Governed AI Analytics
We help data-heavy companies turn fragmented BI, tribal knowledge, and stalled AI experiments into governed, agentic analytics systems — reliable answers, automated analyst workflows, and control over definitions, trust, and governance.
Fixed-price diagnostic · defined-scope builds · a practice small enough to be accountable
The principal: open to senior roles and select engagements
Diagram: scattered analytics artifacts — a board deck, KPI tracker, old dashboard, tribal knowledge, conflicting definitions, manual requests — converge under a governed layer where each becomes owned, certified, and documented; above it, planner, governor, analyst, and evaluator agents answer questions with citations, and unsafe questions are stopped by design.
fragmented estate→governed layer→agentic system
0
years
across data, analytics, and the decisions they serve
0
from zero
BI functions built inside operating companies
0+
governed models
carried through two platform migrations, zero historical gap
0/22
published
the practice copilot's own evaluation result
No. 01 — The problem
Companies bolt AI onto messy data, undefined metrics, undocumented dashboards, and tribal knowledge. The model isn't the problem — the foundation is.
LLMs don't fix messy metrics. They amplify them.
Your analysts are becoming human APIs — re-answering the same questions, re-validating the same definitions.
Your executives want answers, but your definitions are scattered across dashboards, docs, and Slack.
An AI pilot on an ungoverned semantic layer doesn't fail loudly. It fails credibly — with confident, wrong answers.
No. 02 — The fix
Between your data foundations and any AI system sits a layer of governed knowledge: metric definitions with owners, a business glossary, a data dictionary, lineage, trust scoring, and evaluation. Build it, and AI answers become reliable. Skip it, and every copilot inherits your ambiguity — fluently.
No. 03 — Working proof
Three live surfaces on this site, built the way we build for clients — two staged on labeled fictional data, one grounded in the practice’s real knowledge base. Same discipline, three angles.
Planner, governor, analyst, evaluator over a governed registry — running deterministically in your browser. Ask a question, watch the orchestration, switch the governed layer off, and watch the same question fail credibly.
Operate the modelGrounded answers from the practice’s published knowledge base, refusals as designed behavior, a human handoff at the right moment — with its governance ledger and evaluation report published beside it. It’s also one click away on every page, lower right.
Ask the Practice CopilotIngest to bronze to silver to gold to the semantic contract — then a bad definition change arrives, and the trust gate holds it before either consumer serves a wrong answer. The model above shows trust being used; this walk shows it being built.
Ian is the author and maintainer of Iris, an open-source MCP-native evaluation server for AI agents — grounding, quality, and safety scoring for agent behavior, listed in the official MCP Registry. The same discipline this practice sells, maintained in public where anyone can read it.
The reference behind the practice: the working vocabulary, the essays from the field, and the whole stack, mapped.
No. 04 — Field patterns
Twenty years operating inside these environments — not consulting on them from the outside. Sectors named, employers and clients not. References available in a working session.
Tax software — $350M+ operator
Leadership wanted AI answering business questions. The definitions lived in dashboards, documents, and analysts' heads — and AI on ungoverned definitions doesn't fail loudly. It fails credibly.
A versioned knowledge layer — canonical metric definitions with hard guards and automated review gates, readable by people and consumed by AI agents — behind a grounded Q&A surface that cites its sources and refuses what it can't answer safely.
Grounded, cited answers in production — refusal as a designed behavior
Tax software — $350M+ operator
Finance and CS leadership were deciding from stale exports out of three disconnected systems. The gap wasn't technical — it was architectural: no defined grain, no conformed dimensions, no single source of truth.
A governed semantic layer and executive scorecard became the peak-season operating rhythm — rolling variance vs. plan, six time-horizon comparisons, one set of definitions.
16+ executive dashboards through a single peak season — with the diagnostic catches to show for it
Private-equity real estate
A fast-scaling operator making acquisition and renovation decisions on gut feel — not because data was unavailable, but because nobody had ever connected it.
A BI function from zero: semantic model, company-wide data catalog with lineage, ML-assisted cost prediction, and a KPI framework leadership actually used.
First BI function in company history — Excel to governed BI, end to end
Enterprise contact-center migrations
Platform migrations threatened years of comparative history. The real risk wasn't the cutover — it was that nobody had ever documented the legacy schema.
Complete schema documentation, field-by-field mapping, and a continuity layer that projects each new platform into the definitions the business already trusts — so reporting never re-learns its own numbers.
Zero historical gap — two migrations, three platforms, 100+ governed models
Tax software — customer operations
Hundreds of thousands of customer conversations held the answers — why customers struggled, where calls escalated. Far more than any team could read, and too sensitive to ship to an outside AI service.
A warehouse-native LLM pipeline — issue taxonomy, sentiment, escalation flags — run entirely inside the data platform, with a self-serve app so operations could ask questions in plain language.
Hundreds of thousands of transcripts classified — no data left the warehouse
Cross-organization pattern
Documentation, SQL scaffolding, formatting, and stakeholder prep are mechanical — and they consume the analyst hours that insight should get.
AI-native workflow design: automated documentation, SQL tuning assistance, generated briefing decks — with the quality bar enforced, not hoped for.
3–5× faster delivery on documentation-heavy work
No. 05 — Engagements
Start with a two-week diagnostic. Every step earns the next — no seven-figure platform bets, no AI theater.
$9,500
Know exactly why your analytics environment is — or isn't — ready for AI.
from $40K
Scattered business definitions become a trusted knowledge base both humans and AI can use.
from $60K
A grounded AI assistant that answers trusted business questions against governed knowledge — with evaluation to prove it.
A senior AI and data-systems architect on your leadership's side — a part-time seat, a few days a week, alongside your team. Two to three seats at any time.
The principal
Two BI functions built from zero — private-equity real estate, higher education — and a governed AI knowledge platform architected inside a $350M tax-software operator. Since 2023: AI-native analytics with the governance to make it trustworthy. This is a deliberately small practice: senior work, done by the person you talked to.
Ian currently holds a senior analytics seat inside a $350M tax-software operator, where he serves as the data organization's AI ambassador. The practice runs a deliberately small book of work alongside that seat, and he is open to senior full-time roles and select engagements where the fit is right.
Two paths: score yourself in five minutes, or bring the question to a working session.