Not every automation problem needs a large language model. In our AI consultancy work, choosing the simplest technique that works saves money and produces more reliable results. Here is how we decide.
Use rules when the logic is clear
If a process can be written as clear conditions, such as “orders above a limit need manager approval” or “send a reminder seven days before due date”, rules or workflow automation are the right tool. They are predictable, cheap and easy to audit. Tools include ERP workflows, Salesforce Flow and n8n.
Use machine learning when patterns hide in data
When you have historical data and want predictions, such as demand forecasting, churn risk or lead scoring, classic machine learning models work well. They need clean history and periodic retraining, but they are efficient once running.
Use LLMs when language is the work
Large language models are strongest when the task involves reading, writing or understanding text: summarising documents, drafting emails, answering questions about records, or extracting data from invoices. Connected to your systems as an AI agent, they can also complete multi-step tasks with approval.
Combining techniques
Most real solutions mix all three. For example, rules route a supplier invoice, an LLM reads it and fills the fields, and a machine learning model flags unusual amounts for review.
Questions we ask
- Can the decision be written as rules today?
- Is there enough reliable history to learn from?
- Is the input mostly text, documents or conversations?
- What is the cost of a wrong answer, and who checks it?
Frequently asked questions
Are LLMs too expensive for routine tasks?
For simple decisions, yes; that is why we use rules there and save LLMs for language-heavy work.
Related reading
For the product view of this topic, read Rules, Machine Learning or LLMs: AI Options in ERP on the Reckon ERP blog.
Planning a project? Book a free consultation with our team, or see our AI consultancy and agent development service.
