Why Standardising Your Organisation on Claude or ChatGPT Is a High-Risk Bet

Let me be clear about something upfront: Claude, ChatGPT, and their peers are genuinely remarkable. The pace of improvement over the past few years has been extraordinary, and the productivity gains for people who use these tools well are real. I’m not arguing they’re overhyped. I use them daily.

What I am arguing is that standardising on one of them (buying enterprise licences, rolling it out across your organisation, training your staff on it, building workflows around it) is a strategic decision that deserves far more scrutiny than most boards are giving it. There are six distinct reasons why, and taken together they paint a picture of risk that I think most IT directors aren’t fully accounting for.

You’re Locking In on the Model, Not Just the Tool

Here’s a question that rarely gets asked in procurement: are you buying access to a model, or access to a platform? With most enterprise AI today, the answer is both, and that’s the problem.

The model is the underlying intelligence. The tool is the interface, the workflow, the integrations. These are very different things, and they improve at very different rates. Right now, Claude is arguably the strongest model for complex reasoning. Next year, that could easily change. GPT-5, Gemini Ultra, or something from a completely different direction could be demonstrably better. If you’ve standardised on Claude’s tooling, you can’t easily switch. Your organisation has been trained on its interface, your workflows are built around its quirks, and your IT team has built integrations into its APIs.

The principle here is straightforward: tooling and model selection should be separated. The organisations that will navigate this landscape well are the ones that treat model choice as a configuration decision, not a structural commitment.

The model is the engine, not the car. Buying a car because you like this year’s engine is a decision you’ll revisit sooner than you think.

The Economics Are Built on Foundations That Don’t Add Up

The £20-per-month subscriptions that made AI accessible to individuals are, by any reasonable analysis, significantly subsidised by venture capital. It’s not hard to see why: spend a few hours using Claude or ChatGPT on a substantive coding or analysis project and you’ll consume tokens that would cost £100 or more at unsubsidised API pricing. The maths only works if most subscribers use far less than you do, or if the VC runway holds.

The signs that this is already under pressure are everywhere. Large companies are being pushed off flat-rate subscription pricing onto API billing, often ten times more expensive. The highest-capability models are being ring-fenced outside standard subscriptions. Response times have degraded noticeably. These aren’t coincidences; they’re symptoms of organisations under pressure to improve their unit economics.

Meanwhile, Anthropic and OpenAI have both committed to vast data centre infrastructure on the assumption that AI demand will continue to expand at current rates. If demand softens (or if cheaper alternatives capture the lower end of the market), those data centre bets become liabilities, not assets. Organisations that have locked into these vendors as a strategic standard will find that pressure passes directly to their invoices.

The Subscription Model Wasn’t Designed for How Organisations Work

The per-employee subscription model (pay £40 a month for every seat) makes intuitive sense for software that every employee uses every day: email, calendars, document editing. It makes considerably less sense for AI.

The honest reality is that in most organisations, the productivity gains from AI are highly uneven. A small number of staff (typically those with analytical or technical roles) extract enormous value. A larger number of staff use it occasionally, for relatively low-value tasks. And a meaningful proportion of employees, particularly in operational or customer-facing roles, may derive little benefit at all from access to a general-purpose AI chat interface.

Paying £40 per head per month for the latter group isn’t just wasteful. It’s the wrong model entirely. What those employees actually need is specific, well-built tools that use AI where it adds value, not an open-ended interface to a general model. The subscription model optimises for the vendor’s revenue, not for the organisation’s needs.

The Platform Is Built to Maximise Token Consumption — Not Your Efficiency

There is a structural conflict of interest in the mainstream AI business model that rarely gets named directly. Revenue is proportional to tokens consumed: every question asked, every document summarised, every analysis run is a billable transaction. That creates a direct incentive to route as much work as possible through inference, and it shows in how these tools are designed: the interface is a chat box; the paradigm is “ask the AI.” Yet for many of the tasks organisations actually need to accomplish, the better solution isn’t an AI response. It’s code.

