Ask a GTM team what their AI agents did last quarter and you'll get a list of outputs: summaries generated, emails drafted, accounts researched, calls transcribed. Ask what moved, and the room goes quiet.

That gap — between an agent producing something and a deal actually changing state — is the whole problem. It's also the thing we spend most of our build time on. This week's build-in-public post is about the two primitives we're building to close it: Findings and Plays.

Output is easy. Consequence is hard.

The industry has been generous with itself on this. Roughly 3 in 10 agent teams aren't evaluating their agents at all — only about half run offline evals before deployment, and just over a third evaluate live traffic (LangChain, State of Agent Engineering, December 2025). Meanwhile 88% of organizations use AI but only 39% report any EBIT impact from it, and most of those attribute less than 5% (McKinsey, The State of AI, November 2025).

That is not a model problem. Frontier models are extremely good at producing plausible GTM artifacts. It's a plumbing problem: nothing connects the artifact to the pipeline. The agent writes a beautiful account summary into a chat window, and the opportunity record it describes stays exactly as stale as it was that morning.

So when we designed how the motion actually runs on wysdym, we started from the other end — from the CRM write — and worked backwards.

Findings: drift, with the fix attached

A Finding is what the platform surfaces when something in your pipeline has drifted from your GTM truth. Qualification drift. A champion who changed roles. A talk track that keeps underperforming against a specific competitor. Deal risk that nobody has flagged yet because nobody was looking at that account this week.

The part that matters isn't the alert. Alerts are cheap, and GTM leaders already ignore several dashboards' worth. The part that matters is what travels with it: the evidence, the graph paths the reasoning touched, the agent it's attributed to — and the exact CRM write the platform proposes to make.

That last piece is deliberate, and it's the trust boundary. A human approves the write, not the inference. You are not being asked to audit a model's reasoning; you're being asked to look at one concrete change to one record and say yes or no. Each finding type carries a disposition — handle it manually, run a Play, route it to an agent, notify someone, or auto-dismiss — so the volume is governable instead of overwhelming.

This is the opposite of the failure mode Gartner keeps describing, where governance is treated as binary: agents either locked down to uselessness or trusted with everything until an incident forces a rollback. Graduated autonomy is the only version that survives contact with a real revenue team. And the market has already voted — human validation of agent outputs went from 22% to 63% of leaders in four quarters (KPMG AI Pulse, Q1 2026).

Plays: the routine, not the one-off

Findings tell you something drifted. Plays are how the response becomes repeatable.

A Play chains typed skills into a routine — the same way your best AE runs the same sequence of moves on every deal that stalls at the same stage — and every Play is wired to a measurable deal outcome. Not "the agent completed 47 tasks." Stage movement. Because if a routine can't be tied to a deal event, it's an activity metric wearing a costume.

The architecture is in place and the Play surfaces exist, but the backend engine and scheduler that produce and run these on live customer signal at production scale are the near-term build, alongside the Observe SDK and cross-vendor outcome attribution. We'd rather say that plainly than imply a scheduler we haven't hardened yet.

Why this belongs in the layer, not in an agent

Every agent vendor could build its own version of this. Some will. The problem is that a Finding produced inside one agent's walls only knows what that agent saw, and a routine that lives inside one agent only works for teams who bought that agent.

wysdym isn't another agent — it's the operating layer your agents and your team run on. Findings and Plays sit in the layer for the same reason grounding and governance do: so they work across whatever agents you're running, and so what one agent learns doesn't stay stuck inside it. Findings feed the graph. Approvals, edits, and overrides feed the graph. Outcomes grade what was already in the graph. The next agent you plug in starts from all of it.

That's the whole bet, and it's why the tagline isn't about agent count: more agents won't grow revenue. Compounding intelligence will. You can see how the five pillars fit together on the platform page.