Machine learning pays off in a business when three things hold: you make the same decision many times, you have records of past cases where the right answer is known, and a better guess is worth money. Demand forecasts, churn and late-payment scores, fraud flags, email routing and reading fields out of invoices all fit that pattern. In 2026 the work splits in two: models trained on your own tables, where gradient-boosted trees are still the sensible default, and hosted language models for text and documents, which most companies use without training anything. Plenty of problems that arrive labelled "AI" need neither.
The examples below come from client work: an LLM workflow that reads an oilfield services company's field tickets every day, an exercise-program generator for a sports physiotherapy clinic, data and AI engineering for an AI bookkeeping startup, and a hospital rostering SaaS whose engine has no machine learning in it at all. This post is about picking the approach and judging whether a use case deserves a budget; how to run the project with an outside team is in our AI consulting guide.
What a model does, and where the value comes from
A machine learning model learns from examples: these inputs led to this outcome. Given a new case, it returns an estimate, such as the probability that a customer cancels next month, next week's sales of a product, or the VAT rate on an invoice line. The estimate is worth something only when a person or a system acts differently because of it: someone calls the customers with the highest churn scores, or a buyer orders less of a product that is about to slow down.
So the first comparison is with what you do today. Google's Rules of Machine Learning open with "Don't be afraid to launch a product without machine learning", and the advice carries over to business systems. A simple rule (reorder when stock falls below two weeks of average sales, review every order above a set amount) is cheap, easy to explain and gives you the baseline a model has to beat.
Matching the tool to the problem
Most requests fall into one of four rows, and the right tool differs a lot between them.
Problem | Business examples | Usual approach | What you need first |
|---|---|---|---|
Predict a number or a yes/no from your records | Next month's demand, which customers will leave, which invoices will be paid late, risky orders | Gradient-boosted trees (XGBoost, LightGBM, CatBoost) | Past cases with the outcome recorded |
Read, sort or draft text and documents | Pulling fields out of PDFs, routing support emails, answering questions from internal documents | A hosted language model with a careful prompt, plus retrieval over your documents if it needs them | A few dozen to a few hundred real examples with the correct answer, for testing |
Find the best plan under rules | Staff rosters, delivery routes, production schedules | An optimization solver (MILP or constraint programming) | The rules and priorities, written down |
Apply logic that is already known | Parsing structured e-invoices, tax calculations, approval limits | Ordinary code | The specification |
Predictions from your own records
When the data sits in tables (orders, customers, payments, machine readings) and you want a number or a category, start with gradient-boosted trees. In a benchmark published at NeurIPS 2022, Grinsztajn, Oyallon and Varoquaux found that tree-based models remained state of the art on medium-sized tabular data, around 10,000 rows, even before counting their speed advantage over deep learning. They train in minutes on ordinary hardware and can show which inputs matter most. Deep learning earns its keep on images, audio and free text, much less on a customer table.
Forecasting works the same way with a time axis: last year's sales, promotions, prices and holidays become inputs to the same kind of model. The bar to clear is the naive forecast, "the same week last year", and a model that doesn't beat it clearly isn't worth deploying. Holidays that move between years, such as Easter, Ramadan or Chinese New Year, need inputs of their own, because "the same week last year" can fall on the other side of the holiday.
Text, documents and images
For this kind of input, most companies in 2026 shouldn't train a model. Hosted models from OpenAI, Anthropic, Google and others already read invoices, classify emails and summarise contracts well enough to start with. Your work is giving the model the right context, checking what it returns and handling the cases it gets wrong. When it needs your own material (a product catalogue, internal procedures, an exercise library), you add retrieval, usually called RAG: find the relevant passages first and put them in the prompt. Fine-tuning comes in only when tests on your own examples show a gap that better prompts and context can't close.
Custom models can age fast. For Jenesys AI, a London fintech whose product reads invoices and receipts and proposes the bookkeeping entry, we built custom models that frontier models later overtook. Some of the work around a model doesn't depend on which model sits underneath: an alert when the category or VAT field comes back empty, or a pipeline that feeds customer corrections back into the system. With models changing this fast, that is the safer place to spend effort.
