notes · product

Where the synthesis should actually happen

May 21, 2026· Benjamin Dysin

Most research interviews have a hidden second job. The first one is the conversation you scheduled. The second one is everything that happens after the call ends: transcribing it, re-listening, tagging, clustering, drafting the findings, drafting the readout, re-drafting the readout because the PM has a question about slide four.

The first job ends at minute 45. The second job ends roughly two weeks later. By then the decision the research was supposed to feed has already been made another way, usually by Slack and by whoever felt confident enough to call it.

That two-week gap is the bug. Not the friction, not the inconvenience. The bug.

You can shorten it. Otter speeds up the transcription. Dovetail and Notably do AI-assisted tagging on a finished transcript. All three are real products and they all genuinely help. But the premise they share is that there are two phases (conversation, then analysis) and the work is to make the second one faster.

I think the framing is wrong.

The synthesis should happen during the conversation, not after it.

What “during” actually looks like

Inside Lacudelph, every participant turn fans out into four Claude API calls running in parallel.

The first call is the conductor. Given the brief, the objectives, and the conversation so far, what’s the next question. This is the part most AI interview tools stop at.

The second call is the contradiction detector. Does this answer conflict with something the participant said earlier in the same session, and if so, is the conflict load-bearing (they actually changed their mind, or one of the answers was a polite version) or just a phrasing mismatch. If it’s load-bearing, the conductor sees it on the next turn and asks about it.

The third call is the live extractor. Pull a structured finding from this specific turn, scope it to the objective it served, and attach it to the literal sentence in the transcript that supports it. By the end of the call there is a per-finding receipt. You can click the finding and land on the exact phrase.

The fourth call is the closer. It builds the participant takeaway (handed back to the participant on the way out of the call, which is its own quietly useful thing) and the cohort row that the final report will join.

By the time the conversation ends, the synthesis is already there. Not “a draft” or “a starting point”. There. With the quotes, with the contradictions surfaced and resolved on the call, with the takeaway already delivered. The next interview an hour later starts from real prior knowledge instead of a backlog.

Why this changes the operating model and not just the speed

The standard research operating model assumes synthesis takes weeks. Everything downstream of synthesise gets timeboxed by that assumption.

You can’t iterate the brief between participants, because you don’t know yet what you learned from the last one. You can’t pre-empt a decision being made on stale gut-feel, because your update isn’t ready. Continuous research as a discipline keeps not quite working, and the reason it keeps not quite working is that the bottleneck was never the calls. It was the post-call work.

If synthesis is live, the shape of the operation changes. The brief becomes something you adjust between participant 4 and participant 5. The cohort becomes something you stop early when you have enough signal, or extend by two when you don’t. The decision the research feeds gets made when the research is ready, not when the backlog catches up.

That’s a bigger claim than “faster”. A faster version of the old shape is still the old shape. A live version is a different shape.

The bit where I admit what I don’t know

A few things could invalidate most of this.

The biggest one: maybe live per-turn extraction is meaningfully worse than a senior researcher reading twelve transcripts in one sitting. There are things you catch at the cohort level that no per-turn pass will see. Second-order themes. The framing nobody named but everyone was reacting to. The pattern where three participants used the same metaphor and you only notice on the third read. Lacudelph runs a cohort-level pass after the calls close, but the live work is per-turn, and a skeptical researcher could reasonably say the cohort pass is doing the actual interesting work and the live calls are theatre. I don’t think they are. I’m not certain.

The second one: maybe the two-week gap isn’t the bottleneck most teams feel. Maybe the real bottleneck is “we didn’t know what we were trying to learn before we started,” and faster synthesis at the back doesn’t fix a brief that was confused at the front. If that’s the dominant failure mode, the whole argument shifts from “compress the back half” to “fix the brief,” and Lacudelph’s wedge gets narrower.

The third one: the operating-model story assumes the org around the research is already wired to consume insight quickly. A research team that ships into a quarterly planning ritual gets less out of live synthesis than one that ships into a weekly product review. If your org’s metabolism is slow, faster research artefacts just sit in a Notion page longer.

I’ll know more once enough cohorts run on real briefs. For now, the narrow claim I’m willing to defend is: the two-week gap between the conversation and a useful artefact is not load-bearing. You can close it. Whether closing it changes the operating model is downstream of whether the org was set up to use the time.

The conversation itself still has to be good. None of this rescues a brief that didn’t know what it was asking, and no amount of structured extraction makes a participant say something interesting they weren’t already going to say. But the post-call work, the part everyone treats as the price of doing real research, doesn’t have to exist.