The difference matters enormously. A prompt that extracts data from a document on every invocation consumes tokens every time; a well-written function does the same job in milliseconds at near-zero cost, and AI can write that function. The smarter architecture uses AI at the point of construction, to build tools that then run in your own environment, and reserves runtime AI only for tasks where live inference genuinely earns its cost: variable inputs, novel outputs, reasoning that code can’t anticipate. But there’s no commercial incentive to help you get there; if anything, there’s a structural incentive not to.

The practical obstacle is distribution. AI-written tools need somewhere to go: a deployment layer that makes them accessible to users, integrates with existing systems, and operates within your security perimeter. That infrastructure is conspicuously absent from mainstream AI platforms. A robust distribution layer for AI-built tools would systematically shift work away from paid inference, and that’s precisely why nobody offering paid inference is rushing to build one.

AI vendors are incentivised to do your work for you, repeatedly, in the cloud, at a cost per token. The smarter architecture uses AI once to build the tool, then runs the tool yourself.

Token Costs Are Ungovernable — and That’s a CFO’s Nightmare

Software has traditionally been a capital cost or a predictable rental. You buy a licence, or you pay a monthly fee, and your finance team can budget for it. AI introduces something genuinely new: a consumable cost, variable in real time, distributed across every user in the organisation, and essentially impossible to attribute meaningfully without significant additional tooling.

Think of tokens as fuel. Every prompt burns some. A simple question burns a little; a complex analytical task burns a lot. In a large organisation with hundreds of employees all making their own decisions about when and how to use AI, the aggregate cost becomes both substantial and opaque. The CFO sees a large and growing bill. They cannot tell which teams are generating it, which use cases it’s funding, or whether the spend is delivering proportionate value.

This is not a solvable problem within the current model of direct individual access to large AI platforms. The only way to bring token costs under genuine organisational control is to centralise AI access through managed platforms, which brings us straight back to the problem of vendor lock-in, and raises the question of who controls that platform and on what terms.

Token costs are fuel costs: distributed, variable, and opaque. No other software category works this way, and most organisations aren’t financially set up to manage it.

Your Data Is Leaving the Building

Every prompt your employees send to Claude or ChatGPT travels across the internet to servers owned and operated by a US technology company, processed under that company’s data policies, subject to US jurisdiction, and potentially used (depending on how your enterprise agreement is structured) to improve their models.

For a general conversation or a public-domain research task, this may be entirely acceptable. For the kinds of prompts that generate real business value (analysis of client data, drafting of sensitive commercial documents, exploration of strategic options, summarisation of confidential meeting notes), the picture is considerably less comfortable.

Most enterprise agreements include data processing addenda that provide some protection, but they are not equivalent to data remaining within your own infrastructure. For organisations in regulated sectors (finance, healthcare, legal, public sector), this is not a theoretical concern. For any organisation that handles data subject to UK GDPR, the question of where prompts are processed and how long they are retained deserves a proper legal answer, not an assumption that the enterprise tier handles it.

Data sovereignty is rapidly becoming a board-level concern in a way that it wasn’t two years ago. An organisation that has standardised on a US-hosted AI vendor is one that has quietly made a data residency decision it may not have explicitly taken.

What the Right Architecture Looks Like

None of this means organisations should avoid AI. Quite the opposite. The productivity opportunity is real, and businesses that don’t engage with it will fall behind. The argument is about how to engage with it.

The answer, I think, is a platform layer that your organisation owns and controls: one that is model-agnostic (so you can swap underlying AI providers without rebuilding your workflows), that uses AI to build tools rather than routing everything through inference, that deploys those tools as governed applications rather than open-ended chat interfaces, that tracks token usage down to the team and use case, and that keeps sensitive data within your own infrastructure.

That’s the architecture we’ve been building at Blueberry. Our platform, VibeMagic, is designed specifically to give small and mid-sized organisations the benefits of AI without the lock-in, the unpredictable costs, or the data sovereignty compromises. It’s model-agnostic by design, uses AI to build and deploy actual tools rather than encourage open-ended chat, and gives administrators full visibility into how AI is being used and what it’s costing.

The AI landscape will look very different in three years. The organisations that will be best placed are the ones that built their AI strategy on foundations they control, not on a subscription to whoever happened to be winning the model race this year.