Metis is the department's research and systems director. She turns scattered research, data, and process knowledge into a clear system model, decision-ready insight, and a practical next experiment you can actually run.
We have more reports than ever, and somehow less clarity about what is actually happening in the business.
The same process breaks every month, and nobody can tell me where it actually goes wrong.
Every team has data. No two teams define the same metric the same way, so the numbers never reconcile.
We are making gut calls in places where the evidence already exists, but nobody has assembled it.
We want to use AI somewhere that matters, but nobody can say which workflow it would genuinely improve.
Half of how this company actually runs lives in one person's head, and that person is going on leave.
Each one is scoped, has a named deliverable, and ends with something you can inspect and act on. No open-ended research retainers to start.
When a decision is due and the evidence is scattered across reports, interviews, and meeting notes.
You receive: known-facts and unknowns register, pattern analysis, labeled hypotheses, options with confidence notes.
When one business outcome depends on a tangle of people, data, tools, and handoffs nobody has drawn.
You receive: actor and dependency map, feedback loops, failure points, leverage-point analysis.
When a workflow is visibly slow or error-prone and the cause has never been isolated.
You receive: current and future-state map, friction and duplicate-effort analysis, prioritized experiment backlog.
When approved documents exist but nobody, human or agent, retrieves the right one reliably.
You receive: source taxonomy, metadata schema, source-of-truth map, retrieval rules, review and retirement ownership.
When you want AI in the business and need to know where it helps and how you would prove it.
You receive: workflow specification, evidence requirements, failure taxonomy, eval set, pilot measurement plan.
Describe the question. If it belongs with another specialist, Metis routes it there instead of forcing the fit.
Bring the question →That sequence is Metis's whole method. She checks where the evidence came from and how current it is before she draws a single conclusion, and she stops at the smallest test that would actually change your decision.
Reconcile a bounded source set, log the contradictions, and extract the signal a decision turns on.
Examine data quality, definitions, and completeness before any pattern is allowed to become a finding.
Make actors, dependencies, constraints, and feedback loops visible, then locate the point of real leverage.
Find delay, duplicate effort, and information gaps, then design the smallest useful improvement to test.
Structure information so people and agents retrieve the right source safely and consistently.
Define the use case, context boundaries, routing, memory, and the quality checks that prove it works.
Minimum:
Useful, if you have them:
Metis owns the evidence, the model, and the experiment. When the work crosses into another specialty, she hands off rather than improvising:
She checks provenance and freshness before concluding, surfaces missing or contradictory evidence rather than smoothing it over, and prefers the smallest decision-relevant experiment. Correlation is never presented as causation. Metis may specify an AI or information system, but she never implies it has been built, secured, or deployed.
A typical first project, shown as an illustration of the working shape rather than a report on a specific client.
One approved meeting transcript and up to three relevant documents, the workflow you know is inefficient, and two or three examples of where it breaks.
A concise map: known facts, source gaps, actors and dependencies, one clearly labeled working hypothesis, and one measurable next experiment.
Whether the experiment is worth running. The map states where the source set is insufficient, so you can see the limits of the reasoning before you act on it.
One bounded source set, one reviewable map, one experiment worth running. That's the whole first step.
