Trusted AI Analytics Is Possible Today. It Will Still Take Years.
Part 3 of 4: Where analytics goes next
Picture an analyst in 2034, sitting inside the underwriting team. The head of underwriting asks an agent why the loss ratio on a book moved two points. The agent answers in seconds and nobody argues about the number, because its definition was settled years ago and every tool reads the same one. The analyst spends her morning on what the agent cannot settle: why the number moved and whether it will keep moving.
This is Part 3 of a series. Part 1 covered 2026 to 2028, when analytics moves inside the tools people already use. Part 2 covered 2028 to 2032, when analysts supervise AI agents instead of building every analysis. Part 3 covers 2032 onward, which I call Horizon 3.
These are plausible outcomes, not guarantees. Because the claims here are speculative, this installment leans heavily on surveys and citations so you can weigh the evidence yourself.
The short version: The technology for trusted AI analytics works today, inside a single team. Doing it across a whole company, and eventually an industry, requires agreed definitions of what the metrics mean, and that agreement is what takes years. Governance programs, healthcare data standards, and the new semantic standards effort all show the same slow clock. Start the agreement work now, because it sets the pace.
Approximately 1,900 words
Published November 6, 2026
What 2032 Looks Like
By 2032, agents prepare the data and produce the first cut of most analyses. Analysts no longer sit in a central reporting group. They sit inside underwriting, claims, and finance, and they spend their time on judgment: why a number moved, what it means, and what to do. A small central team holds the definitions and guardrails that every analyst and agent reads. Analytics shows up inside the decision itself, in the tools people already use, and stops being a separate request that takes a week.
That picture rests on one thing. Everyone has to be reading the same definition of the metric.
It All Works Today. So Why 2032?
The objection is fair, and half of it is right. Gartner’s March 2026 predictions say universal semantic layers will be treated as critical infrastructure by 2030, in the same tier as data platforms and cybersecurity. A semantic layer is the agreed business meaning of your data, written down where software can read it. Power BI, LookML, Snowflake Semantic Views, and Databricks Metric Views each let you define a metric once inside that platform and serve it to its own dashboards and AI agents.
The accuracy gains are real. A small April 2026 preprint, not peer reviewed and limited to one retail dataset, had three frontier models answer 100 business questions. They scored 45.5% to 50.5% from the database structure alone and 67.7% to 68.7% once a short document defined the measures. Canva reported more. According to a Hakkoda recap of Snowflake Summit in June 2026, its error rates ran from 10% to 40% without a semantic layer, and accuracy exceeded 90% on low and medium complexity questions with one. The layer settles which number is right, but it cannot tell an underwriting head why the number moved.
So a team that writes its definitions down can get trustworthy answers now. Ten metrics inside one firm can be settled in months. Agreement across every function, every business line, and eventually every company in an industry is the part no tool supplies, and it runs on its own clock.
The Agreement Is the Long Pole
In project planning, the long pole is the task that sets the finish date. Here it is agreement. Take “loss ratio,” a metric every insurer reports. Four questions decide which number you get:
- Do losses include a provision for claims incurred but not yet reported, or only reported claims?
- Do claim-handling expenses sit in the numerator?
- Is the denominator earned premium or written premium?
- Is the figure gross or net of reinsurance?
Each answer is defensible, and four yes-or-no choices produce up to sixteen versions of one metric. Actuarial, finance, underwriting, and reinsurance each have a reason to prefer a different one. A May 2026 TDWI piece notes that a semantic layer can hold several related definitions but cannot resolve genuine disagreements about them.
The consultancy Tryolabs, writing in March 2026 about failed semantic layer projects, says that across dozens of data platform engagements the bottleneck is almost never the technology. It is the organizational alignment needed to make sense of the data.
Someone has to convene the argument. In an insurer that is usually the chief data officer, or a finance and underwriting leader who co-sponsor it and have the standing to make a ruling stick.
Consensus Is Slow Everywhere
Every large effort that needs independent groups to agree has run on the same slow clock, even when the technology and the incentive were in place.
Governance programs. A Gartner survey of 223 data and analytics leaders in March 2026 found that cultural resistance outweighs funding constraints as the main reason governance initiatives fail, 60% to 40%.
Healthcare data standards. A mandated standard is still not an adopted one. Regulators require FHIR, a standard for exchanging health data, and the 2025 CAQH Index found that 63% of medical plans are still building FHIR-based electronic prior authorization ahead of a January 2027 federal deadline.
Semantic standards themselves. The Open Semantic Interchange, an effort to let metric definitions move between platforms, was launched in September 2025 by Snowflake, Salesforce, dbt Labs, BlackRock, and RelationalAI. Its first specification followed in January 2026, which Snowflake called a starting point, not an endpoint. A financial services working group held its first formal meeting in June 2026, with practitioners from banking, insurance, and asset management.
