How to restructure data teams and AI priorities in 2027

AI was supposed to make your job easier, but somehow there’s more to do than ever. 

The chat assistant in your BI tool has barely gone live, and leadership already wants AI agents in production by next quarter. But the ad-hoc queue is just as backlogged as before, analysts are still checking AI answers manually, and every new tool adds another set of definitions and integrations to maintain.

So yes, the AI tools are better, budgets are bigger, and the roadmap is more ambitious, but the team delivering the insights are stretched too thin. 

To understand how leading teams are breaking this cycle, we surveyed 201 data, AI, and analytics leaders and sat down with teams at Cisco, Arm, and Property Finder. Here’s what held them back and how they restructured their teams and priorities to pull ahead:

What’s slowing down data teams: Old habits and undefined roles

Getting started with AI analytics is the easy part. Nearly 93% of data leaders are already exploring AI for analytics, and more than half have it in production. Yet only 7% have scaled it across the business. Here’s why:

Data roles are changing faster than anyone can keep up

86% of leaders say AI has changed their team’s role and core responsibilities. Data analysts who used to spend their days building reports are now also expected to check AI outputs, manage a growing stack of tools, and keep AI context up to date.

At the same time, fears about AI taking over data jobs still loom. For most people, though, the most immediate threat is a skills gap. 52% of leaders say their data teams lack the necessary AI skills, and 42% of the surveyed companies do not have the capacity to roll out AI. They’re being handed new work without the essential training or time to grow into it. 

Take context engineering, for example. Managing context is a new kind of work, and very few data folks are explicitly trained for it. It also has no clear owner, so the work gets split across governance specialists, architects, and engineers, usually on top of their day jobs. As a result, they are stuck in fight-or-flight mode, working overtime to keep AI answers accurate while the request queue keeps growing.

Old fixes are failing to solve AI accuracy

Whenever an AI answer comes back wrong, most leaders reach for the fixes they already know. Clean data tops the list, with 60% saying it can help improve AI accuracy.

That’s an understandable instinct, but accuracy is more complicated than it looks. Clean data can only tell AI what the values are and how they connect. However, it fails to convey what a metric means at your company, why a definition changed last month, or which version applies to the question at hand. 

Teams that spot this gap usually turn to the next obvious fix, which is context. They write down their definitions, create a semantic layer, load them into their BI tool, and assume the answers will still be right six months later. The trouble is, a lot can change in six months. 

Schemas get restructured, definitions shift, and new exceptions pile up in Slack threads, so context goes stale without anyone noticing. Even Anthropic watched accuracy slide from 95% to 65% in a single month after treating context as a one-time build.

In the end, neither fix gets to the root of the accuracy problem. That’s why only 19% of leaders are fully confident in their AI-generated answers.

Context lives everywhere and nowhere

74% of data leaders say they don’t have a complete, shared definition of enterprise context. Even those that have defined it rarely keep it in one place. Some of it sits in a semantic layer or a data catalog. Some lives inside whichever BI tool or AI assistant your analyst happened to be using when they wrote down an exception. The result is multiple copies of the same metric, each meaning something slightly different.

Picture this: your CFO asks your conversational BI tool how many active customers you had last quarter. The semantic layer defines an active customer as any account with a paid subscription, while the finance team’s dashboard counts only accounts invoiced during the quarter. 

The AI finds both definitions with no rule saying which one applies, so it picks whichever looks most plausible. Your CFO gets one number, then sees a different one in finance’s quarterly report, and nobody can explain the gap.

Multiply that across hundreds of metrics, and nobody trusts any answer. We ran into the same problem ourselves while building WisdomAI, and we’ve shared what we learned from it.

Tool sprawl is slowing analytics down

The modern data stack is collapsing. For years, companies built their analytics architecture around the dashboard. Each stage of the process, from moving and storing data to defining metrics and visualizing results, required its own tool. Over time, this fragmentation drove up costs and made the stack increasingly difficult to manage.

Now, with agentic analytics, many companies are repeating the same mistake. Rather than rethinking their architecture, they’re layering AI-native tools on top of their existing BI platforms. Meanwhile, business users are turning to general-purpose LLMs like Claude and ChatGPT to analyze business data on the side.

As a result, many data professionals spend more energy keeping tools in sync than helping the business get value from them.

Governance is falling behind AI adoption

AI is spreading across the business faster than the rules around it. New tools go live, business users start asking questions, and most governance decisions still go unresolved: which data can the AI access, which definitions should it trust, and who is responsible for validating the answer?

That leaves data teams in a difficult position. They’re still accountable for every number, but often have limited visibility into how the AI produced them. Which is why 48% of leaders say they want more transparency into how AI analysis is done, while 44% see human-in-the-loop review as a way to improve accuracy. However, adding human review just brings back the bottleneck AI was supposed to fix. 

How are data leaders restructuring their teams?

The leaders making real progress don’t start with technology. They start with people, processes, and the foundation underneath AI.

