AI consulting guide
Machine learning consulting: for the problems LLMs can’t solve
The generative AI wave made everyone forget that most high-value business predictions — demand, risk, price, churn — are classic machine learning problems. Here’s when you need real ML, what it costs, and how to vet past the notebook portfolios.
Machine learning consulting builds predictive systems trained on your data: forecasting, risk scoring, pricing, recommendations, and computer vision. It differs from generative AI consulting at the root — the model is made from your data rather than prompted with it — which is why ML projects live or die on data readiness and why production discipline (pipelines, drift monitoring, retraining) is half the engagement. Senior rates run $130–260/hr; feasibility assessments $15k–40k; first production systems $75k–200k.
The 2026 irony: LLM hype made classic ML talent scarcer— a generation of engineers skipped straight to prompting — while the business problems it solves didn’t go anywhere. If your question is “what will happen next and what should we do about it,” this is still the discipline that answers it.
Where classic ML beats an LLM
The pattern across all five: structured history in, decision-grade prediction out, thousands of times a day, at a cost per prediction LLMs can’t touch.
Forecasting
Demand, revenue, capacity, cash — anywhere the future is a number and history is a guide.
Risk & anomaly detection
Fraud, churn, credit, quality defects — rare events with expensive consequences.
Pricing & optimization
Dynamic pricing, inventory, routing, scheduling — decisions with measurable dollar outcomes.
Recommendations & ranking
What to show, offer, or prioritize next — per customer, per case, per queue.
Perception
Computer vision and signal processing: inspection, document layout, sensor streams.
The shape of a well-run ML engagement
Feasibility before commitment. Two to five weeks against your actual data: is the signal there, is history deep and clean enough, and — crucially — what does the simplebaseline achieve? A surprising number of “ML projects” end here happily, when a well-tuned rule or a seasonal average captures 80% of the value at 5% of the cost. A consultant who volunteers that finding is one you keep.
Baselines and honest validation. Every model earns its complexity by beating the last simpler thing on a properly held-out test — with leakage checks, imbalance-aware metrics, and evaluation windows that mimic real deployment. This is where amateur ML fails silently: models that “test” brilliantly because tomorrow’s information leaked into yesterday’s features.
Production from the start.The deliverable is not a model file; it’s a pipeline — data in, predictions out, on schedule, with monitoring for drift and a retraining plan. Ask any prospective consultant what happened to their last model six months after launch; the quality of that answer predicts your outcome better than any credential.
Decisions wired in, not bolted on. A forecast nobody’s ordering process consumes is a chart. The engagement should end with the prediction feeding an actual decision — reorder points, review queues, pricing rules — which is where ML work hands off to integration and implementation.
The data-readiness conversation, had honestly
Every ML engagement begins with a gap between the data you believe you have and the data you have. Fields repurposed over the years, outcomes recorded only when someone remembered, history that resets at each system migration, labels that encode three different definitions of “churn.” None of this is unusual — it’s the normal state of business data, and good consultants price a data-preparation phase (commonly 30–50% of the budget) instead of discovering it as a change order.
Two practical implications. First, start capturing outcomes nowfor the predictions you’ll want next year — labeling discipline is the cheapest ML investment available. Second, treat vendor claims of “works with any data” as the tell it is: models are only as honest as the history they’re trained on, and anyone who doesn’t ask hard questions about yours is planning to find out on your budget.
Work with ML engineers whose models are still running
The ML specialists in our network are vetted on production evidence — pipelines, monitoring, and models that survived contact with drift. Brief us on the prediction you need and we’ll scope it honestly, including whether you need ML at all.
Start a ProjectFrequently asked questions
What does a machine learning consultant do?+
They assess whether your data can support the prediction you want, establish honest baselines, build and validate models, and — the part that separates professionals — put them into production with pipelines, monitoring, and retraining plans. Good ML consulting is as much data and operations engineering as modeling.
How much does machine learning consulting cost?+
Senior ML consultants bill $130–260/hr. A feasibility assessment with baselines typically runs $15k–40k; a first production model with pipelines and monitoring $75k–200k; complex systems (vision, real-time optimization) $200k–500k. Data preparation commonly consumes a third to half of any budget — plan for it rather than being surprised by it.
Do we need machine learning or a large language model?+
Rule of thumb: if the task is predicting a number or a category from structured history — demand, churn, risk, price — classic ML wins on accuracy, cost, and explainability. If the task involves reading, writing, or reasoning over documents and language, LLMs win. Many modern systems combine both: an LLM interface over ML predictions. A consultant fluent in only one will force-fit your problem into it.
How much data do we need for machine learning?+
Less than folklore says, more than optimists hope: hundreds to thousands of labeled examples can support useful models for many business tasks, especially with modern techniques. The binding constraint is usually quality and history — consistent labels, enough seasonal cycles for forecasting, and outcomes actually recorded. A real consultant tests this in a feasibility phase before you commit to a build.
What is MLOps and do we need it?+
MLOps is the operational discipline around models in production: versioned data and models, automated retraining, drift monitoring, and rollback. You need it in proportion to how much you rely on the model — a monthly forecast refreshed by hand needs little; a fraud model scoring every transaction needs the full treatment. Proposals that end at "model delivered" are leaving you the hard half.
How do we vet a machine learning consultant?+
Ask for models running in production today and what happened after launch: drift incidents, retraining cadence, business metrics moved. Ask how they validate (holdout discipline, leakage checks, baseline comparisons) — anyone who can't explain a train/test split or names accuracy on an imbalanced problem is a red flag. Notebook portfolios without deployment stories are the ML equivalent of demo-ware.