Classic AI vs New AI: Which Does Your Business Need in 2026?
Classic AI vs new AI: when traditional machine learning beats an LLM, when it does not, and how UK mid-market teams pick between them without wasting a budget.
By Ivan Pylypchuk, CEO of SoftBlues. Has led Claude and automation projects for finance, legal and healthcare teams across the UK and Ireland. Last updated: 29 July 2026.
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating costs and unclear business value (Gartner, June 2025). A good share of that waste comes from one avoidable mistake: pointing a large language model at a job that a much cheaper, much more reliable classical model already solved twenty years ago.
SoftBlues is an AI consultancy at softblues.io, a registered Anthropic Partner Network member and a Google Cloud Partner, working with regulated mid-market companies across the UK and Ireland. We spend a good part of every discovery call separating the two, because the answer changes the budget by an order of magnitude.
Classic AI means traditional machine learning: models trained on your own structured, labelled data to predict or classify something specific. New AI means large language models such as Claude and GPT, trained on enormous general corpora to read, write and reason over messy input. Classic AI is cheaper to run, easier to validate and far more predictable. New AI handles language and variation that no classical model could touch. Most real systems need one of them clearly, and a fair number need both.
Key facts
What is classic AI, and what is it still best at?
Classic AI covers the machine learning that has been running quietly in businesses since long before ChatGPT: regression, decision trees, gradient boosting, clustering, classical computer vision. You give it labelled examples, it learns the pattern, it predicts.
It is still the right answer for anything where you have structured historical data and want a number or a label out. Forecasting next quarter''s demand. Scoring a transaction for fraud risk. Predicting which customers are about to leave. Spotting a defect on a production line.
Three reasons it keeps winning those jobs. It is cheap enough per prediction to run millions of times. It is deterministic, so the same input gives the same output and an auditor can follow it. And it can be validated properly, with a confusion matrix and a held-out test set, rather than argued about.
What is new AI, and where does it earn its cost?
Large language models are trained on general text rather than your data, and they arrive already able to read and write. That is the difference that matters. You do not need a labelled dataset to get started, and the input does not have to be tidy.
They earn their cost anywhere the input is language a human would otherwise have to read: contracts, emails, invoices, case notes, support tickets, policy documents. They handle variation gracefully. A new supplier''s invoice layout breaks a rules engine and does not trouble a model.
They are also the only sensible option when the task needs interpretation rather than calculation. Deciding whether a complaint is about a product fault or a delivery failure is judgement. No amount of structured data makes that a regression problem.
Classic AI vs new AI, compared
| Classic AI (traditional ML) | New AI (LLMs) | |
|---|---|---|
| Input it wants | Structured, labelled, consistent | Messy language, documents, mixed formats |
| What it outputs | A number or a category | Text, extracted fields, a judgement |
| Training data | Your own historical data, labelled | Pre-trained. Your data only for context |
| Cost per decision | Fractions of a penny | Meaningfully higher, every call |
| Same input, same answer? | Yes | Not without constraints |
| Time to first version | Weeks to months, data work dominates | Days, if the data is accessible |
| Explaining a decision | Feature importance, standard practice | Harder. Needs logging and human review |
| Breaks when | The input format or distribution shifts | The task needs exact arithmetic or hard rules |
How do you choose between them?
Map the task, then answer two questions.
Is the input structured or messy? Rows in a database point to classic AI. Free text, PDFs, emails and scanned documents point to an LLM.
Does the task need judgement or repetition? If a competent person could write down the rule, you do not need a model that reasons. If two competent people would sometimes disagree, you do.
That gives you four outcomes. Structured plus repetition is classical machine learning, or often just rules. Structured plus judgement is classical ML with a human approving the edge cases. Messy plus repetition is an LLM doing extraction, with deterministic validation after it. Messy plus judgement is an LLM with a human signing off anything consequential.
The last of those is where most of the failed projects sit, because teams skip the human sign-off and then cannot defend the output. We set out where the human belongs in human-in-the-loop AI workflows.
The pattern that actually ships: use both
Almost every system we put into production uses both, in sequence.
Take invoice processing. The LLM reads the document, whatever layout it arrives in, and extracts the fields. Then deterministic code checks the arithmetic, validates the supplier against your ledger, and applies your approval thresholds. The model does the reading, which it is good at. The code does the maths and the rules, which it is good at and the model is not.
Ask an LLM to total a column and you have chosen the expensive, unreliable way to do addition. That single design mistake accounts for a surprising number of pilots that never reached production. There is a fuller version of the argument in AI agents vs RPA for mid-market operations.
You can see the combined pattern working in our compliance file review automation for a financial advice firm.
Where neither is the right answer
Sometimes the honest answer is that you do not need AI at all.
If the rule is stable and writable, write the rule. A conditional statement costs nothing to run, never hallucinates and needs no monitoring. If your data is not accessible or not trustworthy, fix that first, because a model trained on bad data produces confident nonsense faster than a person could. And if nobody can say what decision changes as a result of the output, the project has no business case, whichever technology you pick.
We would rather tell you that on a discovery call than three months into a build.
Frequently asked questions
Is classic AI obsolete now that LLMs exist?
No. For structured prediction on your own data, traditional machine learning is cheaper per decision, more predictable and easier to validate. Fraud scoring, demand forecasting and quality inspection are still classical machine learning problems in 2026, and running them through an LLM would cost more and work worse.
Which is cheaper, classic AI or an LLM?
Per decision, classic AI, by a wide margin, because it does not call a model each time. The comparison changes once you include the cost of building it: an LLM can be working in days without labelled training data, while a classical model needs a dataset that often takes months to assemble. Cheap to build and cheap to run are different questions.
Can an LLM replace a machine learning model we already have?
Usually not, and usually you would not want it to. If the existing model works, is cheap and can be audited, replacing it adds cost and removes predictability. The better move is to leave it alone and add an LLM in front of it to handle inputs it cannot read.
Do I need my own data to use an LLM?
Not to start. That is the main practical difference from classic AI, which cannot function without a labelled dataset. You will need your data for context and grounding to get useful business answers, but you do not need it to train the model.
How do we decide without a data scientist?
Map the process step by step and label each step either "rules" or "judgement", and each input either "structured" or "messy". That gets a non-technical team most of the way. SoftBlues runs exactly this exercise in a discovery call, and it usually takes under an hour.
Which is safer for a regulated process?
Classic AI is easier to explain and audit, so it is the softer option where a regulator will ask how a decision was reached. An LLM can be used safely in regulated work with a defined scope, full logging and a person approving consequential actions. For FCA, SRA or CQC-regulated processes we design the model to draft and a person to approve.
What do most UK mid-market companies actually need?
In our experience, an LLM reading documents or email, feeding deterministic rules that make the decision, with a person approving anything consequential. The genuinely classical machine learning problems tend to already be solved, or to need more clean historical data than the company has.
Picking the wrong one of these is not a technical error, it is a budget error, and it usually surfaces three months in. If you want a straight answer on which your process needs, and we will tell you when the answer is neither, book a discovery call. SoftBlues is a registered Anthropic Partner Network member and a Google Cloud Partner, and we put working systems into production in 90 days at a fixed price, with a money-back guarantee if it fails.
See it in production
Systems we have built and run for clients, with the numbers that came out of them.
