Analyst Becomes a Conductor.
Part 2 of 4: Where analytics goes next
Go back to the stakeholder from Part 1. She asked an agent, software that carries out a task on its own, for forty scenarios on a two-point swing in interest rates. She read the three that looked odd and called the analyst to ask whether they could be trusted.
Part 1 stopped at that call. The analyst is not rebuilding the model. She is deciding whether the odd results are a finding or a fault, and she answers for it if the board acts on the wrong one. That is a conductor’s job, and most analytics teams are not set up to hire or train for it.
The short version: Before agents run whole workflows, name one person accountable for what they produce, define what “vetted” means, and make sure analysts can see what each agent did. All three cost less now than after the first wrong answer reaches a decision.
Quick recap: In February 2026 I presented four themes across three horizons. The themes are Landscape (what is happening around us), Design (how we structure to respond), Skills (who we need to become), and Technology (what we build on). Horizon 1 runs 2026 to 2028, Horizon 2 runs 2028 to 2032, and Horizon 3 begins in 2032 and is more speculative. These are plausible outcomes, not guarantees, and the timing may move forward.
Approximately 1,850 words · 9 minute read
Published October 29, 2026
The Foundation Horizon 2 Inherits
Part 1 scored Horizon 1 with two months of 2026 left. Agents are spreading at large organizations. Governance lagged adoption plans. Question framing, the skill Horizon 1 expects to be scarcest, was still unmeasured. Data readiness was the weak point, with 63% of organizations lacking, or unsure they have, the right data management practices for AI. Horizon 2 sits on top of that foundation and hands the work itself to agents, which changes who asks for an analysis, who checks it, and who answers for it.
Horizon 2, 2028 to 2032: Agents Take Whole Workflows
The technology matures quickly, early pilots show value, and organizational models start to emerge.
Landscape: agentic analytics becomes mainstream. Multi-agent systems handle end-to-end workflows: data pull, exploratory analysis, modeling, and monitoring, with humans in the loop at key decisions. Stakeholders stop asking an analyst for a deck. They ask an agent for scenarios, and the analyst validates and guides.
Design: outcome squads and continuous insight loops. Cross-functional squads of product, operations, analytics, and engineering, organized around a result instead of a function, own a metric and run continuous test-and-learn cycles powered by agents. Instead of requesting reports, teams own a key performance indicator (KPI) and run dozens of agent-driven experiments every quarter.
Skills: from analyst to analytics conductor. Analysts design and supervise agent workflows, set guardrails (the limits agents work within), and focus on hypothesis design and validation. The job moves from doing the analysis to designing the analytical system and deciding when to trust it.
Technology: domain-tuned models and natural language interfaces. Models tuned to a specific industry make conversation the default way users reach data. Most of them never open a business intelligence (BI) tool. They ask a question and get a vetted, contextual answer.
When Nobody Opens the BI Tool
The Technology theme expects users to ask a question in plain language and get a vetted, contextual answer. Writing the query is getting solved fast. When the Spider 2.0 benchmark of 632 real enterprise text-to-SQL problems (software turning a plain-language question into a database query) launched in late 2024, the best model solved 21.3%. Its public leaderboard now lists systems above 90% on one subset, although vendors submit those scores themselves and the data-transformation subset tops out at 65.6%.
Meaning is the harder part. In a small April 2026 preprint, not yet peer reviewed and limited to one retail dataset, three frontier models answered 100 business questions. They got 45.5% to 50.5% right from the database structure alone (its tables and column names), and 67.7% to 68.7% once a short document defined the measures and conventions. The models did not differ from one another, so the gain came from supplying meaning, not from a stronger model.
A model cannot guess what your company means by “revenue” or “active customer.” When nobody opens a BI tool, the analyst stops being the gatekeeper to the data and starts owning those definitions. Part 3 returns to them.
By 2030, the Analyst Moves to the Front Room
In Horizon 2 the stakeholder begins, the agent does the first pass, and the analyst enters where judgment is needed. A conductor knows what each section should sound like and hears the one that is off. For an analyst, that means knowing the business well enough to see that a clean-looking result cannot be right.
A lot of analytics work today happens in the back room. A request goes in, the analyst builds, and a deck often comes out days later, after the decision conversation has moved on. Once an agent does the first pass, building is no longer the scarce part. The value moves to the front room, where the stakeholder is looking at forty scenarios and needs someone beside her who can say which three to believe and why. Part 1 said analytics would come to people inside the tools they already use. In Horizon 2 the analyst has to come along with it.
