The business benefits of machine learning come down to three things: it takes over repetitive judgement that people currently do by hand, it spots patterns in volumes of data nobody has time to read, and it makes a reasonable prediction where you would otherwise guess. Those benefits are real, but they apply to particular kinds of task. This article describes where they show up in an ordinary business, what has to be in place first and how to find out cheaply whether a given idea is worth pursuing.
What machine learning does
Ordinary software follows rules a person wrote: if the order is over a set value, ask for approval. Machine learning is used when the rules are too many or too vague to write down. It is shown past examples with known answers and works out the pattern for itself. Given enough examples of invoices and the fields that were keyed from them, it learns to find the fields on a new invoice.
Two kinds are in everyday business use:
- Predictive models, trained on your own records, that output a number or a category: how many orders next week, whether this payment looks unusual.
- Language models, trained by a large provider on enormous amounts of text, that read and write: extracting details from a document, summarising a long email thread, drafting a reply.
Both are wrong some of the time. That fact shapes everything that follows.
Where the benefits show up
Reading documents into your systems
Orders, invoices, delivery notes, application forms and emails arrive as documents and somebody types their contents into a system. In our experience this is the most dependable place for machine learning to earn its keep. A model reads the document and fills in the record, and anything it is unsure of goes to a person. The benefit is hours of keying returned to staff and fewer transcription mistakes.
Sorting and routing
Incoming emails, support requests and enquiries can be classified and sent to the right team or queue. The benefit is speed of response, and a specialist’s time not spent reading messages meant for someone else.
Forecasting
With a few years of history, a model can often forecast demand, stock requirements, staffing or cash collection more accurately than last year’s figure plus a guess. The benefit is less overstocking and fewer shortages. A forecast is only as good as the history behind it, and it will not foresee an event that has never happened before.
Flagging the unusual
A model can learn what normal looks like, for payments, expense claims, meter readings or machine sensor data, and flag what does not fit. This is the basis of fraud detection and of predictive maintenance, where equipment is serviced when its readings start to drift instead of on a fixed schedule or after it fails. The benefit depends on having the data: predictive maintenance needs sensors and a record of past failures to learn from.
Finding answers in your own records
Staff spend a surprising amount of time looking for things: the clause in a contract, the last time a fault occurred, what was agreed with a customer. A language model connected to your documents and records can answer a question in plain English and point to the source. The benefit is time, and knowledge that no longer depends on asking the one person who remembers.
What it needs before it pays off
In our experience, projects that deliver have the same few things in place.
- A specific task. “Use AI in the business” is not a project. “Cut the time to enter a supplier invoice” is.
- Volume. The task has to happen often enough that saving a few minutes each time adds up.
- Examples. For a predictive model, a history of past cases with known outcomes, held somewhere it can be extracted from. For a language model, a sample of real documents to test against.
- A baseline. How long the task takes now, and how often it goes wrong. Without that you cannot tell whether the new way is better.
- A person in the loop. Someone who checks the output, at least for the cases the model is unsure of.
- A system for it to plug into. A prediction that sits in a separate tool nobody opens is worth nothing. It has to appear in the screen where the work is done.
That last point is where many attempts stall. If the underlying data is scattered across spreadsheets, or the systems involved cannot exchange information, that has to be fixed first. It is ordinary system integration work, and it usually brings benefits of its own.
The costs and risks that get left out
Errors. No model is right every time. The design question is what happens when it is wrong. A misread invoice that a clerk corrects in seconds is cheap. A wrongly declined customer is not. Machine learning suits tasks where mistakes are easy to catch and cheap to put right.
Data protection. If the data is about people, UK data protection law applies to how it is used, and there are specific rules about decisions made about individuals by automated means. The Information Commissioner’s Office publishes guidance on AI and data protection. Establish where the data will be processed, whether the provider keeps it and whether it could be used to train their models, and take advice before using personal data.
Running costs. Hosted models are charged by usage. The cost per document or per question is usually small, but it should be estimated from your real volumes before you commit.
Upkeep. A predictive model reflects the period it was trained on and drifts as the business changes, so it needs checking and retraining. Language models are replaced by their providers, and the service you built on may be withdrawn. Our comparison of machine learning as a service providers shows how often that has happened.
Trust. Staff who are handed a tool that is confidently wrong every so often stop using it. Showing how sure the model is, and where an answer came from, matters as much as raw accuracy.
When ordinary software is the better answer
Machine learning is sometimes proposed for a problem that plain software solves more cheaply and more reliably.
- If the rule can be written down, write the rule. “Flag any order above the customer’s credit limit” needs no model.
- If the trouble is re-keying between two systems, connect the systems.
- If the trouble is a process run on spreadsheets and email, replacing the spreadsheets with a proper system will do more than adding a model on top.
A good supplier will tell you when this is the case. Machine learning is worth its extra uncertainty only when the task involves judgement that cannot be reduced to rules.
A first step that costs little
Pick one task that is frequent, tedious and currently done by hand, where a mistake would be caught. Collect a few dozen real examples, with the correct answers. Measure how long the task takes now. Then have someone test a model against those examples and report how many it got right, before anything permanent is built.
If the result is good, you have evidence to justify building it into the system. If not, you have spent little finding out. That is how we approach adding AI to existing systems. If the test points towards training a model of your own, our guide to machine learning frameworks explains the options.