Inside companies the pattern repeats. Strategy, a vendor in this market, surveyed 100 senior data and technology leaders at companies with $2 billion or more in revenue and found that 99% struggle to define consistent business metrics across tools and departments.
The multi-year estimate is mine, and it comes from experience more than data. I have spent 16+ years in a large global professional services firm, and far too many hours of that sitting in meetings that should have been easy to resolve and stayed unresolved for years. Sound familiar?
A new tool can be configured in a quarter. A finance team and an underwriting team deciding whether a ratio is gross or net can take much longer, because the conversation tends to surface other disagreements along the way. Across a large company I expect years, and across an industry longer.
2032 is six years away.
What Skipping It Costs
A September 2026 VentureBeat survey asked enterprises running AI agents on company data whether any confident but wrong answer in the past six months traced to missing or inconsistent business context, such as a wrong metric definition. 69% had at least one, and 41% had more than one. Among those who reported a failure, 19% said a decision had to be reversed and 12% said an incorrect figure reached a financial, regulatory, or board report. The sample is 113 respondents from VentureBeat’s readers and panel, so read it as direction.
The Skeptic’s Case
Several sources above are vendors, and the surveys are self-reported. Models will also get better at handling ambiguity, which could shrink the accuracy gap.
But the preprint’s gain came from supplying definitions, not from changing models. A stronger model can guess better what “revenue” means at your company, and it still cannot know a choice your company never made. A confident wrong guess is also harder to catch than an obvious error.
The Layer Beneath: Data Contracts
A data contract is an agreement between the team that produces a dataset and the teams that consume it, covering structure, meaning, and quality, and checked automatically. It keeps a definition true after the source system changes. Gartner predicts that 50% of organizations will use AI agents by 2030 to convert governance rules into data contracts software can verify.
However, a person still has to write the policy first. The small central team in the 2032 picture is where that policy lives, and it needs standing with finance, underwriting, and claims, which takes time to earn.
What This Means for the Future of Analytics
Without agreed definitions, an agent prepares data quickly and wrongly, which is the confident, fluent, incorrect answer the respondents described, produced at higher volume. The analyst’s job in the 2032 picture is to own the definitions and answer for what the numbers mean.
That shift sets a skills agenda for the analysts working today. Four skills matter most over the next two quarters, because the agreement work depends on them:
- Business acumen. You cannot settle what a loss ratio means without understanding why actuarial and finance each want a different version.
- Critical thinking and AI validation. Someone has to check an agent’s answer against the agreed definition and notice when it drifts.
- Data storytelling. Turning data into a story that explains what happened, why, and what to do next, so a decision-maker acts on it. A story built on a number two teams dispute ends in an argument about the number.
- Prompt engineering. Agents read the definitions you write, and unclear instructions produce unclear answers.
Two more take longer to build and pay off under any horizon. Stakeholder engagement and cross-functional collaboration let an analyst convene finance, underwriting, and claims in one room. Domain expertise gets that analyst believed when they do.
Three Moves for the Next Two Quarters
Start the argument on ten metrics. Choose the ten an agent is most likely to be asked about. Put the right people in one room per metric, surface the edge cases like the four loss ratio questions, and name one owner. Expect several sessions, because people often turn out to be using the same word for different things.
Write the result where agents and humans read the same version. The preprint got its gain from a short document of definitions. A versioned, shared document is a legitimate first step and can migrate into a semantic layer later. The agreement travels with you even if you change tools.
Add a contract to the sources behind them. For the tables feeding those ten metrics, name the producing team, state what a valid record looks like, and check it automatically. Start with the two or three that fail most often.
If you are an analyst rather than a leader, you can start Monday without a mandate. Pick one metric you report today, find every place it is defined, and write the differences on one page. Bring the page to whoever owns the number.
How Ready Are You for 2032?
Answer yes or no:
- Definitions: Can you name the metrics an agent is most likely to be asked about, and say where each one is defined today?
- Ownership: Is there a named owner or small team for definitions, with standing across finance, underwriting, and operations?
- Skills: Can your analysts explain in plain language why two numbers for the same metric differ?
- Data: Do the datasets behind your most important metrics have a contract that names an owner and checks quality automatically?
Two or fewer yeses means the agreement work comes first, and the three moves above are your starting point.
Next Up: The Plan
Part 4 closes the series with what to do about your own skilling and your team’s. In my view, the skills worth building first are the ones that give the most competitive edge, so I sort them by edge and by how long they take to build, to show where to start now and what to build over time.