On the stakeholder’s call, the analyst does not open the agent’s code first. She asks what the three odd scenarios have in common, checks whether the inputs shifted between runs, and compares the results with how the business behaves when rates move.
Part 1’s chief curiosity officer frames the question that creates a new option. The conductor keeps that skill and adds judging the answer once the agent returns it. The best analyst in 2026 is often the fastest builder. The best one in 2030 catches the confident, wrong answer before it reaches the board.
Who Answers When the Agent Is Wrong?
Stanford’s 2026 AI Index counted a record 362 reported AI incidents in 2025, and ownership of AI governance is scattered: information security is the most common primary owner, at only 21%.
SecurityHQ and Datadog, both vendors in this market, surveyed 204 cybersecurity leaders. 76% said their organization has formally defined accountability for AI-driven decisions, and 45% had already experienced an incident where accountability was unclear. SecurityHQ’s read is that defining accountability on paper is the easy part, and that visibility into what an agent touched is what makes it answerable.
The conductor cannot answer for a section she cannot hear. If the analyst has no record of what data an agent pulled, what it assumed, and what it ran, the stakeholder’s call has no good answer.
Will the Conductor Role Last?
The conductor role depends on agents staying supervised, and the forecasts are mixed. Gartner expects at least 15% of day-to-day work decisions to be made autonomously by agentic AI by 2028 and also predicts that over 40% of agentic AI projects will be canceled by the end of 2027.
Stack Overflow’s April 2026 survey of 1,100 developers and working professionals, a group close to analysts but not the same, found that 59% use agents at work, up from 31% in its 2025 developer survey, and 63% rarely or never let agents run entirely on autopilot. Gartner’s 15% and this survey measure different things but point the same way: a growing minority of decisions made autonomously, with most work still supervised.
The role may not last. If agents become reliable enough, review could shrink, or the work could pass to engineers and business owners. That may happen for routine questions. For decisions with consequences, someone still has to catch the wrong answer, and whether that person is an analyst, an engineer, or a business owner, they need to know the business well enough to doubt a clean-looking result.
What to Practice Before 2028
The World Economic Forum reports that 63% of employers name skills gaps as the biggest barrier to transformation. Analytical thinking remains the top core skill, named essential by nearly 70% of employers, and leadership and social influence ranks third at 61%. A conductor needs analytical thinking, enough technical fluency to inspect an agent’s work, and the influence to be heard in a room.
Start with AI validation. Write down what you check before you trust a result: the inputs, the assumptions, and whether the answer fits how the business behaves. Ask to sit in on one decision that uses an agent’s output, and find out where its logs live. Then practice data storytelling: give the answer in two minutes to someone who does not do analytics, because the front room does not wait.
Four Decisions to Make Before Agents Run Whole Workflows
Decision rights. A squad that owns a KPI has no request queue, because the people who need the answer and the people who produce it are the same people. Governance has to move from approving individual analyses to approving the guardrails inside which squads run their own.
Accountability. When an agent produces a wrong scenario and a decision follows, someone answers for it. Name that person before the first incident, because the alternative is settling it during one.
Definition of “vetted.” Someone has to define what vetted means, including the record an agent must leave of the data it pulled and the assumptions it made, test against it, and own the result when it fails. Part 1’s self-check asked whether each of your ten most important metrics has one definition and one owner, and vetting is the same test applied to answers.
Pace. Dozens of experiments a quarter means a dozen decisions a month about which results to act on, and approval processes built for a handful of annual analyses will not keep up.
How Ready Are You for 2028?
Answer yes or no, as in Part 1:
- Landscape: Can you name a workflow where a stakeholder could ask an agent for scenarios today, and say who receives the results that look odd?
- Design: Does at least one KPI have a named owner with the authority to approve experiments inside agreed guardrails?
- Skills: Have you written down what an analyst checks before trusting an agent’s output, and trained anyone to defend that judgment to a stakeholder live?
- Technology: Can you see, after the fact, what an agent pulled, assumed, and ran?
Count the yeses. Two or fewer means accountability and visibility come first, and Part 4 covers the skills to build on top of them.
Next Up: 2032 and Beyond
Part 3 looks ahead to Horizon 3, which begins in 2032 and is still experimental today, and to why one of its requirements is already a Horizon 1 problem. Part 4 closes the series with what to do about your own skilling.