At 3pm someone asks your analyst: why did new patient starts drop? They pull the trend. Down nine percent. They cut by region. Nothing. Then by specialty, where oncology looks soft. Deeper into oncology by site of care, then by start month. And there it is: a cluster of community sites that all stalled after February.
Three hours and eleven queries later, they write one memo: oncology starts at community sites fell by half after the February formulary change. The memo is probably true. It is also one branch: the one they had time to walk.
That is the whole problem. A question like this opens too many places that need investigation, and an afternoon covers less than a tenth of them.
This has not gone unnoticed. Most analysts already have a mitigation: a text-to-SQL product, a BI copilot, a visualization agent. These collapse the cost of asking one question and getting a decent answer back.
But that is not a sufficient fix. They automate query writing. The analyst still has to read the result, validate it, and decide what to ask next. The bottleneck is the decision-making, and automating that brings its own problems.
Problem 1
Agentic systems mostly solve this by planning up front: break the question into five or six smaller ones, run them, and stitch the answers together. For retrieval that works, because those smaller questions do not depend on each other.
However, analysis is not retrieval. Question four only exists because of what questions one through three found, so a plan written up front is a plan written blind.
Problem 2
The obvious fix is a supervisor that reads the agent’s outputs and decides what to ask next. That solves the blindness, but only partly. The supervisor sees only results or summaries, and the nuances of the investigation (the assumptions, the underlying distributions, the discarded segments) get heavily distilled, or lost. So the supervisor confidently steers an analytical agent that has the full picture, using a distilled view of it. The same failure as before, one level up.
Problem 3
Even if systems were able to go deeper in an analysis, the breadth would still not be covered. The memo on the drop in new patient starts only focused on one branch because the investigation anchored on specialty, and everything after was constrained by that choice. This is the core of the problem: the highest-leverage decision gets made first, with the least information. No amount of depth can rescue a poor decision at the top, and the unexplored branch never speaks up.
Problem 4
This is the most direct blocker to solving problem 3. In practice, good starting points or anchors are rarely in the data. Nothing in a table says that payer mix was ruled out in the last business review, that the brand is near loss of exclusivity, or that shipments slipped because of a vendor issue. That lives in the organization: team meetings, who owns a territory this quarter, assumptions that shift, and what has driven a pattern before.
The first two problems share a root: a decision got made somewhere other than where the evidence lives. The other two are also a pair: both are about where an investigation starts. Hence, the following four principles are created.
The first two principles settle the architecture. The machinery around an investigation has to be systematic. It carries out a decision already made in front of the data, runs long enough to finish the thread, and stops before it runs away with itself.
So how do we fix the planner? Remove it. Let the analytics agent drive the direction itself. That provides a two-fold advantage. You cover the cost of the answers and nothing else, with no planning overhead on top. And the investigation is exactly as smart as the agent, so every improvement to the agent lands inside it for free, skipping the cost of training a supervisor separately from an agent.
The fix for breadth is not to pick a single vertical at all. The investigation is not one deep dive. It is several, fanned across breadth points, each digging on its own. Anchor one on specialty, one on referral timing, one on payer mix, one on patient support, one on the competitive picture. Let each go as deep as its thread deserves.
Agreement and disagreement both become evidence. When two investigations reach the same conclusion through unrelated data, that is far stronger evidence than one reaching it through five careful steps. And when the timing branch says formulary while the support branch says prior-auth delays, that tension is the most useful sentence in the report.
Branches can be abandoned without ego. Sunk cost is a property of the analyst, not the analysis. A branch that dies in two steps costs a parallel system nothing to drop. That makes the marginal branch cheap enough to actually try. That is the unlock: not that the machine is smarter, but that it can afford to be wrong four times out of five and still find the one branch that matters.
How we built it
In Orcana this is an Analytical Loop. Point one at a question and it takes the depth off your hands. It decides what to ask next in front of the result, runs each step as a full analysis, and keeps going until the thread is answered rather than until time runs out.
And you do not fire just one. Loops run concurrently, anchored on different breadth points. A single question comes back as one multi-level analysis across every angle worth opening, instead of you slicing the work into pieces and driving each piece by hand.
You also do not have to scroll through every question-and-answer pair. Each loop ends with an artifact written for three readers: the bottom line for a VP, the method for an analytics lead, and the full list of technical decisions (assumptions, definitions, derivations) for whoever has to defend the number in the room.
✦ Bottom linefor the VP
Oncology starts fell nine percent, concentrated in community sites after the February formulary change.
Recommendation: Treat the formulary change as the primary driver and the prior-auth backlog as a compounding factor. Brief the field on Q3 access before the next start wave.
start_date, not a sample or bridge fill.start_date, not first_fill_date, which lags by up to three weeks and would misassign 214 patients.We argue hardest for that last layer. A loop, like every question in Orcana, is deterministic and inspectable end to end. Every assumption, every filter, every condition, definition and cohort rule behind a number stays on the record and can be shown on demand.
Nobody has to trust an investigation they cannot audit. Open any step and see exactly what it did and why. That is the difference between a result you can defend in front of a committee and one you can only quote.
Orcana captures the context that decides where to look. Which team you are on. What they are accountable for this cycle. The lines of questioning this group always ends up wanting. The nuances your data operates under: the six-week restatement, the cut settled last quarter.
That is what makes anchors proposable instead of generic. A market-access lead asking why volume slipped gets a fan pointed at formulary tier movement, prior-authorization rejection, and abandonment at the pharmacy counter. A field-operations lead typing the same five words gets call-plan attainment, territory vacancy, and targeting fit. Same question, different user, different openings.
And it compounds. Orcana learns from the investigations that came before: which anchors paid, which have moved recently, what was asked last month and what came of it. Every fan-out starts closer than the last, which leaves our users doing the only part that was ever theirs: glance, decide, ship.
One of our early customers arrived with a problem we had not anticipated. Not I don’t trust the answer. Something more basic.
“I don’t know what I can ask, or how this helps me find things in my data I don’t know are there.”
That turns out to be common. Plenty of business users badly want insight out of a tool like this and have no idea where to start. And once they start, they do not know how to go deeper, or wider.
A fan answers this in a way a question box cannot. They did not need a question list. They needed a topic. Pick something you care about, ask one open-ended question about it, and let the loop take it from there.
We had already ingested a good deal of their context: quarterly business reviews, past and upcoming drug releases, the things that only exist inside the organization. So an open-ended question did not come back generic. It came back as a deep dive that had already covered several bases, chosen from what we knew about their business rather than from the shape of their tables.
The depth of it surprised them. That was the part that mattered. Reading it, they could see which threads deserved more, and they fired four more loops in parallel. Between them, five cuts of the data settled a purchasing decision they had been unsure about.
The only thing they supplied was direction: the business judgment that was theirs and nobody else’s. Everything after that was coverage they did not have to plan, run, or stitch together.