth an hour of someone’s time to configure once.

Response three: be intentional about which provider you use, given what you already pay for. Most freelancers and small businesses already pay for Google Workspace — the same Google Workspace that now bundles Gemini into Gmail, Docs, Sheets, and Meet at no additional per-token cost for the in-app features, and integrates with the Google Cloud API for custom workflows. For the kind of work that small businesses actually do — document drafting, email triage, meeting summaries, knowledge-base search — running it through Gemini via your existing Workspace subscription is often substantially cheaper than running it through Claude or GPT via separate APIs. Not because Gemini is necessarily better, but because the bundled-in usage avoids the metered-API line item entirely for a lot of common tasks. The principle is broader: be deliberate about which provider you route which workflow to, based on what you already pay for and how the marginal cost actually lands.

Response four: buy your API credits at a discount where you can. This is the response most small-business operators have not heard of and the one that requires the least technical work. There is a secondary market for AI provider credits — unused or bulk-purchased credits resold through marketplaces at a discount to the providers’ retail pricing. The discounts typically run between 15 and 40 percent, depending on the provider and the volume. The credits behave identically to credits bought directly through the provider’s billing portal; the only difference is the procurement channel and the price. For a small business with an annual API bill in the low thousands of euros, even a 20 percent discount is real money — the difference between an annoying line item and a manageable one. Cheap Gemini API credits through one of the established marketplaces, for example aicreditmart.com, can land at substantially below what Google charges through Cloud Console for the same Gemini tokens running on the same infrastructure. For a freelancer or small business already running Gemini workflows through their Google Workspace setup, this is the closest thing to free money in the AI cost-management toolkit — same product, same behaviour, lower price.

§ 04 · What not to do

Three responses I see small businesses try that do not work.

Do not just roll back AI usage to where it was a year ago. This is what panicked small businesses do when the bill arrives. They turn off integrations, cancel API access, go back to manual workflows. Six weeks later they realise that some of those workflows were genuinely saving them time, and they turn the integrations back on without doing any of the cost discipline work, and the bill comes back larger. The right move is to keep the workflows that earn their keep and apply the cost-reduction patterns to them, not to abandon the tools.

Do not try to switch to self-hosted open-source models on your own laptop. Every few months I get an email from a freelancer who has spent three weekends setting up a local Llama instance because they read that it was free. It is technically free of inference cost. It is not free of your time, the electricity, the hardware constraints, or the quality gap that becomes painful for any serious workflow. For a small business with an AI bill under €2,000 a month, the math almost never justifies the time investment. Self-hosting becomes worth considering above that level, and even then only for specific workflows where the local model is genuinely adequate.

Do not adopt every new AI feature your existing tools ship. The accounting software, the CRM, the project tracker, the email client — all of them are shipping AI features now and most of them route to a frontier model under the hood at a per-action cost that you are paying somewhere in the subscription bundle. The features are sometimes genuinely useful. They are also sometimes just there because the vendor wanted to advertise AI on the marketing page. Be intentional about which ones you turn on, and pay attention to whether the subscription cost on the underlying tool has crept up to subsidise them.

§ 05 · The honest bottom line

The AI line is now a real expense category. Treat it like one.

For the last three years, AI API spend has been small enough that it lived in the “miscellaneous software” line of most small-business P&Ls, unwatched and unmanaged. In 2026 that has stopped being true for any business that has actually integrated AI into its workflow. The line item is now real, it is variable, and it grows with usage in ways that are not obvious until the quarterly review.

The good news is that managing it does not require giving up the tools or learning new skills. It requires treating the AI line the way you already treat payment-processor fees, accounting software, or business banking — as a procurement category where supplier choice, configuration, and the right purchasing channel meaningfully change the number. Run the cheaper model where it works. Cache the repeats. Be intentional about which provider you route which workflow to, especially given what you already pay for through Workspace or similar bundles. And buy the credits at a discount through marketplaces where the math supports it.

The Berlin consultancy that emailed me did the four things above over a two-week period. Her June bill came in at €640, down from €1,900 in April, with no measurable change in what the team can do. That is not a 100 percent solution; it is a 65 percent reduction with an afternoon of plumbing work and a procurement decision. For most small businesses reading this, something similar is available, and the cost of not doing it is real money out the door every month.