Leanne Fitzpatrick recommends designing data teams around their actual budget, staffing, architecture, and business needs before pursuing an ideal structure.
2
At the Financial Times, adding engineering capacity mattered more than simply increasing the 11-person data science team because batch systems limited near-real-time products.
3
Data teams need embedded people who can connect technical capabilities with business users, especially when models must change how teams work rather than feed isolated campaigns.
Summary
Leanne Fitzpatrick discusses how to build data teams when budgets, existing architecture, and business maturity constrain the options. She argues that leaders should start with those constraints, then decide where to add capacity. At the Financial Times, the data science team has 11 people inside a company of about 2,000. Rather than growing data science directly, she sought engineering support to move beyond batch work and improve low-latency experiences. She also describes the role of embedded people who can translate between data teams and commercial users. The conversation covers when to buy third-party capabilities, when to build internally, and why infrastructure changes should happen behind a stable developer experience. Fitzpatrick explains her long experience putting R models into production and discusses R's place beside Python. Her answers are practical and candid about the limits imposed by organizational maturity, unclear role definitions, and the difficulty of winning trust for data investments.
Data teams should start with constraints before designing an ideal structure
Leanne Fitzpatrick says a team must be built for the business in front of it. Headcount, operating expenses, contractor budgets, and the scale of the wider company limit what a data leader can do. At the Financial Times, roughly 11 data scientists work in a business of about 2,000 people. The team originally focused on subscriptions and decision-making for business-to-consumer and business-to-business work. Fitzpatrick chose to grow engineering support instead of simply adding more data scientists, because the main limitation was the ability to turn models into operational products. Her approach is to understand the available resources and architecture first, then design the team around those facts.
The Financial Times needed engineering capacity to move beyond batch models
The Financial Times works with business intelligence engineers to run batch models through its BigQuery layer. Fitzpatrick says this works for much of the current work, but it limits growth because the team cannot easily operate in real time. Some low-latency personalization exists, such as work on the payment barrier shown to readers, but the capability is still limited. She obtained two additional people from the data platform organization to embed with data science. Those teams sit in different parts of the company, under the CTO and chief analytics officer, but already worked closely together. The staffing decision followed the need to make data science operational rather than the assumption that more data scientists would solve every problem.
Data teams still have to explain roles that businesses understand poorly
Fitzpatrick says businesses can usually explain why they need an engineer, product manager, or project manager. It is harder to explain the difference between a data scientist, data engineer, machine learning engineer, and MLOps engineer. She describes data as having often received less trust and authority than other technology functions. Business leaders may allow technology teams to decide how engineering resources are used while questioning whether data teams should make the same decisions. In her view, this comes from the relative immaturity of data work inside many companies and from limited understanding at board level. Even organizations with senior data executives can have friction over what different data roles do.
The route to investment depends on whether data is the product
Leanne separates organizations whose revenue comes directly from data capabilities from organizations where data supports another business. A data platform provider or a company selling data services must invest in its data function because that capability keeps the business operating. In other companies, a model can remain a small input to a campaign without changing how the wider organization works. Fitzpatrick sees the larger opportunity in changing business behavior. A churn or propensity model should not only feed a few marketing campaigns. It should help teams rely on data services when making commercial decisions. That requires organizational change as well as a model in production.
Embedded translators connect data builders with the people who use their work
At the Financial Times, data science sits within analytics, and the organization has people embedded with business teams and their tools. One person works with the customer data platform used by the business-to-consumer marketing team and acts as a contact between that team and data science. Fitzpatrick says this middle layer understands how a capability will be used and helps champion it inside the consuming team. These people may come from technical, visualization, product, or data backgrounds. They also need strong listening and storytelling skills. She does not see them as mythical hires. The career path is simply still underdeveloped, so many people have to create this kind of role for themselves.
Near-real-time work is about transparency as well as speed
Most Financial Times models run in batch, and Fitzpatrick does not argue that every company should force work into a two-hour window. The problem is that low-latency filtering and model execution are difficult to see in the current architecture. For example, the company may need to avoid sending a reader a newsletter they already receive, yet the relevant data assets and filtering layers are not transparent to the person maintaining the model. Another example is personalization on the payment barrier, where consented user data needs to become model input quickly enough to select an appropriate page. Fitzpatrick describes the existing setup as a reliable but complicated system that needs a clearer wrapper and better visibility.
Third-party tools are useful when they cover a narrow capability outside the business's core work
Fitzpatrick chooses vendors according to return on investment and the business's immediate needs. The Financial Times used a fraud vendor because payments required that capability before its data science function existed. It also uses third parties for some natural language processing work because creating a large specialist team would not make sense for the company's needs. She sees these tools as additions to internal expertise. The vendor does not know all the details of a company's payment system, content, or operating model. Internal practitioners still provide the domain knowledge that connects a specialized service to the rest of the business. Data companies have a stronger reason to build those capabilities in-house because data is part of their product.
Infrastructure should improve behind the same workflow for data scientists
Fitzpatrick compares her team's current setup to a Ford Fiesta that reliably gets people from one place to another. She wants to put a better engine and upgraded components underneath it without forcing the team to learn a completely different way of working. Data scientists should still push commits to GitHub, get them approved and merged, and use the same deployment path. Engineering can change the continuous integration, continuous delivery, serving, hosting, and scaling layers behind that interface. She argues that this work needs engineers who understand data science and can move the team forward gradually. A sudden switch to Kubernetes or another platform can create a large migration and training burden for a small team.
R remains useful in production, although Python has stronger momentum
Fitzpatrick started building models in R after working with SAS and C++. In 2014, she found R's community and packages more mature for her needs than scikit-learn, while her teams also used Python. She has put R models into production since 2014, including systems that needed low-latency responses. Python became more important when parallel processing and multi-threading were needed for real-time decisions. She thinks R benefits from RStudio's investment and from the tidy models ecosystem, but worries that open-source tooling often treats R as an afterthought. She expects universities and education providers to influence whether R continues to hold its place beside Python and possibly Julia. She also thinks data scientists should be able to move between languages as the underlying concepts remain stable.
"There's got to be something in between the team that's responsible for building and servicing the capabilities, the model, and the team that's actually consuming it."Leanne Fitzpatrick25:35
Who should watch
You are deciding whether to add data scientists, engineers, embedded analysts, or outside vendors to a small data organization.
Your team has working models but batch architecture, unclear ownership, or a difficult path from development to production.
You need to explain data roles and model value to business leaders who understand conventional software roles more easily.
