Machine learning for mid-size business: what is real, what is not
An honest look at which problems machine learning actually solves for companies without research labs, and which ones remain academic.
The conversation around machine learning is getting louder. After the ImageNet results in 2012, after the rise of interest in predictive analytics, after several large companies started publicly describing how algorithms help them make decisions - the topic has moved from academic circles into the business press.
If you run a mid-size company and you read about this, the first question usually sounds like: "Do we need this? And where do we start?" I will try to answer honestly.
What machine learning can actually do right now
Without exaggeration: there are a few classes of problems where it works reliably and produces measurable results.
Classification and categorisation. If you have a stream of objects - applications, transactions, emails, support requests - and you want to automatically assign them to categories, algorithms handle this well given enough labelled examples.
Forecasting from historical data. Predicting customer churn, forecasting sales by segment, estimating default probability - problems where a company has accumulated several years of data are solved considerably more accurately than by expert estimates alone.
Anomaly detection. Identifying suspicious transactions, deviations from normal in production data, unusual patterns in behaviour - an area where algorithms supplement human attention where there is simply not enough of it.
Where realistic expectations end
Three assumptions that most often turn out to be wrong.
"The algorithm will find patterns in the data on its own." Partly true, but the data must exist, be structured, and be reasonably clean. An algorithm trained on three years of chaotic exports from different systems with incompatible formats will produce chaotic results.
"This will replace the analyst." No. It will change what the analyst does - less manual processing, more interpretation and quality control. A person is still needed.
"Train it once and it runs itself." Models age. If the business environment changes, what the model needs to predict changes too. Without regular retraining and quality monitoring, accuracy declines quietly. The tension between model quality and data freshness is worth understanding before committing to any approach.
What you need to get started
Before thinking about algorithms, answer a few questions about your data and your problem.
Do you have a specific business task with a clear definition of success? Not "do something smart", but "reduce churn in this segment by X percent" or "automatically handle Y percent of incoming requests."
Do you have enough historical data? Most tasks need thousands or tens of thousands of examples, ideally labelled. If the data does not exist, you need to accumulate it first.
Is there someone on the team who understands statistics and can work with these tools? Or is there a budget for an external specialist?
A practical filter
I use a simple test to assess whether a task is ready for machine learning.
Write the task in one sentence: what the algorithm should predict or classify, and how the quality of its output will be measured. If you cannot write that sentence, the task is not yet formulated clearly enough.
If the sentence works - check: do you have training data, do you have a mechanism to collect feedback on the quality of results, and is there someone who will own this process.
If all three answers are yes, the task is likely genuinely solvable. If even one is no, start by closing that gap - not by choosing an algorithm.