Four decisions to build a data company
A data company is built in layers, and each layer forces you to choose between options that cost different things. This article organises what we covered in the Business Data and AI Strategy module as four decisions: where the use case comes from, which data strategy to follow, which culture to build and which governance model to adopt.
In each one you will find the options, what each offers, a criterion for choosing and, at the end, a box with what our team chose. The example is Clink, a fictional home-search startup we used as a class case: it illustrates, it is not the subject.
The four decisions at a glance
Each layer builds on the previous one and is decided with a different question.
| Layer | What is decided | Options | The question that helps you choose |
|---|---|---|---|
| 1. Innovation | Where the use case comes from | Start from the problem, or start from the data | Is there someone with a clear pain, or do I have data with potential? |
| 2. Strategy | Which data strategy to follow | Control, optimisation or transformation | Am I protecting a business, improving it or creating a new one? |
| 3. Culture | Which culture to build and how | Constructive, passive or aggressive; rites; literacy | Which behaviours do I want repeated without anyone ordering them? |
| 4. Governance | Which operating model to adopt | Decentralised, networked, centralised, hybrid or federated | How much control do I need and how much speed am I willing to give up? |
Decision 1, innovation: where the use case comes from
A use case comes from a problem or from data, and it pays to make both passes.
| Aspect | Start from the problem | Start from the data |
|---|---|---|
| How it works | Traditional design thinking: empathise, define, ideate, prototype and evaluate | You start from data as an asset with potential and look for its fit in a process: data potential, product fit, design and rollout |
| Strength | The value is clear from the start | It makes use of what you have and what you can reach |
| Risk | The data to solve it may not exist | A solution looking for a problem |
| Starting question | What does the person we design for need? | What new knowledge can we build with this data? |
There are more choices within each path. Who you solve it for: the same sector has different customers, and in housing they are the person looking for a home and the estate agency, each with their own pain points and metrics. How you make money: B2B if companies pay, B2B2C if you reach the user through another company and B2C if the user pays with a subscription.
The tool. The workshop was run on a Miro board with eleven boxes, grouped in three steps: understand the problem (problem and alternatives), design the solution (candidate solutions, solution and key objectives) and turn it into a business (unique value proposition, who it is for, challenges, revenue, high-level concept and key metrics). The exercise ends with a four-minute mini pitch per team.
Step 1 of 3
Understand the problem
Problem and alternatives: who suffers what, how much it costs them and what they do today to solve it.
Step 2 of 3
Design the solution
Solution and key objectives: which solution we choose among the candidates and what it must achieve.
Step 3 of 3
Turn it into a business
Value, income and metrics: the value proposition, who it is for, how it makes money, the high-level concept and how it is measured.
A useful test. Ask the most uncomfortable question: why would the biggest company in the sector want to work with me? If there is no answer, the idea needs another approach.
Criterion. Start from the problem if you already have someone to listen to; start from the data if your asset is going unused; and then make the other pass, because the path is not linear.
In Clink
We started from the problem of the person looking for a home, focusing on two stretches of their journey: searching on portals and getting in touch to arrange viewings. The solution is a conversational agent that crosses the listings of several portals and returns an explained match percentage. The data would arrive through APIs, under mutual-agreement contracts with each portal. We answered the uncomfortable question by bringing the small portals together instead of competing with the big one.
Decision 2, strategy: which data strategy to follow
A data strategy is a purpose plus a plan, and its type depends on whether you want to protect a business, improve it or create a new one.
A strategy describes the starting situation, the destination and the path between them. It boils down to four questions: where do we want to get, what do we need, what could stop us and what will others do.
The tool. The workshop was run on a Miro board with eight boxes, in three steps: starting situation, purpose, vision and three objectives and key results; action plan and management control indicators; and risks and mitigation measures.
| Type | Key question | Analytics | What it seeks |
|---|---|---|---|
| Control | What happened and how much did we win or lose? | Descriptive and diagnostic, classic BI | Protect value: avoid losses and comply with regulation |
| Optimisation | What will happen and how do we do it better? | Predictive, prescriptive and AI | Operational efficiency and more margin |
| Transformation | What new business can we invent? | Ecosystem and data monetisation | New, non-traditional revenue lines |
Each business question is answered with a different type of analytics.
What makes it good or bad. A good strategy has a diagnosis, a chosen approach and coherence between actions and resources. A bad one has vague objectives, keeps changing priorities, ignores its weaknesses and does not measure progress.
What sustains it. Three elements: the strategy, the change plan and management control. Hard changes are easy to plan; cultural inertia is not. And control means being hard on the task and soft on people, with a continuous cycle of planning, executing, measuring and controlling.
Before deciding, a diagnosis. Two tools: dataMat, a 105-question survey that scores the organisation across 21 areas of the data world, and dataToolkit, a card game, one card per area, in which you place yourself on one of three levels in order to design the strategy from the weak points. I describe them in more detail in Data strategy: purpose first, then the plan.
Criterion. Control if the urgent thing is not to lose what you already have; optimisation if you have a business that works and data to improve it; transformation if the business does not exist yet.
In Clink
We chose transformation, because it does not improve an existing business: it creates one. Three objectives: organic traffic, a solid team and a reliable AI model. The most critical risk is legal, if the portals stop sharing listings. The second is hallucination: hard data, such as the price, always comes from the database, and the agent cites its source and knows how to say it does not know.
Decision 3, culture: which culture to build and how
The main brake on a data project is not technology; it is people, culture and change management. In a 2026 data and AI leadership survey cited in class, 93.2% pointed to that brake, against 7% who pointed to technology.
A weak culture is hard to define and explain, has little alignment with values and is controlled with lengthy procedures and bureaucracy. A strong culture is understood and can be explained by anyone, with high alignment and few rules.
