You need a machine learning consultant when three things are true at once: you hold labelled records of the past, you repeat one decision again and again, and a wrong decision costs you a number you can name.
Condition: all three must be true together. Two out of three does not make a model project sensible.
Limit: if your real problem is a slow manual process, you need an AI workflow consultant, not a new model. I wrote this because half of the requests that reach my desk are not machine learning problems. That is my observation, not a measured figure.
Start with the symptom, not the technology
Find the row that looks most like your situation today. Read the reason column before you read the middle one.
| Symptom | What you need | Why |
|---|---|---|
| Your team copies data between apps each day | AI workflow audit | No prediction happens here. Work only changes hands. |
| You guess next week stock from 3 years of records | Machine learning consultant | Labels exist, the decision repeats, the cost is clear. |
| Customer chat replies are slow after hours | AI agent with reply scripts | The answers already exist. Speed is what you lack. |
| Order data sits in 5 Excel files | Data plumbing work | Any model fails without one source of numbers. |
| You want to know which customers will leave | Machine learning consultant | Customer history can carry a left or stayed label. |
| Your order form keeps getting filled wrong | Form fix and validation rules | Plain rules close this case with no model at all. |
When you need a machine learning consultant
Machine learning is a program that learns patterns from examples. You do not write the rules. You hand it thousands of past cases, and it builds the rules itself.
Here is the analogy I use. You can train a new cashier in two ways. In the first, you write a price rulebook. In the second, you show 5,000 old receipts and let the pattern sink in. Machine learning is the second way.
That mechanism creates one hard requirement. The scikit-learn documentation states that the fit method takes 2 inputs: the samples matrix X and the target values y, and that y does not need to be specified for unsupervised tasks. So a prediction needs a column of answers that already happened. That column is your label.
I order the three tests like this:
- Labelled data. You hold an outcome column for past events. For example: this order was cancelled, this customer left.
- A repeated decision. The same decision appears daily or weekly. A once-a-year decision is not worth training for.
- A measurable cost of error. You can name the unit of loss: money, staff hours, or lost customers.
If one test comes back empty, stop there. Fill that gap first, and it costs less than a model project.
When your problem is really a workflow problem
Many requests arrive with the words machine learning and ask for 3 other things. I name them like this.
An agent. A program that reads an incoming message, pulls data from one place, then writes a reply or a record. No new prediction happens inside it.
A rule. One if-then line written by a person. If the order value passes a threshold, ask for approval. A rule wins when you already know the pattern.
A better form. Dirty data is often born at the point of entry. Fix the fields, the validation, and the fixed choices, and half of the complaints disappear.
If you cannot tell which case you are in, start with an AI workflow audit. That step costs less than a model project, and the result tells you whether a model is needed at all. To put a number on the work itself, read AI cost per operational task.
What a model project really costs in time and data
Training the model is not the longest part. The longest part is preparing the data and proving the labels are right. The table below is a rough estimate to set expectations, not a measurement.
| Stage | Estimated share of time | What you prepare |
|---|---|---|
| Decision and label definition | 15 percent | One decision sentence and a label definition |
| Data collection and cleaning | 45 percent | Source system access and a person who knows the data |
| Training and testing | 15 percent | A baseline number and a separate test set |
| Fitting it into the workflow | 15 percent | The people who will act on the prediction |
| Monitoring after launch | 10 percent | An owner who reads the monthly report |
Length of time is not the only difference. A workflow project ends when the work is done. A model project never fully ends, because new data keeps arriving.
Four questions to ask before you sign
Ask these 4 questions in the first meeting. The answers separate a person who has shipped a model into production from a person who has only read about it.
First: which baseline are we going to beat? The scikit-learn metrics guide notes that dummy estimators are useful to get a baseline value of those metrics for random predictions. Without a baseline, 85 percent accuracy means nothing.
Second: which test data was never used for training? The scikit-learn cross-validation guide calls learning the parameters of a prediction function and testing it on the same data a methodological mistake, and names that situation overfitting. Common practice holds out part of the data as a test set.
Third: how will we know the model has weakened in 6 months? The Google Cloud MLOps guide states that models can lose performance because data profiles keep evolving, so you need to track summary statistics of your data and monitor online performance to send notifications or roll back.
Fourth: who owns the data, the model, and the code when the project ends? Put the answer in the contract, not in meeting notes.
