Scaling Analytics with AI

Turning Analysts
into AI Solution
Builders

Once agents can explore schemas, write competent SQL, and pull from a governed semantic layer, the “human as query engine” role shrinks. The higher-leverage job is designing, directing, and validating the systems that do the analysis.

📖 ~14 min readOperating-Model Deep-Dive
scroll
The Monday patternExec packs, cohort deep-dives, variance explanations, data-quality checks. The backlog only grows.

The request volume never slows down. Every Monday the analytics channel fills with the same pattern of asks: refreshed executive packs, cohort deep-dives, pipeline variance explanations, data-quality checks. The team is competent and fast, yet the backlog only grows. Ad-hoc work crowds out the higher-judgment analysis the business actually needs.

That pressure forces a decision most analytics leaders eventually face in 2026. Keep treating analysts as the people who write the queries and build the dashboards, or deliberately turn them into designers and validators of agentic workflows. The second path is harder in the short term and the only one that scales.

The analyst moves from doing the analysis to directing and governing the system that does the analysis. Teams that don't make the shift simply drown more elegantly.

The Role Is Being Rewritten

The old loop breaksReceive request → write SQL → deliver → next ticket. It collapses under its own weight.

The classic analyst loop (receive request, write SQL, deliver answer, move to the next ticket) is breaking under its own weight. Once agents can explore schemas through MCP, write competent SQL, pull from a semantic layer, and format a summary, the pure “human as query engine” role shrinks. What emerges instead is a higher-leverage version of the job:

Spot the recurring, high-volume workflows.

Design the agentic version: tools, guardrails, evaluation criteria, output contract.

Wire them to governed data through MCP servers and the semantic layer.

Define what “good” looks like and build the verification steps.

Monitor, improve, and retire agents over time.

Teams that make this shift handle significantly more demand with the same headcount. Teams that don't simply drown more elegantly. Routine reporting, many data-quality checks, and first-pass anomaly investigations are already being automated end-to-end in production, and the analysts who can design, direct, and validate those workflows are already more valuable than those who only write the queries.

The Practical Stack That Is Winning

The convergenceCursor + Git + LLMs + MCP. Agent outputs become PRs. Nothing reaches prod without a trail.

The combination most analytics teams are standardizing on is straightforward and powerful:

Cursor (or equivalent agent-native IDEs) as the primary workbench, where agents read the project, edit multiple files, and propose coherent changes.

Git as the source of truth, so agent outputs become pull requests, humans review, and nothing reaches production without a trail.

LLMs as the reasoning engine inside the agents.

MCP servers as the controlled interface to the warehouse, dbt project, semantic layer, internal docs, and APIs.

dbt's MCP server comes up constantly among analytics engineers because it lets an agent inspect models, understand upstream dependencies, and trigger builds while respecting project structure. Pair it with a warehouse MCP and a semantic-layer connection and an agent can move from question to governed insight without the analyst hand-crafting every step.

Non-negotiableVerification. Agent work is only as good as the human who defines the validation contract and checks it.
You are no longer writing every line yourself; you are learning to direct a set of capable clones. The clones remain only as good as the original.An analytics engineer on the new workflow
PREVIOUS MODEL Ticket arrives Write SQL Build dashboard Next ticket 2026 MODEL Identify recurring workflow Design agent (tools+evals) Connect via MCP + semantic Review agent PR Monitor & improve executing → directing the work
Figure 23.1: From Query Writer to Workflow Designer. The unit of work changes from a single answer to a reusable, governed agent the analyst designs and reviews.

How PMs Drive the Shift

Not a workshopThis needs product thinking and change management, not a training budget and good intentions.

This doesn't happen through a training budget and good intentions. It requires product thinking and deliberate change management. The effective moves:

  1. 1

    Freeze or heavily constrain new ad-hoc dashboard requests for a defined period so the team has oxygen.

  2. 2

    Choose one or two painful, high-volume, well-understood workflows and turn them into reliable agents first.

  3. 3

    Make agent design and PR review visible in standups and performance conversations.

  4. 4

    Build shared scaffolding early: project rules, evaluation checklists, approved MCP servers, SKILL files, so the next agent costs less.

  5. 5

    Pair stronger AI-native analysts with others on the first projects instead of running generic workshops.

Artificial scarcityPausing the old work for three weeks is what creates space for the new work.

The three-week freeze. One PM running an analytics platform at a mid-stage SaaS company took the blunt route: pause most new dashboard work for three weeks, redirect the team onto three recurring executive and growth workflows, and make successful agent hand-off the only way the backlog would shrink. The first two weeks were uncomfortable. By week six the volume of repetitive production work had dropped sharply and the team was spending more time on novel questions. His reflection: “We had to create artificial scarcity of the old work to make space for the new work.”

Workflows That Convert Well

Convert-cleanly testRepetitive, explicit “done,” and a wrong answer is manageable once verification is in place.

Not every analysis should become an agent. The ones that convert cleanly share three traits: they're repetitive, the definition of “done” can be made explicit, and the cost of a wrong answer is manageable when verification is in place. High-value examples teams are successfully agentifying:

Weekly executive metric packages with automated data-quality checks and variance commentary.

Cohort retention or activation investigations that stay inside the semantic layer.

Pipeline or forecast anomaly detection with a first-pass root-cause summary.

Data-quality monitoring that opens tickets or PRs when freshness or volume breaks.

Board or investor reporting packs assembled from governed sources.

Constrained exploration agents that can't invent metrics because they only see the semantic layer.

Start narrow. Earn trust. Expand only after the first agents are reliably reviewed and merged.

