The life and times of data teams in the AI era

As if on schedule, the data world seems to have an identity crisis every several years or so.

A new technology emerges, we declare our old ways of working obsolete, and we come up with a new name or job title to help us chart our way to a beautiful new future.

The first time you experience it, it feels energizing and empowering, but once you’ve been through a few more, it’s easy to roll your eyes. All the same, there’s only so much you can gain from being cynical about it. New technologies can seem incremental or transformative depending on when you decide to adopt them, but sooner or later, the profession converges on what’s genuinely useful.

And the leverage they give us ends up fundamentally changing the way we work. AI in data is reaching that state—where the technology has evolved enough that it’s worth taking a step back to talk about how we operate—and build our teams.

The fundamentals of data teams

Also, hello! If you don’t know me already, I’m Katie, the head of data here at Hex, which is not just a ceremonial title. I run a real data team here at Hex, where we deal with data quality issues, chase down metrics when they go haywire, and yes, think about vendor costs on occasion.

I have also built and lead data teams at small startups all the way to publicly traded companies, which has given me a lot of perspective on what works in which environments.If I had to boil down the purpose of most data teams as simply as possible, I’d say that our number one job has been getting data into people’s hands.

On a very small or early data team, one person might own all of this work, but regardless of how many people are doing it, I tend to group the work into two distinct buckets: platform and distribution.

Platform work is all about making sure you have the right data supply to meet your organization’s data demands. It involves ingesting and integrating data into one place, modeling it for consistency and quality, and making it available in all your relevant consumption surfaces (like Salesforce, a BI tool, or even directly from the warehouse itself).

Distribution work, on the other hand, is about the last mile, and it is similarly multifaceted. Some of it is high effort and deeply stakeholder entwined, such as running analyses for someone or facilitating a regular business review, and some of it is more disintermediated, such as building canonical dashboards or programmatically generated reports.

If this feels very high level, it is because I’m trying to be a little vague on purpose. As we have noted before, job titles are somewhat imprecise when it comes to describing someone’s work, and anchoring discussion in the tech-du-jour can feel shortsighted. The important thing is, whether you're writing a MapReduce job or a SQL-based transformation, what's most important is that you’re still using data to model concepts important to your business.It might also feel artificial to divide the world in two like this, especially if the argument is that both of these things are a part of the mandate of data teams, but the reason for the divide has more to do with mindset than skillset.

There’s always been a tremendous amount of leverage in being able to think both in platform and distribution, but the incentives of each mode are different enough that you need to be conscious of which one you’re working in.Platform thinking is more about abstraction, composability, and the long term.

Distribution thinking is more about direct applications, relevance to a specific context, and timeliness. Neither is inherently better or more important because both are required to successfully make data a part of an organization’s heartbeat. Anyone building a data team needs to know how much they are responsible for each buckets, and when they are contributing to one or the other.

What AI changes

The good news is, data teams still largely have the same mandate: getting data into people’s hands. The biggest shift is that now most data access is mediated by agents. This is not the same as being for agents, but it does fundamentally change a lot of things for both modalities of data work. We’ll start with distribution, and then move backwards to platform.

Perhaps the most significant change for data teams on the distribution front is that where there was once a substantial technical barrier, whether that is through direct text-to-SQL or something more abstract like an MCP or a purpose-built agent, there no longer is.

Because of this, self-service (with all its promise) is now possible, and whether you like it or not, your stakeholders are going to be using AI to pull, analyze and interpret your organization’s data.

The liberating part of this is that a lot of distribution work in the past has been very toilsome. Sure, there can be moments where writing a query or assembling another dashboard for someone else is a pleasantly rote task, but it gets demoralizing quickly when it crowds out everything else.

Now an agent can work directly with that stakeholder who needs to go through multiple iterations of their request to get what they really want. On top of this, analytics agents can now unlock truly personalized distribution.

Anyone who has had to maintain a dashboard on behalf of their organization has had the experience of trying to build something well-scoped and specific, only to have a number of small requests that add up to the point that your laptop is set on fire every time you try to load it.

There's a ton of value in people having their own personalized view of the business, but until the dawn of AI, it was completely infeasible to actually create them. The same is true about highly customized docs or decks of data.

As I type this, though, I can feel the anxiety rising. The mediation of data distribution frees us in many ways, but it also means that data teams will be cut out of the loop. This has always happened to some degree , but now it’s at a whole new scale that feels impossible to supervise.

An infinite supply of vibed coded dashboards and analysis also means that there is boundless opportunity for two numbers to not match and for that work to be eventually thrown over the wall for the data team to figure out.

When it comes to the platform side of the house, the world has also been changed by near-universal data accessibility. Suddenly, we're starting to see the problems of much larger companies appear much sooner than you’d expect.

Things like ineffective metrics governance, exploding vendor costs, compliance headaches, all come to the forefront when data is available to everyone.

So, what do you need to do? Now that so much data usage is agent-mediated, observability and access management need to become first class concerns essentially from the start. When anyone can pull data and throw it into an LLM, they can pipe it into all sorts of places you might not want it to go.

You need to determine what parts of your data are sensitive and set up RBAC to restrict access to the right folks, and get systems in place to help detect when something is awry.

Widespread usage of your data will also mean that more people will find the edges of defined concepts faster. If you’re seeing a lot of adoption of AI for data work, you’re also likely seeing it accelerate the pace in all other parts of your company, which can make it even harder to keep up with all the modeling needs created by an influx of new and constantly mutating entities and business processes.