Building AI skills across the data team

Closing the skills gap doesn’t mean you have to start hiring from scratch. Most leaders are investing in the people they already have, with 57% planning to upskill or reskill their teams over the next two years.

At Arm, the nearly $5 billion chip and software company, the first step was changing how people on his team spend their time. Tom Smith, who leads procurement governance, policy, and reporting, says his group used to focus on building reports and collecting requirements. “That’s the easy part now,” he says. Today, the higher-value work is helping AI understand the business by reviewing how people use it, examining the queries it generates, and finding where context breaks down.

The lesson is simple: don’t just throw AI at people. Find out where they’re struggling, and give them the skills to handle it.

Laying a stronger data and context foundation

Clean data still matters, but it shouldn’t be treated as the finish line.

Arm learned this firsthand. When it launched its AI analytics program in procurement, Smith didn’t wait for the data to be perfect. “If I was waiting for perfect procurement data, we wouldn’t have ever started,” he says. Instead, his group focused on giving AI enough business context to interpret the data correctly and let data quality improve alongside the rollout.

Cisco’s finance team went a step further and rebuilt its foundation entirely. Rather than adding new dashboards to reconcile old ones, Raj Kamra, Head of AI Strategy at Cisco Finance, says they built a single context layer where semantics, governance, and institutional knowledge work together to help AI reason correctly. As a result, Cisco improved time to insight by more than 90%.

So before you fund another data migration project, ask whether it will make your AI more accurate. In most cases, investing in how you manage context will get you further, faster.

How data leaders are resetting their AI analytics priorities

It’s tempting to say yes to every AI initiative that lands on your desk, from new assistants and agents to yet another tool to evaluate. But without a clear way to measure ROI, each new project only drains people who are already stretched thin. The leaders we spoke to take the opposite approach, narrowing their focus to 3 priorities:

Boosting trust in AI answers

Speed was the first win of AI analytics, with 56% of leaders naming it as the biggest benefit they’ve seen so far. However, faster answers don’t matter much if nobody trusts them enough to act on them. Here are 3 of the most common ways to close that gap:

• Tracking domain health: Company-wide accuracy scores tell you how AI is doing on average, not where it’s failing. Tracking Domain Health for each part of the business, such as finance, sales, or procurement, shows where context is strong, where it’s thin, and where it needs work.

• Running evals before and after launch: Before business users rely on AI answers, test your business context against real business questions, and keep testing. Evals catch missing definitions, ambiguous terms, and conflicting logic early, while they’re still easy to fix. 

• Using an analytics harness: An analytics harness sits between the AI and your data. It controls which definitions and sources the AI can use and how it reaches an answer, so the same question follows the same path every time. And because the harness draws from the same context layer and is applied in every AI tool, adding a new AI assistant or agent doesn’t mean maintaining another copy of your business logic.

Unifying governance and explainability

Governance can’t be a checkpoint that AI answers pass through at the end. It needs to be built into every stage, from how context is created and maintained to how answers reach the people who rely on them. Two practices make that possible:

• Treating context like code: Each domain should have an owner, an update schedule, a review process, and version history, so context stays current instead of being uploaded once and forgotten. For instance, Property Finder has a dedicated context owner for each domain, including commercial, product, and marketing. They help keep definitions consistent and stop one’s logic from quietly overriding another’s.

• Making answers explainable: Every AI answer should come with a clear trail showing which data it used, which definitions it applied, and how it reached the result. This transparency gives data teams the observability to review generated queries, spot patterns, and trace any number back to its source. Each mistake they catch becomes better context, so accuracy compounds with every new question. 

Redesigning people and processes

Actual results come when every rollout supports people in how they learn, work, and measure success. Here’s how to reshape the way your organization adopts AI:

• Plan for change management: 34% of leaders say the lack of a change management strategy is a major challenge. That’s why every new AI tool should launch with clear guidance on who should use it, what it’s for, and where to go when an answer looks wrong. 

• Train the business users too: Even with the right guidance in place, business users need the skills to use AI well, especially for more complex analysis. 31% of leaders say people lack the AI literacy to judge whether an answer is right. Once business users learn to ask better questions and sanity-check results, fewer verification requests land on the data team.

• Measure the right outcomes: Reports shipped and dashboards built show how busy your data folks are, not how much value AI delivers. Realign your metrics to track whether the business actually gets answers it can act on, such as time to a trusted answer, accuracy by domain,and the share of questions business users answer on their own. That way, people are rewarded for impact rather than volume.

Where to go from here

The teams pulling ahead aren’t seduced by the latest wave of tools. Instead, they’re investing in what makes AI answers trustworthy enough to act on, from skilled people to governed context to metrics that capture real impact.

The good news is that none of this has to happen all at once. You can start with one domain, one context owner, and a few important metrics. From there, the work comes down to deciding where AI fits in your stack and how it will earn trust. Get that right, and AI starts making the job easier, as promised.

Want to see how your team compares? Download Meet the Modern Data Team for the full findings from our survey of 201 data, AI, and analytics leaders. Inside, you’ll find benchmarks and real-world stories on AI adoption, accuracy, context management, and hiring.