Red flags that justify walking away
These 4 signs show up before money moves. You can spot all of them with no technical background.
| Red flag | What it means | Your reply |
|---|---|---|
| Accuracy promised before any data review | The number is a sales guess | Look at the data first, then give a range |
| No baseline mentioned | The model result cannot be judged | What score does a random guess get? |
| No separate test set | A training score is reported as the result | Show the score on data it has not seen |
| No monitoring plan | The model is treated as done on handover | Who reads the report in month 7? |
Monitoring is not my own idea. Amazon SageMaker Model Monitor lists 4 monitoring types: data quality drift, model quality drift, bias drift, and feature attribution drift. Large providers build dedicated tooling for this problem. A consultant who skips it is skipping the expensive part.
Simulation with dummy data: 5,000 order rows
The numbers below are a simulation with made-up data. They show the order of work, not a client result.
Starting condition. An online parts shop takes 40 orders a day. The owner reports that some orders end up cancelled after the goods are packed.
Input. 5,000 order rows from the last 2 years. Each row carries a final status column: completed or cancelled. That column is the label.
Steps. First, compute the baseline: guess that all orders complete, then record the share you got right. Second, set aside the last 1,000 rows as test data and do not touch them. Third, train on the remaining 4,000 rows. Fourth, score the model on the 1,000 rows you set aside.
Observable output. Two numbers side by side in one table: the baseline score and the model score on the test data. The gap is what you are buying.
Decision. If the gap is small, cancel the model project and fix the order form. If the gap is wide, price it: how many packing runs can you hold back before the goods leave the warehouse. Only after that calculation do you talk about rollout.
Checklist before you contact anyone
- Write the decision you want to automate in one sentence.
- State how many times that decision happens per week.
- Point at the label column in your data, and name the file it lives in.
- Price one mistake in the unit you use day to day.
- Note where the data lives now and who can open it.
- Set a baseline today, even if that baseline is your staff guessing by hand.
- Decide who will act on the prediction inside their own job.
- Prepare one closing question: what would make this project a failure?
Our limits, stated plainly
Rama Digital does not sell machine learning research and does not train models. Our work sits in 3 areas: applied AI, agent operations, and data plumbing.
Here is what that means for you. If your case truly needs a trained model, I say so in the first meeting. After that I help you scope it: the label definition, the data requirements, the baseline, and the questions to ask a modelling team. You still hire that team elsewhere.
Many cases stop before that stage, and that is good news for your budget. If you want the sequence for a first 3 months, read the 90 day enterprise AI transformation plan.
Ready to take apart your own case? Bring one decision and one data file to our AI consulting service. We start from the symptom, then decide whether you need a model at all.
Questions and answers
What is the difference between a machine learning consultant and an AI consultant? A machine learning consultant builds a prediction model from your labelled data. An AI consultant fits existing tools into your workflow, including language models and agents.
How much data do I need before training a model? The amount depends on how many columns you have and how rare the event is that you predict. Ask a candidate to look at your data first, then give a range instead of a fixed number.
Can I use a language model instead of training my own? Often yes, in particular for reading text, summarising, and routing messages. Language models are weak at numeric prediction that depends on your internal records.
What is drift and why should I care about it? Drift is a change in data patterns after the model goes live, such as a price rise or a new type of customer. The Google Cloud MLOps guide states that models can lose performance as data profiles evolve, so monitoring belongs in the budget.
Why does a consultant refuse to promise accuracy up front? Accuracy depends on data they have not seen yet. A number promised before a data review is a sales promise, not a technical estimate.
What is the cheapest first step I can take? Set a manual baseline this week. Record how often your staff guess correctly, then use that figure to compare against any proposal you receive.
Sources
- scikit-learn, Getting Started - states that fit takes the samples matrix X and the target values y, and that y is not needed for unsupervised tasks.
- scikit-learn, Cross-validation - calls training and testing on the same data a methodological mistake, and recommends holding out a test set.
- scikit-learn, Metrics and scoring - notes that dummy estimators are useful to get a baseline value for random predictions.
- Google Cloud, MLOps - states that models can lose performance as data profiles evolve, so data statistics and online performance need monitoring.
- Amazon SageMaker Model Monitor - lists data quality drift, model quality drift, bias drift, and feature attribution drift.
