Why it works
Successful AI and intelligence needs the right data.
Every report, model, agent, and dashboard is downstream of the data it can reach, and nothing built above the data can add information the data does not contain. Metis is the Enterprise Intelligence OS built on that belief: it measures the data first, raises it where it's short, and only then runs what the data can carry.
How we found out
Years of building AI taught us where it breaks.
For years, we built AI systems and agents for organizations, noting where they succeeded and where they failed.
Better models didn't fix it. Better data did. Grounded answers are capped by what retrieval can find. An index over a stale corpus answers fluently about a world that no longer exists. A question about a fact nobody recorded has no answer, however good the search. Every fix on offer sat above the data: a bigger model, a different embedding, an agent framework. Some helped at the margin. Few of them gave the system what it was missing.
“Successful AI projects need a real problem to solve, and the right data.” — Andrew Ng.
Fiscally responsible AI is anchored to ROI, and AI investments that are not will fail. No processing can make an output carry more information about the world than the data it came from. And because the same data is available for one decision and not another, that ceiling can be scored, per decision, before anything is built.
We stopped asking which AI to buy. We started asking what problem it solves, whether the data can carry it, and what it returns. Metis is designed around this central principle. We target a 400% ROI (over 3 years) on every engagement, and we measure it in the unit's own metric, with a signed chain from the investment to the returns.
The chain
From data you have to returns you can measure.
Intelligence is a chain, and it breaks at its weakest link. Data availability comes first because it is the ceiling; decisions come next because availability is only ever measured for a decision. Metis makes every link a thing you can see and score.
- 01
Data availability
Does the data a decision needs exist, can it be reached and lawfully used, is it true, current, and understood, and is it fit for the instrument. Scored per decision, on ten dimensions.
- 02
Decisions
The unit everything hangs off. The same data is available for one decision and not for another, so availability is always scored against a decision, never in the abstract.
- 03
Instruments
Reports, diagnostics, predictive models, recommendations, grounded answers, and agents. Each tier demands more of the data than the one before it, and the tolerance for missing data falls toward zero as instruments move from describing to acting.
- 04
Outcomes
Every decision is logged with what was expected. Outcomes are measured at the horizon and compared, in aggregate, so skill is separated from luck.
- 05
Returns
Return on investment (ROI), stated in the unit's own metric: what was invested, how far availability rose, which instruments that made possible, what the metric did, and what came back.
01 · Data availability
Ten dimensions, three questions.
Availability is not a single score and it is not on or off. It is ten measurable dimensions in three bands, and a decision's data is only as available as its weakest one.
Can the instrument get it, lawfully
Existence: was it recorded at all, at the grain the decision needs. Accessibility: can the systems and people that serve the decision reach it. Permission: may it be used for this purpose, at this grain.
Can the instrument trust and understand it
Quality: is it true to what it claims. Currency: is it current when the decision is made. Provenance: where it came from, who owns it, what happened to it. Meaning: does everyone, human and machine, compute the metric the same way.
Does it serve this decision through this instrument
Relevance: would knowing it change the choice. Sufficiency: is there enough history, enough labeled outcomes, enough of the rare cases. Shape: can the instrument consume it without a preparation project.
Scores combine as a weakest link, never an average. A dataset with perfect quality that the instrument may not lawfully use has an availability of zero for that decision.
02 · Decisions
Find data for a purpose, not a purpose for data.
Relevance is not a property of data; it is a relation between data and a decision. So the ceiling stays where the theory puts it, on data availability, and the decision is what availability is measured for: Metis inventories the decisions an organization has to make, derives what each one needs, and scores the data against that. That is decision-driven, not data-driven. Data-driven in the wrong sense starts from what exists, looks for uses, and ends up answering only the questions the data happens to be able to answer.
Inventory the decisions
For each unit, the few decisions with the largest value at stake: alternatives, objective, frequency, the owner, and the metric the unit already reports.
Derive the variables
What, if known, would change the choice. Priced by the value of information, so data that would not change anything is not collected.
Score and admit
Each decision's data is scored on the ten dimensions against the instrument it needs. An instrument runs when the data can carry it, and the gap is named when it cannot.
This is why the pivot away from starting with a model matters: a decision two gaps from working beats a model with no data underneath it.
03 · Instruments
Six tiers, tightening toward action.
Every instrument answers a different question and demands a different profile of the data. The action-facing demands rise monotonically.
Descriptive
What happened. Reports and dashboards over conformed data with one definition per metric.
Diagnostic
Why. Joining across domains, with the causes recorded and not only the effects.
Predictive
What will happen. Models trained on representative, labeled, sufficient history, with calibrated confidence and explained drivers.
Prescriptive
What to do. Recommendations that need something the others do not: the record of past decisions and their outcomes.
Generative
Explain and answer. Grounded answers whose accuracy is capped by what retrieval can find, not by the size of the model.