How it is built
- Metaphors that compare the organisation to a machine, a group or a family.
- Stories that teach how to act in each case.
- Rites and ceremonies, which combine the two above. They can be rites of initiation, of improvement (with public recognition), of integration or of conflict reduction.
How it is looked after
With a data literacy plan:
- Distinguish data literacy from technology literacy.
- Understand that each person learns differently.
- Reward curiosity.
- Define which data skills each person needs.
- Involve the leaders.
- Use a common language.
- Define what success is.
- Remember that literacy is only the beginning.
| What is organised | One option | The other option |
|---|---|---|
| The data platform | Monolithic view: a data lake that centralises everything, from ingestion to use | Data mesh: domains, self-service and data treated as a product |
| The data people | A centralised team | Embedded in each area, with a data lead per domain |
Who sponsors. Not only whoever has the authority the org chart gives, potestas, but also whoever has legitimate influence, auctoritas: reference people in their area who represent poles of informal power.
Criterion. Choose the culture by the behaviours you want to see repeated and design rites that reinforce them. Literacy works when people can explain, communicate and discuss, and ask questions instead of just accepting.
In Clink
- Motto: “The first brick is a piece of data. The rest, together.”
- Principles: five, about arriving together, truthful data and transparency without ego.
- Ceremonies: a weekly stand-up with the whole company; demos and retros; a monthly ten-minute tech talk; a quarterly award for the best data-based solution; a resilience party.
- Sponsorship: the CEO, the CDAIO and the CFO, together with data experts from each area.
Decision 4, governance: which operating model to adopt
Data governance decides who can decide what about data; management carries it out day to day. The operating model sets out roles, responsibilities and decisions, and it comes in five versions.
| Model | How it works | Advantages | Drawbacks |
|---|---|---|---|
| Decentralised | People collaborate in committees and there is no single owner | A flat structure, aligned with the business and easy to set up and grow | Many participants; deciding among many is hard; ownership of the data stays unclear |
| Networked | The decentralised one, formalised with a RACI matrix | Gives the data an owner without creating a central team that becomes a bottleneck | Requires defining who is responsible, who approves, who is consulted and who is informed |
| Centralised | The most formal and mature: one person in charge | Strong leadership and simpler decisions | A big organisational challenge and a change of paradigm; you have to cover everything from the start; hard to grow step by step |
| Hybrid | Combines advantages of the centralised and decentralised models | Combines a flat structure with alignment to the business | Inherits some of the decentralised model’s problems: hard collaborative decisions and unclear ownership |
| Federated | A variant of the hybrid: centralised strategy and decentralised execution, with a center of excellence that sets the guidelines | Ideal for large corporations; gives subsidiaries autonomy; allows one CDO per subsidiary, is stakeholder-oriented and supports an incremental approach by unit or region | An ambitious, complex process; a balance has to be found and it needs many resources |
The roles. Every piece of data needs an owner and someone to look after its definition: the data owner, the data lead or data steward and, at the top, a position that brings together data, analytics and AI, the CDAIO.
What gets governed
- Business glossary: defines each concept in business language.
- Technical dictionary: says how each piece of data is stored.
- Business and technical metadata: connect the two above.
- Quality: measured along eight dimensions: accuracy, completeness, consistency, integrity, reasonableness, timeliness, uniqueness and validity.
What to lean on
- DAMA-DMBOK: organises the knowledge areas of data management.
- SAM model by Henderson and Venkatraman: sets out two types of alignment, strategic and tactical.
- Amsterdam Information Model: known as the nine-cell model.
Criterion. The operating model is not forever: it is chosen by the size and maturity of the organisation and evolves as it grows. Choosing a model is choosing between control and speed. Decentralising is the easiest way to start; centralising means covering everything from the start.
In Clink
We chose the networked model with a RACI matrix, which is also what the teacher recommended for a team like ours. The CDAIO approves the data and AI standards, area heads are consulted and the whole company is informed. There are five teams, each with an owner and a data lead, and two mechanisms against silos: data contracts between teams and a shared dictionary and catalogue. The glossary started from the funnel of contact, appointment, virtual attendance and sale. As it grows, it would evolve into a federated model.
The thread that ties the four decisions together
Each decision answers a different question: what for (strategy), how we behave (culture) and who decides (governance), with innovation as the starting point. None of them works without the others.
A strategy without culture stays a document, a culture without governance turns into good intentions, and governance without purpose is bureaucracy.
How to decide: a checklist per layer
These questions help you reach a decision in each layer before writing anything down.
- Innovation: who suffers the problem and how do I know? What data do I have or can I get? What is the most uncomfortable question for my idea?
- Strategy: what are we doing it for? Do we play in control, optimisation or transformation? What could stop us getting there?
- Culture: which behaviours do we want to see repeated? Who sponsors, with potestas and with auctoritas? How will we know literacy is working?
- Governance: who decides what? Who owns each piece of data? How do we define and measure its quality?
What I take away
- Each layer is a decision with options that cost different things: there is no single right answer, there is a criterion.
- You can start from the problem or from the data, and it pays to make both passes, because the path is not linear.
- A strategy is not a wish list: it is diagnosis, approach and coherence.
- Culture is practised and measured; it is not declared.
- Data without an owner and without a definition is not an asset, and the model that gives it an owner changes with the size of the organisation.
A personal reflection. In most businesses, the business team is the owner of the data: it knows the company’s needs and can also speak the language of technology. That makes me think I am well placed to keep learning from both sides, business and data.
Thanks to David Rey, Rafael Fernández Campos, Ubaldo González and Alberto Turégano for the classes this article is based on.
Follow the blog by RSS