What should your AI cost?

AI is unarguably the hottest tech trend right now, but ask around and most professionals aren’t clear on what constitutes a ‘fair price’. It’s not because you haven’t asked or didn’t understand the memo; it’s because we’re in the early stages of a new market. Everyone is still finding their way, and transparency and consistency are tricky concepts when nothing stands still for long.

That comes with risks: some decision-makers are worried it’s all wildly expensive (or reassuringly expensive, depending on your worldview). Others are piling in on free trials without calculating the costs that stack up later — in people’s time, in tech debt (the extra cost or effort you’ll have to deal with later), in missed opportunities, or in tools you can’t get rid of.

So… what should AI cost?
The short answer: it depends.
The longer (and more useful) answer is below.

This article is for anyone trying to make good decisions about AI in their organisation — especially if you're not the IT person but want to be a good commissioner, and especially if you don't want to get stung. We’ll walk you through what we’ve learned so far about what things really cost, what's fixed, what's flexible, and how to ask the right questions. Including when you talk to us.

What people are paying right now: mainstream AI

Let’s start with the stuff people are buying personally or in small teams. This is the mainstream “get started quickly” chunk of the market, where everything seems to cost $20 a month, or thereabouts.

These are decent starting points, especially for early exploration and individual professional roles like policy, comms or product teams on a journey to familiarity with new AI and the art of prompting.

Public service announcement: If you're feeding these tools your organisation’s data – or indeed anything sensitive –make sure you know what happens to the information you give it. Most tools will use it to ‘train their model’ by default, unless you’ve opted into a paid plan with proper data controls and flipped the switch. Training sounds like such a lovely idea — and we all like to be helpful — but in generative AI, it means your data becomes part of the training model. The big sea of data feeding the AI new content. And it’s there forever. For everyone. Your IP might become someone else’s chatbot answer.

If everyone were being honest, the checkbox would say:
☑ Yes, I agree to sell my innermost thoughts and all team chat logs to the algorithm overlords.

This matters even more in public sector settings, where the rules often prohibit not just data sharing but the processing of data in other countries — like those where the government has a legal right to poke around in any cloud server it likes.

The enterprise versions: what big organisations are buying

Once you move beyond individual use, things get pricier — and less clear.

Enterprise plans usually offer better security —but not always. Check what “enterprise” actually means, or give the Ts & Cs to your most diligent colleague with a cup of tea, a biscuit and a highlighter pen.

Some public sector bodies and education providers can get sweetheart deals, either because they’re early adopters or because the vendor wants the relationship. Don’t be shy about asking.

What about industry-specific AI tools?

There’s a growing list of AI tools on offer at the trade fairs, targeted to specific sectors — legal AI, bidwriting bots, planning assistants, AI for fundraising, and many more. Prices vary wildly — and so does value.

Many really do offer extra functionality, like integration with sector-specific data or formats, automation of fiddly workflows, or specialised prompts and fine-tuned guidance.

Sometimes you’re paying for speed, polish and convenience — but sometimes you’re paying a hefty markup to avoid learning how to write a basic prompt, an art you can definitely master. And it’s going to be the new life skill.

A lot of specialist tools (including ours) are, at their heart, a top-tier LLM (like Claude or GPT-4) behind a tidy interface and well-crafted training prompts. That might be all you need, or you might need a bit more, like a partnership with a team who will co-design your AI and give you proper control over where your AI sources data from and where it sends your data. (That’s usually where we come in.)

What are people charging for specialist AI?

We are, inevitably, fascinated by this question. Here are some real examples (names are available on request, but I’ll spare their blushes here):

Ask yourself whether these tools are really doing something that off-the-shelf Claude or GPT-4 cannot. Will they save time and make outcomes better?

Then ask why some (most) of these tools range between 4 and 10 times the price of ours, per user, especially if they’re not providing onshore data processing or trained on a curated library of reliable content.

What you’re actually paying for

It’s hard to price something when the “thing” doesn’t sit in a box. Capital expenditure makes more sense when you’re buying land. But your new AI solution has real, actual costs that have to be addressed. You think clouds are free; they’re not.

