Money is annoying; it takes up too much space and makes your wallet hard to close. Want a secret 10 second hack that 9 out of 10 dentists agree on? Treat your AI solution like it’s your CRM, give it your credit card details and away you go. “What do you mean, Will? ChatGPT is a monthly subscription. The cost won’t change and keep be budgeted.” Yes, but also really no.
Think of AI as an adopted puppy. Aww isn’t it cute? But what you didn’t realise is that it needed specialised food (cleansed data). It needs obedience training so it’s not barking at hallucinated cats (fine-tuning). Crucially, it needs treats; constant treats just to get it to sit (tokens). This fundamental misunderstanding is why 80% of enterprises miss their AI infrastructure forecasts by more than 25%, and 84% report that AI is significantly eating through their gross margin (Mavvrik, 2025). With so many getting it wrong it’s clear that the mindset of the past doesn’t cut it anymore.
What’s a Few Tokens Between Mates?
Tokens are how LLMs like ChatGPT breakdown words. Some short common words may just be a single token others could be broken up into multiple tokens. The words you enter are the cheap kind, Input Tokens. The ones that the model generates for you are the much pricier Output Tokens.
You’d be right in thinking that a given prompt has the same Input cost. But ask the model to ‘think harder’ and deliver a Michelin Star response rather than Maccies Saver Menu? The price rockets. Those internal workings burn through hundreds of tokens reasoning out the response, all billed at Output Token rates.
Say we use 200 input tokens. A simple task might return 50 tokens with zero reasoning. A complex task returns 150 tokens but burns 2,000 more thinking about it. That’s 2150 overall. But how much more expensive are Output Tokens in reality?
The Model Selection Mania
Choosing the right model is a harder task than Tuchel selecting England’s starting wingers at the 2026 World Cup. At the time of writing, OpenAI’s current line-up has GPT-5 mini for small tasks and GPT-5 for the ones requiring deep reasoning. The costs are (OpenAI, 2025):
- GPT-5 mini – Input = $0.25/M Tokens, Output = $2.00/M Tokens
- GPT-5 – Input = $1.25/M Tokens, Output = $10.00/M Tokens
In our example, GPT-5 reasoned and GPT-5 mini didn’t. GPT-5 mini cost 0.015¢, whilst GPT-5 is 2.175¢. The answer from GPT-5 is way better, but it has just cost your 145 times more.
But that’s just basics. What if you solution needs to code? Process images? Draw information from multiple sources versus just 1? Yes the GPT-5 sledgehammer will squash the fly, but depending on the use case, a cheap flyswat specialised for the task may work better and save a chunk of change.
Okay fine, so there’s a bit of work choosing the right model. That’s it, though… right? Oh boy, we’ve just skimmed the top off Everest. Because even if you pick the perfect model, you still need to feed it something. And this is the bit where most projects collapse.
Death, Taxes and Quality Data
If AI is your new office employee, quality data is the coffee. To get them functioning right it needs quality data. There’s an upfront cost in getting your data into a useable state. Think about your Power BI dashboard. Feed in duplicate customer entries or missing revenue values and it won’t fall over. Instead, it’ll present a beautiful colour-coded lie (insert something to do with glitter and rolling). An experienced analyst will spot the false narrative.
AI is different. It is not just the dashboard; it’s also the intern who doesn’t know better, analysing results and giving recommendations. Feed AI messy customer policies and it’ll tell a customer they have 90 not 30 days to return an item. Give it flawed revenue stats and it’ll recommend axing a well performing product. These issues are hard to notice once deployed, but the ramifications are you’re paying for an expensive service causing even more expensive problems.
This is why we always start with a data audit before any AI conversation. Not just because we’re boring, but because deploying AI on messy data is like paying a child to review Shakespeare before they can read, they’ll make something and it’ll be utter nonsense.
Ensuring AI has integrated, cleansed data from various sources is the most critical step. But that time is frequently overlooked. According to a PwC study, 42% of AI projects required unforeseen spending on data quality initiatives, adding an average of 30% to initial budgets (Piccolo, 2025).
SMEs Have The ROI High Ground
It may seem like I’ve just framed an AI project to be the next HS2, but it doesn’t have to be. Hughes (a subsidiary of EchoStar) developed 12 production apps, from automated sales call auditing to customer retention analysis. This saved roughly 35,000 man hours and boosted productivity by at least 25%. Or take Blix, a US retail company. They increased revenue by 150% in a year with AI-powered chatbots unlocking real-time support and product recommendations.
You’re in a prime position being an SME. It used to be that big data won, but now your size is your advantage. Large enterprises are likely drowning in their data swamps of duplicated, siloed data. You may only have a puddle, but you can see the bottom. You don’t need a data lake, just a golden dataset to tune a chatbot to talk like your business. You can make decisions in days, not months of committee meetings. In the dawn of AI, it’s data for show, quality for dough.
Closing Pandora’s Box
Start small and targeted. Don’t get excited by a specific AI tool and shoehorn it into a workflow. Instead, pick that painfully repetitive task that’s sucking the soul out your employee like an overenthusiastic primary schooler with a Frube. Things like filling out process forms based on email information or finding quick answers from customer data like ‘all customers in Hampshire’ without badgering an Analyst. Why? These free up talented staff for complex work, and that time saving stacks up against token costs. Plus, removing drudgery boosts mood and adoption.
Use a pilot scheme. Don’t jump straight to live. Set a strict budget, run it for a quarter against the current process and measure key statistics. Think about the specific data you will need, a customer record from 1987 isn’t likely useful and chews into the time cost to build. Enterprises missing their forecasts by huge chunks likely skipped this step; they try to scale before they learned.
AI isn’t just for Christmas; it’s part of your working life now. Budget for improvements and fixes, not just initial implementation. Get this right and you’ll be boring people at parties about your AI wins. Get it wrong and you’ll be a story for my blog.
The Bottom Line
AI doesn’t need to be a budget black hole. But it will pick your pocket if you let it. AI might seem like a chaos causer, but the truth is it’s opening your eyes to existing chaos. The bad results come from rushed, poorly planned projects. Someone picks the most expensive model because “more money equals better,” powers a process that sounded good in a meeting but solves no real problem, feeds it disconnected data, and acts surprised when the results are useless. All just to tick the AI box for the board.
PS: We have built a free AI Data Readiness Diagnostic Assessment so you can see in 5-minutes where your estate stands.
References
Mavvrik. (2025, September 10). 2025 State of AI Cost Management Research Finds 85% of Companies Miss AI Forecasts by >10%. Retrieved from PR Newswire: https://www.prnewswire.com/news-releases/2025-state-of-ai-cost-management-research-finds-85-of-companies-miss-ai-forecasts-by-10-302551947.html
Microsoft. (2025, June 6). EchoStar and Hughes save thousands of work hours, cut costs with Azure AI. Retrieved from Microsoft: https://www.microsoft.com/en/customers/story/24300-hughes-azure-ai-foundry
OpenAI. (2025, November 28). Pricing. Retrieved from OpenAI Platform: https://platform.openai.com/docs/pricing?latest-pricing=standard#text-tokens
Piccolo, V. (2025). Cost of implementing ai in 2025. Retrieved from callin.io: https://callin.io/cost-of-implementing-ai/