The exercise-program generator we build for a sports physiotherapy clinic in the US is an example of retrieval over a client's own material. Its first engine, in 2023 and 2024, combined rules with Q-learning, a reinforcement-learning method. In 2025 we rebuilt it as an LLM generator that retrieves from the clinic's library of more than 1,800 exercise entries, searching by exercise, goal and body region. The clinic's coaches have used it with real athletes since 2025: they add notes about the athlete, regenerate a program, swap exercises for alternatives and rate the results.
If your prompts will carry personal data, check what your data-protection law (GDPR in the EU, KVKK in Türkiye) says about sending it to a provider in another country before you build, and strip out whatever the task doesn't need.
Plans under rules, and logic you already know
Oplist, a hospital shift-scheduling SaaS we build for, generates rosters that must respect rest periods after night shifts, department staffing targets and a fair spread of unpopular shifts. That is the nurse rostering problem, a classic optimization problem, and Oplist solves it with a MILP solver in Python. There's no machine learning in it. The rules are known, and a solver either finds a roster that meets the hard rules or reports that it couldn't. A model trained on past rosters would copy last year's habits, including the ones that broke the rules.
Structured data works the same way. An e-invoice that arrives as XML needs a parser, which is exact and never guesses. Keep models for the messy inputs: scanned delivery notes, free-text emails, PDFs that look different from every supplier.
Six questions before you fund a project
Write the answers down before you talk to a vendor or hire anyone.
- What decision changes because of the output, and who makes it?
- How is it done today, and how often is it wrong? That's the baseline a model has to beat.
- How many past cases do you have where the right answer is known? For a prediction, what counts is how many examples you have of the outcome you care about. Forty fraudulent orders among 200,000 is a thin dataset, however big the table looks.
- What does a wrong answer cost in each direction? A false alarm costs someone's review time; a miss costs the loss you were trying to prevent. The two are rarely equal, and their ratio decides where you set the model's threshold.
- How many decisions a month? Saving two minutes on 30 cases a month won't pay for a model's upkeep.
- How will you know it still works six months after launch?
If you can't answer the second and fourth, measure the current process for a month or two before you spend anything on a model.
Getting the data right
Expect most of the effort to go into data: getting it out of other systems, agreeing what fields mean and cleaning up what the source gets wrong.
Define the outcome first. "Churned" could mean cancelled the contract, stopped ordering for 90 days, or downgraded. Each definition gives a different model, and the business has to pick one before anyone trains anything.
Then watch for data leakage, which the scikit-learn documentation describes as building a model with information that wouldn't be available at prediction time. It produces models that look excellent in testing and disappoint in production. The classic case is a churn model that learns from a "cancellation reason" field, which only gets filled in after the customer has left. Split by time: train on older records, test on a later period, and treat any result that looks too good as a bug until you've proved otherwise.
Labels need the same care. If past invoices were categorised by three bookkeepers with three different habits, a model learns a blend of all three. Have the people who do the work agree on the right answer for a sample, and keep that sample as your test set.
Duplicate and conflicting records cause quieter damage. A customer who signs up again with a new email address appears as one cancelled account and one new one, so a churn model trained on that table learns from a customer who never left. Merge records like these before training and keep the cleaning steps in code, because the same steps have to run on the data you score in production.
What it costs
For models trained on your own tables, compute is a small line. A gradient-boosting model on a few hundred thousand rows trains on a laptop, and scoring a million rows takes seconds. The money goes on engineering time: extracting data, defining outcomes, building features, getting predictions into the screens people already use, and monitoring.
Hosted language models are billed per token, which changes the arithmetic at volume. At Anthropic's list prices in October 2026, Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, and Claude Sonnet 5.5 costs $2 and $10. Two examples, with token counts assumed:
- Reading 3,000 documents a month, about the volume of the field-ticket workflow described below, at 5,000 input and 500 output tokens each: roughly $2 a month on Haiku 5.5 and $45 on Sonnet 5.5.
