At a simple level, many businesses haven’t really chosen. They simply allow (or not) employees to leverage Microsoft Copilot with Office 365 or Google Gemini with Workspace. These two models are synonymous with office productivity suites. However, as many white-collar workers have experienced, it isn’t always easy to replicate the complexity of work tasks in such models. Other businesses have developed an end-to-end data and AI strategy to move tasks onto various models. This could include Llama (Meta), Grok (SpaceX), GPT (OpenAI), Nova (Amazon) and Claude (Anthropic) to name a few. However, there is a multi-dimensional question to answer before pressing ahead. So, let’s discuss how to choose an AI model for your business.
The token economy behind AI models
Broadly speaking, tokenisation is the universal currency for AI billing. Purchasing tokens, which is roughly equivalent to 4 characters or ¾ of a word, gets you outputs from your inputs. Inputs include anything that you put into a model, from text prompts to documents. Outputs include everything that the models generate and serve to you. Whilst there is some efficiency for ‘cached’ tokens, particularly for repeating the same prompt, output is generally more expensive than input.
However, whilst £1 GBP buys you a certain amount of something in the UK, the same wouldn’t buy you very much in Switzerland. The same is true for AI models. The capabilities and tiers of models differ substantially – and not just the cost. Some estimates suggest that the cost can vary by up to 50x from the cheapest to the most expensive model. Unfortunately, it isn’t easy to compare and price is only one part of the problem. There are now examples of organisations who are spending more on AI tokens than on the people who carried out the tasks. Uber famously blew their annual AI budget in 4 months, forcing them to halt spending and cap spend per software engineer.
Other factors to consider: how to choose an AI Model
There are many factors that influence the choice of AI model in an organisation, such as (in no particular order and not exhaustive):
- Sovereignty
- Where AI models originate matters more in some countries than others. For example, the USA is investigating a few companies for potentially ‘leaking’ trade secrets by using Chinese-developed AI models such as DeepSeek.
- Pace of improvement
- OpenAI’s GPT is one of the most frequently updated models, which may or may not cause challenges for software engineers and the users who depend on them, especially if the audit trail for how a decision was made changes.
- Pre-training
- Models are typically pre-trained on large datasets and their parameters are fine-tuned. The more pre-training occurs, the less complex the prompting and the less localised compute is required. It may also influence the quality of outputs and probabilistic accuracy.
- Reasoning depth
- A model that has a longer chain of logic steps between input and output is more likely to produce an answer with higher accuracy. You may also wish to control depth to limit token spend, depending on decision complexity.
- Alignment/suitability
- A model that is biased towards mathematical and repeatable results is most likely to suit rigid and highly-structured workflows. The model also needs to align to your values, rather than the creators in some circumstances.
- Workload type
- Summarisation-heavy workloads with long inputs tend to favour low-end models. Generation-heavy with long output suits lower token prices. Finally, reasoning-heavy workloads typically suit ‘frontier’ models that reduce ‘retries’.
- Weighting
- Some models are based on an ‘open weight’ approach, such as Llama and DeepSeek. In such models, decision weightings are public to download, run and tweak. This allows some visibility, without being open-source, boosting trust for low-risk, high-volume tasks.
- Prompt caching
- Simply defined as re-using text prompt prefixes to reduce the cost of input tokens and reducing latency of follow-up prompts.
- Hallucination rate
- This does not refer to errors explicitly. It refers to untrue, fabricated or misleading information that the AI presents as 100% true fact.
- Ethics and bias
- How business information is used, stored and processed by models (including where) matters. Similarly, how the model was trained and the parameters (even weightings) used can influence the outputs – often reflecting the implicit biases of the developers.
AI models in your business
The question that many CEOs are asking is, “How can I deploy AI in my business to replace existing, routine tasks?”. However, just as many are asking how to structure work for transition, how to test various models and how to limit cost exposure. For those testing the water with local instances or with free versions, the real issue is not where to start, but how to scale it. There is also the ethical question of what to do with the millions of white-collar office workers, especially those in shared services globally, whose routine tasks could be automated by AI.
The industrial revolution shifted the balance from hard labour to machine-assisted labour. The AI revolution is shifting tasks from people to computer models. How much of this work will remain in auditing, checking and monitoring the AI is questionable. However, we also know that the gnarly, complex array of non-uniform tasks across the corporate world is creating a headache. Some 90% of AI projects have failed to make a return on the investment. Companies from Ford to Klarna have re-hired some of the people that they shed. This is because the AI could only handle repetitive, basic prompts rather than judgement and escalating tensions. The worsened customer experience and quality forced rehiring.
Support on how to choose an AI model
Busy CEOs and IT departments should think of AI like trying to catch a bullet. Catch it at the right moment and you can rapidly accelerate. Catch it at wrong moment and it can lead to a world of pain. Since business leaders already juggle everything from investor demands to compliance and customer issues to fundraising, do you have the time to fully explore AI?
If you would like to get support with how to choose an AI model for your business, why not reach out to our team.
Alternatively, read a little more about our approach and about what we do.
Finally, why not also check our thoughts on staff reviews and AI and driving down costs pre-AI.