What actually delivers a spot-on insight?
AI analytics is finally real, and everyone now claims you can point an LLM at your data and get answers. So we ran the comparison that matters. We put Perch next to a capable general AI, both pointed at the same real program data, and looked at one thing: which one produces an answer your team sits with, checks against what they already know, and calls spot on.
That reaction, not a benchmark score, is what decides whether an insight gets acted on or quietly ignored. A number can be accurate and still get waved off in the room. The answer that gets used is the one a leader trusts enough to act on.
What “spot on” actually means.
You cannot always name why an answer is good. You just feel it. When we look closely, that feeling comes down to a handful of moments, and every one of them has a version that lands and a version that makes you close the tab. This is the vocabulary of trust, in the words your team actually uses.
Watch for these six in the two questions that follow. They are the whole difference between a nod and a shrug.
The same questions, read by the person who has to act on them.
We compared two things: Perch on your data, and a capable general AI on your data. Same program, same prompts. For each question we show the strong, spot-on answer first, then read both the way your team would.
Total sales fell 18.4%, driven mainly by a contact rate collapse, 56.0% down to 48.2%, with the rest downstream. Agent effort was flat: calls per agent per day did not move.
The drop is uniform across all 10 states, all 17 vendors, every hour of the day, and even the earliest attempts. That signature points to spam-labeled outbound numbers, not lead quality.
Sold-lead volume fell 22.7% month over month. The top of the funnel looks like the bottleneck: opportunities worked dropped 10.7%, and interactions per sold lead rose 21%, so agents are working harder per sale.
This reads as a lead-supply or lead-quality softening. Ticket size and bundle rate held.
Close rate fell 100 bps to 16.1%. Your three cuts, plus what they open up:
Agents, done like for like: before comparing producers, I separated the 8 agents who started in January from the tenured team. Much of the apparent agent decline is just new hires ramping. On a tenured-only basis the real drop is smaller, and points elsewhere. Lead sources, with the link you did not ask for: the vendor decline likely traces to last month’s spam-labeling finding. If flagging hit those vendors hardest, the contacts getting through are less interested, which reads as a lower close rate. The lever nobody flagged: 30% of open opportunities got no follow-up attempt at all. Bigger and more fixable than the vendor mix.
Close rate fell from 17.6% to 15.4%. Across the three cuts you asked for:
Agents: the decline concentrates in a few producers; a couple actually improved. Lead sources: about two-thirds of the drop traces to four vendors, with an estimated sales gap for each. States: NE and OH fell hardest.
Same model. The difference is what it can see.
An AI is only ever as good as the picture it reasons over. Perch reasons over a clean one: every call joined to the lead it was made against, one definition of every metric, the business itself encoded, who is new, what the cap is, what to leave out.
Point the same capable AI at your data as it sits today, and it sees noise: calls that do not connect to leads, three definitions of close rate, no idea which agents are new, test records mixed in with real ones. Same reasoning. Opposite result.
So the answers above are not different because one model is smarter. They are different because of the layers Perch puts underneath the reasoning. Here is what each layer does, and the exact difference it made on the very answers you just read.
Perch builds these layers and, the harder part, keeps them current as campaigns launch, dialers migrate, agents turn over, and product mix shifts. That is what you are buying: not a cleverer model, but a living system of knowledge and structure, maintained so that the AI reasoning on top of it produces spot-on insights continuously, not just on a good day.
It answers the questions you did not know to ask, over a model that already understands your business, the same way every time you open it. Your team will not grade it on a benchmark. They will grade it on whether the answer makes sense, and that reaction is the whole product.