- Scoring 1,000,000 transactions a month with one LLM call each, at 1,000 input and 50 output tokens: roughly $125 a month on Haiku 5.5 and $2,500 on Sonnet 5.5, plus the delay of a network call on every transaction.
For documents, the model bill is the smallest part of the project. For high-volume predictions on structured data, a gradient-boosting model costs next to nothing per prediction and was built for exactly that kind of input. The Batch API halves token prices for work that can wait, which helps overnight document runs but doesn't change that comparison. Measure token counts on your own documents, because layout, images and language move them a lot.
After launch
The first thing that goes wrong is drift. Customers, prices and suppliers' document layouts change, and a model trained on last year gets slowly worse without raising a single error. Track accuracy on recent cases once their true outcome is known, and retrain or re-test on a schedule. Our MLOps guide covers the tooling.
The model or service you depend on can also be withdrawn. Anthropic gives at least 60 days' notice before retiring a publicly released model; it told developers in April 2026 that Claude Sonnet 4 would be retired and switched it off on its API on 15 June 2026. Packaged ML services go too: AWS closed Amazon Forecast to new customers in July 2024 and pointed them to SageMaker Canvas. If you've kept your test set, you can run it against the replacement before the deadline and see what changed, instead of hearing about it from users.
Providers also throttle you. The field-ticket workflow we built with an automation agency partner for an oilfield services company has read about 100 documents a day since September 2025. What broke was mostly the LLM API's rate limits (HTTP 429 responses), and the fix was falling back to alternative models, so a ticket still gets processed when the first model turns the request away. Plan that fallback before go-live.
Then there are wrong answers that look fine. A model that returns an empty field or an impossible value doesn't throw an exception. Validate outputs against a schema, alert on missing or out-of-range values, and send uncertain or high-value cases to a person. Where experts review the output, store the model's proposal together with the expert's final choice; that is what we're adding to the physiotherapy generator in 2026, so its logic can be checked against past athletes. You get evaluation data almost for free that way.
How to start
- Pick one decision with a measurable baseline and a known cost of error.
- Collect test cases: past records with known outcomes for a prediction, or 50 to 200 real documents with agreed answers for a language-model task.
- Build the simplest version that could work: a rule, a gradient-boosting model, or a hosted model with a careful prompt.
- Run it next to the current process for a few weeks and compare the two on the same cases.
- Ship it, adjust it or stop. Stopping with numbers in hand is a good outcome, and far cheaper than stopping a year later.
If you'd like a second opinion on a specific use case, our data and AI page describes the work we take on.
Frequently asked questions
Do we need to hire a data scientist?
Usually not for a first project. An engineer who knows your data and has shipped a gradient-boosting model or an LLM integration can take it a long way. Specialist depth pays off once models are central to your product, or when errors are costly enough that someone should own model quality full-time.
Can we run a language model on our own servers?
Yes. Open-weight models such as Meta's Llama, Alibaba's Qwen, Mistral's models and OpenAI's gpt-oss can run on your own hardware or in your own cloud account, so prompts never leave your infrastructure. You pay for GPUs and the work of running them instead of paying per token, and the strongest hosted models are still ahead on hard tasks. Whether that gap matters for your task is something only your own test set can show.
Is machine learning the same as AI?
Machine learning is the part of AI where systems learn from examples instead of following hand-written rules. Large language models are machine learning models too; what changed in the last few years is that you can use a very capable one without training it yourself.
Related Articles
MLOps Guide: What to Set Up First, and What Can Wait
MLOps gets a model into production and keeps it useful there. Here is the order we'd set it up in, with current MLflow and DVC examples, and what changes for LLM features.
Read More: MLOps Guide: What to Set Up First, and What Can WaitComputer Vision Applications: Where It Works and Where It Struggles
Where computer vision works today, from inspection lines to invoices and medical scans, what makes some applications hard, and when a general vision model saves you training your own.
Read More: Computer Vision Applications: Where It Works and Where It StrugglesAI Agent Development: When You Need One and How to Build It
An AI agent is a language model working in a loop with tools. When you need one, the loop in Python, choosing a framework, keeping it safe, and what each run costs.
Read More: AI Agent Development: When You Need One and How to Build It
