Most small businesses treat AI like a subscription. Pay the monthly fee, use it when you need it, move on. That model is changing.
The industry is moving to pay-as-you-go, and every AI interaction has a price. The unit of that price is the token – and for businesses running on disconnected systems, spreadsheets, and scanned PDFs, that cost is higher than it needs to be.
What is an AI token, and why does it matter?
Tokens are the unit of consumption for every AI interaction: every word you send, every word you get back, every document, image, or agent loop running in the background costs tokens.
For a long time, that cost was hidden inside a flat monthly subscription. That worked when people were mostly chatting with AI.
Chatting with an AI uses a few hundred tokens per paragraph. But agentic AI – where models execute multi-step tasks, browse the web, and act on live business data – can consume thousands more in a single session.
How SMEs accidentally burn through tokens
Most token waste isn't intentional. It happens because businesses haven't thought carefully about what they're sending to the model. Four common patterns:
Noise attached to documents
PDFs and scanned files require OCR processing before the model can read them. All that noise (headers, footers, formatting) consumes tokens that generate nothing useful. Plain text or CSV gives the model exactly what it needs, nothing more.
Bloated prompts
Dropping a whole manual into a prompt when a concise markdown instruction would do. More input is not better input. A well-framed, specific prompt will consistently outperform a large, vague one – and cost significantly less.
Agent loops with no output
When agents talk to other agents, cycle through the same data repeatedly, and produce no useful result, the token meter keeps running. This is one of the more invisible ways costs escalate. A useful rule: if you can't describe the process clearly in plain language, an agent isn't ready to run it.
Not using caching
Most major AI platforms handle caching automatically. If the model has seen the same prompt before, it retrieves the answer rather than reprocessing from scratch. Cached tokens can cost up to 90% less (1). The catch: it only works if you're consistent. Build a simple prompt library so your team uses the same wording every time.
Clean data costs less to run through AI
This is where ERP foundations matter.
A business exporting a formatted Excel report sends rows, columns, cell formatting, and metadata – most of which the model discards. A business querying a structured inventory movements table in an ERP like SAP Business One sends a clean, precise dataset to the model. Same insight, fewer tokens.
The better approach: connect AI securely to structured data
By securely integrating AI with your source data, instead of generating a report and handing it over, you enable the model to work from structured information – inventory movements, aged debtors, sales history – and find the insights itself.
Using an ERP like SAP Business One runs on SAP HANA or Microsoft SQL Server – industrial databases built to process large volumes of data almost instantly – makes this process easy. A secure AI integration can query database fact tables in real time, retrieving only the relevant data for analysis while keeping appropriate controls around sensitive business information.
For example, you could analyse 12 months of customer ageing data and ask the model to identify credit risks or customers worth watching.
But the right architecture matters. Before connecting AI to business systems, make sure your approach protects confidential information and aligns with your security and governance requirements. An experienced ERP partner can help you design an AI integration that's both efficient and secure.
The pricing shift you can't ignore
The cost of running AI is dropping, but usage is scaling to fill the gap. Jevons Paradox tells us that when technology gets cheaper, total consumption rises to match – because suddenly it makes economic sense to apply it everywhere.
It happened with cloud computing, and it's happening with AI. The price per token will keep falling. The number of tokens your business consumes will keep rising.
Anthropic was the first major provider to make the move, shifting away from flat-rate enterprise pricing toward per-token billing in April 2026 (2). The rest of the market is following:
-
DeepSeek dropped its cached token pricing by 98% in 2026 (3)
-
Large enterprises are now routing routine agentic work through lower-cost models and reserving the premium LLMs for tasks that genuinely need them
For SMEs, the message is straightforward: not every task needs the most expensive model, and not every process should be automated just because it can be. The value case has to come first.
Where to start: lowest token, highest value
For a business evaluating ERP and thinking about AI, start with a process that already has clean, structured data behind it and a measurable outcome at the end. Good candidates:
-
Margin analysis
-
Cash flow review
-
Aged debtor risk flagging
These are repeatable, the data is well-defined, and the output is something you can act on immediately.
Use generative AI first to prove a use case is worth pursuing, then layer in agentic AI to remove the friction once that's established. Start small. Prove it works. Build from there.
How Key Business Solutions can help
If your business is still running on spreadsheets and PDFs, you're not just behind on ERP, you're starting every AI interaction at a disadvantage.
SAP Business One gives SMEs the structured, connected data that makes AI work properly and cost-effectively. Key Business Solutions helps businesses get that foundation in place, so that when AI enters the picture, it delivers.
Ready to explore? Start a conversation with our team today.
(1) https://platform.claude.com/docs/en/about-claude/pricing
(2) https://www.cnbc.com/2026/04/17/ai-tokens-anthropic-openai-nvidia.html
(3) http://developer.puter.com/tutorials/deepseek-api-pricing/
