Phase 01 of the Data Science Lifecycle

Business Understanding:
Translating Business Goals into Measurable Questions

Business understanding is the first phase of the data science lifecycle after CRISP-DM: business goals become measurable questions. Together we define what questions your data should answer.

Business understanding

Why business understanding is critical

Without a written goal there is nothing a model could pass or fail against.

Goal Definition

Collaboratively developing clear, measurable business goals with defined success criteria.

Stakeholder Alignment

Involving all relevant perspectives, from specialist departments to executive leadership.

Feasibility

Realistic assessment of data availability and expected value.

Goal definition

From Business Goal to Analytical Question

A business goal is a direction: fewer returns, fewer unplanned downtimes, fewer resignations. A model needs a question that can be answered from data and predicts exactly one quantity. This translation, the goal definition, is the actual work in business understanding, the first phase after CRISP-DM. Three examples from our worked use cases, all three model calculations for fictitious companies:

Business goal, as stated in the initial callAnalytical questionTarget variable
"We want fewer returns."What share of fit-related returns on jeans can be avoided if the customer is recommended the right size at checkout?The size kept or sent back per customer and cut; the success criterion is the return rate per category
"Our machines fail unplanned too often."Which asset will fail in the coming weeks with what probability, and how many days of warning remain for a maintenance window?The time of failure from the maintenance history, plus the sensor trace before it
"Too many drivers are quitting."In which driver groups does the risk of resignation rise in the next 90 days? Evaluated at group level only, with the works council from the start.Resignation within 90 days per driver, reported per group

Every row fixes three things: the target variable the model is to predict, the population the prediction applies to and the time horizon. What is missing in the row is missing later in the model. If the target variable was never recorded, because for instance nobody noted which size the customer kept, there is no model, only a statistic. How the first row becomes a model in the ordering process is shown by the use case reducing return rates with size recommendations.

Predictive maintenance: seeing the failure before it happens Predicting driver turnover: spotting the group about to tip
Success criteria

Success Criteria: What a Good Model Means for Your Business

A technical criterion says how often the model is right: hit rate, recall, mean error. A business criterion says what that means for you: avoided returns per category, planned instead of unplanned downtime hours, prevented resignations. In phase 1 we fix both, together with the translation from one to the other. In the driver example a recall of 60 percent means that the model recognises three out of five upcoming resignations in advance. Whether that rate is enough is decided not by statistics but by your calculation: what a departure costs and what rescheduling the shifts costs.

The baseline rule

The second decision is the baseline: how good is today's solution without a model? The dispatcher's experience, the size chart in the shop, the fixed maintenance interval. In phase 1 we write down how good this practice is today. Otherwise the model's success cannot be measured. Against "no support" every model wins; that is why we measure against established practice. The result is a written, pre-registered success endpoint: which metric on which population over which period against which comparison. It is checked in phase 5, unchanged.

How the success criteria are checked in phase 5 Why 95% of GenAI pilots fail
AI potential analysis

The Discovery Workshop: Process, Participants, Outcome

At myBytes, business understanding starts with a free 30-minute initial call and leads into a discovery workshop. The workshop brings together the people who make the decision today and know the data: the business department that knows what a field in the merchandise system actually means, and the person responsible for the project in-house. Involving stakeholders means here: whoever is meant to use the prediction later sits at the table.

The outcome is a target picture, the AI potential analysis for exactly one use case: the business goal, the analytical question, the target variable, the success criteria with baseline, a first assessment of data availability and the project scope. Whether the project pays off stands beside it as a business case: what does implementation cost, what does it bring, from when does it carry itself? If an AI strategy for the Mittelstand exists, it brings this calculation with it, and phase 1 makes it measurable. Without a strategy, the target picture initially records the assessment of the expected value. With the target picture, phase 2 begins, in which the data inventory checks the assessment.

At myBytes, phase 1 also ends with a no. Three stop criteria weigh most heavily. First: the target variable was never collected. Anyone who does not have the size kept, the time of failure or the resignation date in their data cannot train a model on it. Second: the history is too short. For forecasts we typically need historical data over 12 to 24 months, depending on seasonality and granularity. For other questions less is enough: in the tyre example at least six months. Third: there is no process that would use the prediction. A failure warning that lands in no maintenance system changes no decision. In all three cases the no goes into the target picture, together with the proposal of what needs to happen first: record the target quantity, build up history, clarify the decision process. A no after phase 1 is cheap; a no after modelling comes only once data preparation, the most laborious phase, is already done.
The consulting process: from initial call to AI strategy in six steps AI Readiness Assessment: checking data, target variable and use case before the project
Data science project approach

Our Approach

01

Discovery & Goal Finding

In structured workshops we jointly identify pain points, opportunities and strategic priorities.

02

Problem Framing

We translate business questions into precise, analytically addressable problem statements.

03

Success Criteria & Scope

Together we define which metrics will determine project success.

Data science deliverables

Typical Deliverables

Documented business goals and questions
Success criteria and evaluation framework
Initial feasibility assessment
Agreed project scope

Frequently Asked Questions about Business Understanding

What is business understanding in the data science lifecycle?

Business understanding is the first of the six phases of CRISP-DM. It translates a business goal into an analytical question with target variable, population and time horizon and fixes what success is measured by. The outcome is a target picture with success criteria and an assessment of whether the data can carry the question.

Which questions does a discovery workshop clarify?

Which decision should be better, who makes it today and with what? Which quantity would a model have to predict, for whom and how far in advance? Which data exist in-house, and who would use the prediction in everyday work? The baseline the model has to compete against is quantified in the same workshop.

How do you make business goals measurable for an AI project?

With a written success endpoint that is fixed before the project: which metric, on which population, over which period, against which comparison. "Fewer returns" thus becomes the return rate per category, measured before and after introducing a size recommendation, compared with the size chart without a recommendation.

What does phase 1 cost and how long does it take?

The 30-minute initial call is free. We deliberately name no blanket duration for phase 1; it depends on how many candidates you bring and how clear the data situation is. Anyone who wants the feasibility check as a delimited package gets it in the AI Readiness Assessment at a fixed price.

What happens if phase 1 shows that the project is not feasible?

Then that goes into the target picture, with reasons and with the proposal of what needs to happen first: record a missing target quantity, build up history or clarify the process that is meant to use the prediction. A no after phase 1 is cheap; a no after modelling comes only once data preparation, the most laborious phase, is already done.

Do we need an AI strategy before phase 1?

Not necessarily. To get started you need three things: a concrete problem with tangible costs, processes that already generate data on it and a person in-house who is responsible for the project. An AI strategy additionally clarifies the order of several candidates and the business case per project; phase 1 takes up this groundwork where it exists.

Which documents should we bring to the initial call?

No presentation. Three things help: the decision that should be better, in one sentence; the systems in which the data on it live, such as ERP, merchandise management, telematics or control system; and the person who would use the result in everyday work. Our questionnaire asks for your starting situation and the existing data sources before the call. You bring the decision and the person.

Let's talk about your project

Bring the decision that should be better and the systems in which the data on it live. The initial call takes 30 minutes.

Book a free initial call