1. Cloud and compute

Even if a tool feels free to use, something is paying for those servers. Usually, that’s you — through charges for API calls(every time your AI app sends something off to be processed), data storage(think documents, chat logs, logs of chat logs…), and cloud infrastructure costs (typically Azure, AWS, or Google Cloud).

In our experience, it’s hard to make anything work in this space – paying for cloud and compute alone – for less than about $250 a month. We’ve also found that AWS can be cheaper and more flexible than Azure. Either way, you’ll need someone who knows what they’re doing. Pro tip: look after your cloud whisperer.

2. Licenses, tokens and funny money

Some solutions charge by the number of tokens processed (for the uninitiated: chunks of text that are like calorie counts, but for sentences). Others charge by user, or by how many API calls you make. Some will cap your usage each month and start charging extra if you go over.

This is where the dreaded “death by overage fee” comes in. It’s easy to get stung if you don’t know how usage might scale. Either do your sums up front or negotiate a deal where you can flex your usage as adoption grows. And by the way, anyone who can confidently work this out from their Azure dashboard deserves a medal.

3. Our time and effort, your time and effort

If you use a supplier (waves cheerily 👋), you’re also covering the real costs of development, support, and — yes — sales and marketing. Pricing isn’t always straightforward. It’s a bit like drug development: the first prototype was probably too expensive to sell anyone, but once it’s tested and working, scaling it becomes easier.

If you're going the DIY route, don’t forget to factor in the time your team is spending. Has your tech team built a prototype in Copilot Studio? Did they stitch together a few APIs into something that sort-of works? That’s all real cost. It might pay off by building internal capability, but someone has to keep it running.

For context: our dev team rebuilds our RAG assistants every couple of weeks. It’s usually necessitated by needing minor tweaks, upgrades, or adjustments to cloud configuration changes. You can do that yourself (well, maybe not you, but someone who understands the back end of your systems), but sometimes it’s more efficient to work with someone who does it all the time and knows how to get it right first time. It’s like doing your own plastering: technically possible, but maybe not the best use of your weekend(s).

4. The cost of change

AI isn’t usually plug-and-play — not if you want meaningful change. Proper adoption often means redesigning workflows, reviewing policies and compliance, training or hiring people with new skills, integrating legacy systems, and translating all of that into three slides and one metaphor for the board.

All of that needs time and resourcing. When the benefits are there, it’s worth it. And if your solution includes a slick interface, quick wins, and minimal training? Even better. But let’s be honest: that’s not always where we start.

What to watch out for

Cheap per user doesn’t always mean cheap overall. £5 per person sounds like a steal, until you realise that for 200 users, so it’s£12,000 a year. And that’s before VAT, setup and training.

Free trials can easily become time sinks. Without a decent use case, you’re just poking shiny things with a stick. That’s fine if you’re doing it intentionally and you know when the trial ends — but it might be time better spent somewhere else.

Vendor lock-in is real too. Some solution providers make it unnecessarily hard to leave or to extract your data. Always ask: Can we get out of this cleanly later? This matters even more when your management information system starts offering "AI upgrades".

A handy checklist: Questions to ask before you buy

Last week’s blog was all about the risks of sitting on decisions forever. This one’s telling you to pause and think. I see you raising an eyebrow, but both are true. I am prepared to die on this hill of just enough planning and no more.

The biggest costs are usually the ones no one budgeted for —the extra hours, the opportunity cost and half-built solutions, the nice idea nobody quite owned. But there’s good news: you don’t need a huge budget. Ask any of our customers. Then ask yourself:

1. What exactly are we trying to achieve with this tool?
2. Who will use it, and how often?
3. Is this data sensitive (and available to use)? Do we know where our inputs and outputs go?
4. Are we locked into a vendor or ecosystem – have we benchmarked the cost and quality of it?
5. What's the real cost over a year — including licences, support, compute, and time?
6. Is there a better way to test this — e.g. with a small pilot or sandbox?
7. Can we negotiate a better deal — especially if we're public sector or doing something really cool that people want to be associated with?

The big lesson of change – whether it’s AI-powered or just good, old-fashioned, workplace transformation, is that only really need commitment, communication, clear goals, a chance to ask questions and the confidence to walk away from something that doesn’t make sense. You don’t need to write a paper. ChatGPT can do that.

And maybe you need a friendly expert to help. That’s us. We’re always here when you need us.