Machine Learning
Machine learning is software that learns a pattern from historical data instead of following logic a person wrote. For an operator, the distinction that matters is the boundary between problems a rule can solve and problems only a model can. Vendors blur that boundary, because a model call bills per use and a rule bills once. A large share of what gets sold as machine learning is work a lookup, a threshold, or a regular expression handles at a fraction of the cost. Our design rule is that AI layers on top of deterministic software and systems you own rather than substituting for them; the full argument, including what substitution costs in token spend and reliability, is in our model lock-in entry.
How it actually works
A model is trained on examples: rows of inputs paired with known outcomes. It finds the combinations of inputs that predict the outcome, then applies what it found to new rows. The quality of that prediction is capped by the quality of the examples. If the signal is not in the data, training cannot find it.
The boundary question comes first. When the pattern can be written down, a rule is the better choice: it returns the same output for the same input, costs nothing per run, and can be audited line by line. A model earns its place when the pattern space is too big for that: hundreds of interacting factors, or structure buried in qualitative data such as reviews, call notes, and survey responses. Feature extraction, turning that raw material into inputs a model can read, is where most of the real work happens and where most pitched projects have done none.
The industry pattern runs in the opposite order. The build is scoped first, the ROI question is asked after launch, and the project is judged on whether the model runs instead of whether a decision improved. The fix is procedural: name the decision the model is supposed to improve, estimate the value of improving it, and compare that to the cost of building and running the model. If the estimate does not clear the cost, the project should not start. The base rate for skipping that order is documented: in 2025, MIT's NANDA initiative published findings that 95% of generative AI pilots delivered zero measurable impact on the P&L, despite $30 to $40 billion in enterprise investment. AI amplifies whatever you feed it: feed it a single KPI tracked in isolation and it amplifies the mistake faster, at greater scale, and with more confidence than any team member would have had making the same error manually. We cover this in [AI and bad data](/insights/why-strategic-initiatives-fail#ai-and-bad-data).
In practice
For one client, customer acquisition cost went from $52 to $13.50 in a matter of weeks. Repeatable patterns identified across the market were weighed by machine learning over seven hundred tests, so budget followed the factors that predicted acquisition instead of the factors a platform reported.
Where we come in
This work sits inside our Revenue Systems Architecture practice, built for businesses from new startups to the Fortune 15, including Digital Realty and Equinix. We demonstrate ROI before each build. Where a deterministic rule outperforms a model, we ship the rule and reserve models for pattern spaces that genuinely need them. The systems are portable, so you are never locked to a single model, platform, or vendor.
Start a Revenue Health Pre-Assessment →See it in action
Related terms
- Deterministic rule
- Logic that produces the same output for the same input, a lookup, a threshold, a regex. When a rule can do the job, it beats a model on cost, speed, and auditability.
- Feature extraction
- Turning raw and qualitative data into the structured inputs a model trains on. It is where most of the effort in a working ML project goes.
- Training data
- The historical examples a model learns from. A model's ceiling is set by these examples rather than by the choice of algorithm.
- Overfitting
- A model that memorized its training examples instead of learning the pattern. It scores well on history and fails on new data.
- Inference cost
- What each prediction costs to run in production. Rules cost near zero per run; model calls are metered, so the economics depend on volume.
- Large language model (LLM)
- A model trained on text that can read, summarize, and generate language. Priced per token, which means cost scales with usage rather than with outcomes.