AI Costs for Small Business: Why AI Got So Cheap

The most important development in AI for your business is not a model release: it is that AI costs for small business have collapsed, and that price drop matters more than GPT-4, Gemini or any agent demo. If you run a small or mid-sized UK firm wondering whether AI is finally affordable, this post is for you. By the end you will understand why the collapse happened, what it changes for an SME, three rules for spending well, and an honest note on where this thinking could be wrong. We are a Manchester technology business that advises and builds AI systems for UK small businesses, and we design caching-aware pipelines for clients because the economics, not the hype, decide what is worth doing.

Why did everyone focus on the wrong thing?

The headlines went to capability because capability is photogenic and price is not. A model that passes an exam or drives a browser makes a great demo; a graph of falling cost per token does not.

The story that did not make the news

So the storyline that matters most to a small business barely got covered. Work that cost pounds per query on 2023’s frontier models now costs fractions of a penny. That collapse, not any single clever model, is why AI automation stopped being an enterprise toy and became something a Manchester trades firm or clinic can run profitably.

Why the mechanics matter to you

Here is the thing: the mechanics of that drop tell you where prices go next, which is exactly what you need to plan. If you know why something got cheap, you can judge whether it will stay cheap.

What actually made AI cheap?

AI got cheap because of a handful of unglamorous engineering breakthroughs, not one headline model. Each one chipped away at a different cost, and together they reset the economics.

Models stopped using all of themselves

The most important architectural idea of the period was the sparse mixture-of-experts, where each query activates only the relevant sub-networks. Mistral’s Mixtral in December 2023 put it into the open-model mainstream, and the effect was big-model quality at a fraction of the running cost, because most of the model sleeps through most questions.

Customisation stopped being a lab project

Training and tuning got radically cheaper too. A string of research results, QLoRA squeezing large-model fine-tuning onto a single GPU, DPO simplifying preference training, made customisation a line item rather than a data-centre project.

The labs learned to stop recomputing

Prompt caching, remembering the unchanging part of a request so it is not reprocessed on every call, sounds like an accounting trick, but at scale it routinely cuts a high-volume workload’s cost by an order of magnitude. Every serious automation pipeline leans on it, ours included.

Small models grew up

Models like GPT-4o mini, Gemini Flash and Claude Haiku handle most real business tasks (classify this email, extract these fields, draft this reply) without needing a frontier intellect. Vendors such as Anthropic and OpenAI now publish whole tiers built for exactly this, and matching the model to the task is the single easiest cost optimisation available.

Competition did the rest

DeepSeek-R1’s January 2025 release, frontier reasoning, fully open, built for a fraction of the assumed cost, forced everyone’s pricing assumptions down. The full open-versus-closed picture is worth its own read in our guide to open-weight versus closed AI models.

What does cheap intelligence change for a small business?

Cheap intelligence moves the constraint from the AI itself to the system you build around it. The cost of being clever has stopped being the bottleneck.

Run the arithmetic

An AI assistant that answers website enquiries, qualifies leads and books consultations handles a conversation for pence at frontier quality, and a fraction of that on a small model. A missed-call text-back automation costs effectively nothing per event. A document pipeline that summarises and files paperwork runs for less than the coffee of the person who used to do it.

The cost moved, it did not vanish

What remains is the cost of building the system well: wiring it into your tools, setting boundaries, testing it, keeping it reliable. That is a one-off craft cost, not a scaling cost, which is exactly the shape of investment small businesses do well with. Pay once to build, then run at near-zero marginal cost. If you want concrete examples, our overview of what AI automation can actually do walks through them, and our workflow automation service is where it becomes a project.

What about the AI you already pay for?

Before buying anything, check what you own. A lot of small firms are already paying for AI features sitting unused inside software they bought for other reasons.

The obvious place to look

If you are on Microsoft 365, the Copilot documentation is worth an hour of someone’s time. The same is true of the AI features quietly added to most accounting packages, CRMs and helpdesk tools over the last two years.

We check this first on client projects because it is embarrassing to build something a customer is already paying for. If your Microsoft setup has not been reviewed in a while, that is a job in its own right, and our Microsoft 365 and IT support service covers it.

Cheap does not mean free of duty

Falling prices do not change your obligations. Whatever you run, you are still accountable for the data going into it. The ICO’s guidance on artificial intelligence sets out what that means in practice, and the NCSC’s briefing on AI and cyber security covers the risks that arrived with the savings. The rules side is covered in more depth in our plain-English AI regulation timeline.

How should an SME spend on AI now?

Three practical rules follow from the economics, and they apply whether you are spending fifty pounds a month or five thousand.

1. Right-size the model. Use frontier models where judgement matters, such as customer conversations, and small models for routine classification and extraction. The price difference is often five to twenty-five times for near-identical results on simple tasks.
2. Design for caching. Pipelines structured to reuse stable context cost a fraction of naively built ones. This is invisible in a demo and decisive on a monthly invoice, which is why we build it in from the start.
3. Revisit the maths yearly. Anything that was “too expensive to automate” in 2024 probably is not now. Prices keep falling, so your workflow list deserves an annual review.

The order to spend in

Spend on the thing that is losing you money first, not the thing that demos best. For most service businesses that is the enquiry nobody answered or the invoice nobody chased, both of which are cheap to fix and immediately measurable.

What does this change about hiring?

Less than the headlines suggest, and not in the direction most people assume. For a small business, cheap intelligence rarely removes a role. It removes the worst parts of several.

The pattern we actually see

The jobs that get absorbed are the ones nobody wanted: re-typing enquiries, chasing quotes, formatting reports, copying numbers between systems. What is left is the work people were hired for in the first place, which is usually the work they are better at than any model.

Where it does change the sums

It changes the sums at the margin. A business that would have taken on a part-time administrator to cope with growth can often postpone that hire, not by working the team harder, but because the admin genuinely shrank. That is a real saving, and it is worth being honest that it is also a real decision about how you want to grow. Our AI consulting service exists mostly to help teams make that transition without anyone feeling replaced.

Where could we be wrong?

We could be wrong about how long the price falls last, and it is worth saying so plainly. The collapse so far has been driven by efficiency research and fierce competition, and neither is guaranteed forever.

The case against our own argument

If competition consolidates, or if the next capability jump turns out to be genuinely expensive to run, the cost curve could flatten or even tick up for the best models. The cheapest small models will likely stay cheap, but “frontier capability for pennies” is an assumption, not a law.

Why the advice survives anyway

That uncertainty does not change what to do. Right-sizing, caching and an annual review are sensible whether prices fall fast, slowly, or stall. The strategic takeaway is blunt either way: your competitors’ access to cheap intelligence is identical to yours, and the advantage goes to whoever turns it into working systems first. For the longer view of how we got here, our Story of AI timeline covers the whole arc.

If you want the cost arithmetic done for your specific workflows rather than in the abstract, that is what our AI Opportunity Assessment produces, and you can always just ask us first.