New Division of Labor

Explicit hand-offsAnalysts shouldn't debug MCP auth; engineers shouldn't guess which workflow matters.

Clear ownership prevents the usual friction:

ANALYSTS / BUILDERS

Own the design, prompting, evaluation criteria, and iteration of the agentic workflows. They stay closest to the business questions.

ANALYTICS ENGINEERS

Own the reliability of the underlying data products, dbt models, semantic layer, and MCP servers. They make the tools safe and consistent.

PMs

Own prioritization (which workflows deserve agents), success metrics, stakeholder communication, and the upskilling program itself.

In smaller teams the same person wears two hats. In larger teams the hand-off points must be explicit. Analysts shouldn't be debugging MCP authentication, and analytics engineers shouldn't be guessing which business workflow actually matters.

BUSINESSQUESTIONS PMprioritize ANALYSTSdesign + validate AGENTSexecute + open PRs HUMANreview ANALYTICS ENGINEERSsemantic layer · MCP servers · data products improvements feed back
Figure 23.2: 2026 Analytics Operating Model. Questions flow through the PM to analysts who design agents; engineers keep the governed tools safe; humans review and merge. It's redirection, not replacement.

Real Stories from Teams Making the Transition

Sofia's reframe“I stopped being the person who produced the pack and became the person who decided whether it was good enough.”

The executive-pack agent. A senior analyst named Sofia took the weekly metrics package that used to consume most of her Monday and designed an agent that pulled from the semantic layer, ran the standard data-quality checks, generated variance commentary, and opened a PR. After two iterations the PR required only light editing. The biggest change, she said, was psychological: she stopped being the person who produced the pack and became the person who decided whether the pack was good enough.

Ryan's rollback“Speed without a verification contract just creates confident mistakes at scale.”

Premature automation. A different team tried to agentify too many workflows at once without shared standards or verification checklists. Several agents produced plausible but incorrect numbers because they bypassed the semantic layer. The team had to roll back and restart with stricter scaffolding. Ryan, who led it, put the lesson bluntly: “Speed without a verification contract just creates confident mistakes at scale.”

Measuring Whether the Shift Is Real

Language & calendarWhen those change, the operating model has changed, and not before.

Easy metrics like number of agents created or volume of agent-generated SQL are easy to game. Better signals: the percentage of recurring reporting and investigation workflows that now run primarily through agents with human review; the reduction in average time from a standard business question to a governed answer; the drop in pure ad-hoc request volume; the proportion of agent-generated PRs that merge with only light editing; and the qualitative shift when analysts spontaneously say “we should agent-ify this” instead of “I'll just write the query.” When the team's language and calendar change, the operating model has changed.

Framework: Analytics Team AI Transition Checklist

Use this to keep the shift grounded.

Ground truthEvery agent output lands as a Git PR with human review. No exceptions.
  • Identified the top 3 to 5 highest-volume recurring workflows?

  • For each, a clear definition of “done” and a verification step?

  • Agents required to use the semantic layer and approved MCP servers?

  • Every agent output landing as a Git PR with human review?

  • Shared project rules, evaluation checklists, and scaffolding that lower the cost of the next agent?

  • Performance and recognition updated so designing and governing agents counts as core work?

  • Current ratio of repetitive vs novel work known, with a six-month target?

Common Mistakes

MistakeFix
Training on tools while measuring ticket volume.Change the incentive: reward agents designed and merged.
Analysts spin up agents with no shared standards or Git.Scaffolding and PR review before scale.
Automating low-volume or high-judgment work first.Start with repetitive, well-understood workflows.
Building agents on messy, ungoverned data.Fix the semantic layer first; agents inherit its quality.
Treating it as a one-time workshop.It's a sustained operating-model change.

Key Takeaways

Redirection, not replacementKeep humans in the judgment seat; free your strongest people for problems that need taste.

The analyst role is moving from executing analysis to designing and governing the systems that execute it.

Cursor + Git + LLMs + MCP is the practical stack most teams are converging on.

PMs must create the conditions (prioritization, scaffolding, incentives, protected time) or the request-driven model reasserts itself.

Start with a few high-volume, well-understood workflows and make them reliably agentic before expanding.

Success shows up as reduced repetitive load, faster governed answers, and analysts spending more time on judgment work.

Ask Your Analytics / Data Team

1. “What are the three highest-volume recurring workflows we still do manually every week, and what would it take to turn each into a reliable agent with human review?”

2. “What shared scaffolding (project rules, approved MCP servers, evaluation checklists) do we already have, and what's the biggest missing piece?”

3. “Of last month's analyst time, what percentage went to novel or high-judgment work versus repetitive production, and what target do we want in six months?”

The analytics teams that will matter won't be the
ones writing the most SQL. They'll be the ones that
turn their repetitive knowledge into reliable agents,
keep humans in the judgment seat, and free their
strongest people for the problems that still
require taste and context.
Cross-references: The MCP infrastructure chapter covers the controlled interface these agents read through. The Semantic Layer chapter defines the governed metrics that keep agents from inventing numbers. The Proactive Intelligence chapter covers the monitoring workflows that convert especially well.

AT A GLANCE

Core conceptAnalysts shift from executing analysis to designing & governing agents that do it
The stackCursor + Git + LLMs + MCP servers, with PR-based human review
FrameworksQuery-Writer→Designer, 2026 Operating Model, Transition Checklist
Key ruleStart narrow, verify everything, keep humans in the judgment seat
Next Chapter
Understanding Evals
How to measure whether an AI system actually does the job you built it to do.
Continue →