The Field Guide · Term
Human-API Analysts
When your best analysts spend their week serving lookups a governed layer should serve.
Human-API analysts is the pattern where an organization's analysts function as a request-response interface to its own data — re-answering the same questions, re-validating the same numbers, re-explaining the same definitions — because no governed layer exists for people or machines to read directly.
The most expensive query engine you own
Watch a data team's intake channel for a week. Most of what arrives is not analysis — it's retrieval. Which dashboard is right. What this metric includes. Whether this number can go in the deck. Each request routes to a person, because the person is where the answer lives. The analysts have become an API: high-latency, salary-priced, and rate-limited by burnout.
The pattern is a symptom, not a staffing problem. Answers recur because they're stored nowhere durable; definitions get re-litigated because none is canonical; validation is re-performed because trust was never made legible. Hiring more analysts adds capacity to the wrong layer — the queue drains faster and refills at the same rate.
It's also the honest test of AI readiness. The questions flooding your analysts are exactly what leadership hopes a copilot will absorb. But a copilot pointed at the same ungoverned sprawl the analysts navigate by memory doesn't inherit their judgment — only their workload, minus the caution. The way out is the same either way: a governed layer that serves the recurring answers, so both humans and machines stop paying the retrieval tax.
How it shows up
A senior analyst's calendar is diagnostic: hours blocked for “data questions,” the same five metrics explained to the same three departments in rotation. The organization's real metric registry is this person's memory — unversioned, unbacked-up, and two weeks of PTO from an outage.
The compounding cost is invisible in any single week: the analysis that never happened. Teams staffed for insight ship lookups instead, and the interesting questions — why the cohort behaves differently, what the margin pattern means — wait indefinitely behind the queue.
How to detect it
- 01
Sample a month of intake requests. Tag each as retrieval (the answer exists somewhere) or analysis (new work). Most environments are surprised by the ratio.
- 02
Ask what fraction of analyst hours goes to repetitive, mechanical requests. If nobody can answer, that is itself the answer.
- 03
Test the bus factor: could the team answer stakeholder questions if the most senior analyst left tomorrow?
- 04
Check whether recurring questions are captured anywhere durable — or re-answered from scratch each time they arrive.
Questions, answered plainly
Isn't answering questions literally the analysts' job?
Answering new questions is. Re-answering solved ones is a systems gap wearing a job description. The distinction to draw is retrieval versus analysis: retrieval belongs in a governed layer both people and AI can read; analysis is what you actually hired for.
Does a copilot fix this?
Only after the governed layer exists. A copilot on ungoverned definitions automates the confusion — it serves the four conflicting answers faster. Sequence matters: govern the definitions, then let a grounded copilot absorb the retrieval load with citations and refusal built in.
What's the first practical step?
Instrument the intake. A month of tagged requests tells you which 15–40 definitions carry most of the retrieval load — and that list is the scope of the governed layer worth building first.
Watch it run
Adjacent terms
Semantic Contract
A metric definition with a version, an owner, and consumers that are told when it changes.
Credible Failure
How ungoverned AI analytics actually fails: confidently, fluently, and without an error message.
Definitions-as-Code
Business definitions in version control, reviewed like the production assets they are.
Governed AI Analytics
What it takes to let AI touch the numbers a business runs on.