Agentic
Agents that execute. The most demanding profile: current data, per-action permission, and full provenance, because the failure is an action, not a belief.
AI is the engine of the last four tiers. It is a small part of the operating system, and it is the part most likely to be funded above data that cannot support it.
04 · Outcomes
Quality decisions, not lucky ones.
A decision is a process. A choice is its output. An outcome is what the world does next, and it includes chance. Metis scores the decision, logs the choice, and measures outcomes only in aggregate. Higher quality decisions mean better outcomes and returns, over many decisions, and that is the reason to score them.
Logged at commitment
Alternatives considered, the choice, the rationale, the expected outcome, and who decided, human or agent. The log is itself data, and it is the data that prescriptive instruments need.
Measured at the horizon
Outcomes are joined to their decisions and reviewed by cause: was the frame wrong, the data unavailable, the reasoning flawed, or the commitment unmet.
Calibrated over many
Across many decisions of the same kind, stated confidence is compared with what happened. Calibration is a measurable, improvable skill, and it is how an organization knows its instruments are telling the truth.
A single outcome says almost nothing about a decision. A hundred say a great deal.
05 · Returns
Invested this much, moved this metric, returned this.
Returns are return on investment (ROI): intelligence and AI investments stated in the metric the unit already reports, with the chain written down before the work starts: the investment, how far availability rose, which instruments that made possible, what the metric did against its baseline, and what came back.
A baseline and a counterfactual
The metric's value and trend before the work, and an agreed way to say what would have happened anyway. Before and after alone is not attribution.
A fixed window and an owner
The measurement period is set at the start, and a named executive defends the metric and the value per unit that finance signed.
Adoption, not deployment
An instrument counts only for the share of decisions actually taken with it. A forecast nobody uses returns nothing, however accurate.
This is what makes outcome-based commitments possible: baseline, counterfactual, window, and validation are objects in the system, not a negotiation afterwards.
What you might be thinking
The objections, answered from the theory.
Each of these is reasonable, and each has an answer that follows from the ceiling.
- We already have a warehouse. Our data is fine.
- A warehouse raises accessibility and quality. It does not raise relevance, and it can be built without a single decision written down. Scored against the decisions that matter, the usual surprises are meaning (how many definitions of one metric exist across teams) and existence (whether the unit records its own decisions and what happened next).
- We will use a bigger model.
- A bigger model approaches the floor the data sets and cannot go below it. Nothing built above the data can add information the data does not contain. Once a model clears adequacy for the decision, the next dollar goes to the weakest availability dimension, not to a larger model.
- Our data is not ready, so we cannot start.
- Scoring is how you start. Reports and diagnosis work at the Foundational tier, and a decision two gaps from working beats a decision six gaps from working, even if the second is more valuable. Run what the data can carry today, and close the named gate for what it cannot.
- Agents will handle the integration.
- Agents are the tier with the least tolerance for missing data: currency, permission, provenance, and accessibility all near their maximum at once, because the failure is an action, not a belief. A standard tool interface in front of an ungoverned system gives an agent fast access to bad data.
- This is data governance by another name.
- Governance is a function of availability, not overhead: for an agent, lineage and permission are how it knows what it may do and whether its inputs are true. But the deliverable is never a catalog. It is a changed decision with a measured consequence, stated as returns in a metric the unit already reports.
- We do not have time for an assessment.
- The assessment is three to seven decisions per unit, scored with evidence, on a fixed cadence. The alternative is finding out after the spend which instruments the data could not carry.
- Is this for finance, or supply chain, or service?
- The journey is measured on the metric the unit already reports: close-cycle days, stock-out rate, cost per interaction, forecast accuracy. The binding gates cluster by function, and they are known.
- How do we know the score is not gamed?
- Evidence per anchor, sampled verification, weakest-gate scoring, an independent scorer for anything tied to money, and instrument outcomes as ground truth: if the score says model-ready and the forecast is not trusted, the score is wrong and is re-anchored.
How it lands
Forward-deployed, inside your organization.
The method cannot be run from outside, and seeing the map is not walking the path: scoring ten dimensions with evidence, against the column each instrument requires, and closing the binding gate in the right discipline is the engagement, not the reading. Orama's forward-deployed engineers land Metis with your decision owners, on your systems, on infrastructure you choose: your cloud, on-prem, or dedicated.
- 01
Inventory the decisions that matter and who owns them
- 02
Score each decision's data on the ten dimensions, with evidence
- 03
Write the returns chain and get it signed
- 04
Raise availability: connect what exists, and create the data that was never captured
- 05
Run the instruments the data can carry
- 06
Log every decision, measure outcomes, re-score on a cadence, and hand the loop to your own teams
Every engagement leaves the platform stronger for the next one, so the second organization's first step is faster than the first's.
See it on your own decisions
Bring the decisions that matter and the data you have. We score availability first, then build what the data can carry.