There’s always something new to give a shape.

What a data leader needs to prioritize

Data work was already not easy, and now we’ve got to do more of it and at a larger scale? It’s really easy to feel overwhelmed by the change brought on by AI, but the good news is, there’s still a lot that’s the same as before.

First, data quality still matters. Agents are capable of figuring a lot of things out through brute force querying, but much like people, they work more efficiently and effectively when they’re handling well-designed and clearly documented datasets. At Hex, we see that the data teams that get to value quickest when deploying AI analytics are the teams that have taken the time to invest in data quality, whether its via the data models in their warehouse or something more formal like a semantic model.

It really stands out that data quality is emphasized even by the big labs when talking about how to use frontier models for analytics. Even in this moment of change, there’s still a tremendous need make sure data is clean, usable and reliable.

Second, incentives still matter. Companies have dreamed for ages of democratized data access, where everyone can get access to the same information and use it scaffold their thinking into something more robust, but that doesn’t change the fact that everyone is bringing their own perspective and biases to data interpretation.

This doesn't mean that all of the non-data-professionals out there are maliciously misinterpreting data, but rather that they don't always have the ability to bring a neutral perspective to their data usage. And they might not have the experience to recognize common pitfalls or identify clearer ways to view the same information. Data teams have a lot of practice at remaining objective, and that’s something that’s unlikely to go out of style any time soon.

Finally, knowing the business still matters. I probably haven’t painted a vision of a world where data teams have any less work to do, which means that you’re still going to have to make choices about which of your organization’s data problems get the precious attention of a human.

This requires data leaders to build relationships with their stakeholder counterparts, to pay attention to what on your roadmap is working and what isn’t, and to relentlessly solve the most important problems facing the business instead of being relegated to doing what someone else thinks “data” is.

We are definitely moving at a faster pace than we were in the pre-AI world. But we may have actually achieved that age old-goal of getting data into people’s hands, and uncovered a new mission in the process. Everyone is using data now, and the new goal of the Data team is to elevate how it’s done.

Advice for building a data team in the age of AI

If you’re leading a data team, you’re always building a data team, even if you’re not building one from the ground up. The organizations we sit in are changing around us, and we need to change along with them.

There has been a lot of talk in the industry broadly about how roles are blending and merging, and there is some merit to it.

Being a full-stack data professional (or at least being full-stack enough that you can reach left or right of where you typically sit) has always had some leverage and with the rise of widespread AI adoption everyone’s default skillset has greatly expanded.

Data roles have gotten highly specialized the past few years, and it has resulted in idle cycles while people wait for the right person to be available to take on the next leg of the work.

I strongly encourage anyone building a data team now to delay specialization where you can, and to focus on cross-training and upskilling folks already on your team to be able to contribute to more parts of the end-to-end data lifecycle.

It might feel intimidating at first, but it’s also highly empowering for folks to have new skills to unblock themselves. It also greatly reduces the bus factor on your team, reducing the need for individuals to engage in heroism to keep things afloat and increasing the likelihood that things like their well-earned PTO won’t be interrupted.

There is some nuance here, though! Technical skillsets are not the same as knowledge of when to apply them, so don't assume that all the members of your team are totally fungible across projects and tasks. (And don’t let your stakeholders assume this either!)

The two buckets of platform work and distribution work are as relevant as ever in a post-AI world, and specialist knowledge is still incredibly important. Don’t ask the analyst who knows nothing about security to be the sole owner of setting up RBAC in your data warehouse, and don’t ask the data engineer who doesn’t know inferential statistics to be the only point of contact for a product manager designing an A/B test.

As a leader, you still need to be mindful of which types of work your team members are happiest and most motivated by doing. Even if AI erodes the technical boundary between roles, make sure that the folks on your team still have a clear, primary focus. If a problem is important to solve, make sure it has a full-time owner and worry less about the specific technical skills that AI can smooth over gaps in.

Above all else, the most critical piece of advice I have for building a data team in 2026 is that your default should be to think about how you can say yes. This is the age of the builder, and whether you personally identify as one or not, it behooves you to act like one, or at a minimum, to figure out how to enable them at your organization.

On a scale of chaotic to lawful, Data teams definitely lean towards lawful, and that has historically earned us the reputation of being stubborn and slow-moving.

We worry about consistency, quality, and compliance, and we are right to because those things are important. But we need to figure out how to allow our colleagues to who are hungrier than ever for data to make good choices by default. This requires you to invest in understanding agent behavior, and to accept that not every agent interaction will be perfect.

Agents are pretty good at retrieving data (given the right context), but the vast majority of interactions with them are not high stakes so it’s okay if they don’t always make the same choices as you would. Figure out what matters, make sure you have the right checks in place to catch bad behavior, and learn the techniques for steering agents to follow the right procedures and policies you’d want a member of your team to follow.

Making data better for everyone

In the past few years, it feels like AI has launched us to the future, but fortunately, sometimes the future is more similar to the past than it might seem.

Even though more data work is done by agents and more people outside data teams are able to participate in it, there’s still a critical need for data teams’ signature combination of technical and business knowledge.

AI is the newest tool to add to our kit, and with the right investments, we can use it to transition to a new era. Instead of worry about getting data into everyone’s hands, we can focus on something more ambitious: making sure we’re using data better.