Analysis co-pilot: findings to insights

Findings are AI's home turf: extraction, restructuring, and summarizing over material you hand it. Fed the transcripts, it will pull every seat-cap mention faster and more completely than any analyst. Insights are where it turns mediocre: an insight requires knowing which decision the client is actually facing, and that context lives in your engagement, not in the model. Asked for insights cold, AI produces findings wearing bigger words. Implications are forbidden - not because the model cannot type "target the 210 first", but because that sentence is a recommendation the firm signs and a human owns. Session 1's rail 2 draws the same line from the liability side.

The practical shape of week 3, then: AI drafts rungs one and two at volume, proposes candidate so-whats for rung three that you accept, reject, or sharpen, and never touches rung four. The engagement's judgment budget - your scarcest resource - gets spent where AI is weakest, which is exactly where it should be spent.

Quick test when you cannot tell which rung a sentence sits on: could a competitor's analyst, given the same data, write the identical sentence? "Retention is 94% vs 71%" - yes, that is a finding. "Churn concentrates in single-seat users, so conversion is worth more than acquisition this year" - only someone who knows what Cadence is deciding this quarter would write that, so it is an insight. Findings are properties of the data; insights are properties of the data plus the decision. AI has the first ingredient whenever you supply it. It never has the second unless you write it into the prompt - and by then, you have usually done the thinking yourself.

Prompted with "summarize the interviews", a model returns three safe themes - fluent, balanced, and useless. Here is the naive output against what a disciplined pass finds in the same twelve transcripts:

Nothing in the flat column is false. That is what makes flattening dangerous: it is not hallucination, it is compression toward the middle, and the middle is where insights go to die. The outlier - the pay-double quote - is precisely the kind of signal week 4's storyline needs, and a naive synthesis buries it every time.

The cure is a sequence, and the order matters - characterization comes last, after the evidence is pinned down:

• 1 · Verbatim quotes first, per interview, with the speaker's role attached. "Extract every quote about seats, pricing, or blockers, verbatim, tagged with the speaker's role." No paraphrase yet - paraphrase is where flattening starts.
• 2 · Tag the dissent explicitly. "Find the interviews that disagree with the majority view and quote them." A model asked for consensus produces consensus; asked for disagreement, it is a superb dissent-finder.
• 3 · Count before characterizing. "How many of the 12 actually said security review was the blocker?" The answer - 2 - kills the safe theme and replaces it with a precise one.
• 4 · Keep the outlier visible. One interviewee would pay double for admin controls. That quote rides at the top of the synthesis, labeled as an n-of-1, until a human decides whether it is noise or the finding of the week.

One rail before any of this: interview transcripts carry names. Strip them before anything enters a prompt - role and account tier are all the synthesis needs. Session b8 drills the full anonymization pattern; week 3 is where the habit starts paying rent.

The quantitative half of week 3 has its own division of labor: AI writes the analysis plan and the checks, your analyst runs the numbers. "Given these case questions, list the data cuts that would confirm or kill each hypothesis, and the sanity checks for each cut" is a superb prompt. "What is our retention by plan type" aimed at the model itself is not - findings come from the data, not from the model's memory, and a model asked for a number it cannot compute will often supply one anyway. The 94% vs 71% in your deck must trace to a query someone ran, logged in b3's source log like every other fact.

The handoff artifact is short: three support arguments, one line each, with the findings that back them. Demand: 62% of trial teams hit the 5-seat cap in 3 weeks, and 210 accounts have entered procurement or security review unprompted. Economics: retention runs 94% on enterprise-style seats vs 71% on individual plans, and an enterprise tier lifts ARPU 2.1x. Window: rivals price at $38-45 per seat, leaving room to enter at $30. Next session, b5's pyramid puts a governing thought on top of exactly these three - this page is where their evidence gets finished.

The quality gate before handoff: every bullet has a number, and every number has a row in b3's source log. That is the whole gate. It takes ten minutes, it is mostly mechanical, and it is the difference between a storyline built on findings and a storyline built on vibes. The disciplines compound: b3's log feeds b4's gate feeds b5's anchors feeds b9's audit.

A small habit that makes b5 dramatically easier: label every line of the handoff with its rung. "62% hit the cap in 3 weeks [finding, source log row 4]" reads differently from "demand is organized into buying units [insight, from findings 3-5]" - and the labels expose gaps before the pyramid does. If an argument has three findings and no insight, the so-what work is still undone; if it has an insight resting on one finding, the evidence is thin and b9's audit will say so less politely. The ladder is not just how you think this week; it is how the handoff document should read, because next week's governing thought can only be as strong as the rungs beneath it.