AI agents · RevOps · build philosophy
10-80-10 is a dial, not a template
The three build patterns are three positions on one dial — how much neural do the edges need? The deterministic core is always there. The neural miles are optional, and you only pay for them where ambiguity exists.
The dial · three positions, one principle
Neurosymbolic
Forecast · Win/loss · Churn · Commit
10 - 80 - 10
Full pattern. Vague ask in, judgment-heavy narrative out — both neural miles earn their keep.
Guarded judgment
Lead routing · Deal risk · CPQ · Hygiene
~5 - 90 - 5
Edges shrink. A structured trigger ate most of the first 10; the output is a short note, so the last 10 shrinks too.
Deterministic pipeline
Pipeline coverage · Comp & quota · Activity capture
0 - 100 - 0
Edges gone. The trigger IS the spec, and a sync has nothing to say. Sometimes a tiny narration (0-100-5) — never computation.
The test · only use neural where language ambiguity exists
Two questions decide the dial position — nothing else does.
Ambiguity coming IN?
Vague human question, messy scope → you need a first 10% to frame the problem
Ambiguity going OUT?
A human needs a nuanced explanation → you need a last 10% to narrate it
Neither?
It's a pipeline. Adding an LLM anyway just adds cost and new failure modes
Purest example — Activity capture: email → CRM sync. No question to interpret, nothing to explain. Forcing a 10-80-10 shape onto it means inventing work for a model to do — the exact anti-pattern that makes agent pilots stall.
The proof · a healthy fleet is a pyramid
4neurosymbolic — true reasoning agents
7guarded judgment — rules validate, model narrates
10deterministic pipelines — code does the work
This is how mature automation is built everywhere: most agents are pipelines, a few are guarded, only the top handful truly reason. An inverted pyramid — reasoning models everywhere — is how you get impressive demos with unpayable bills and unauditable answers.
Read the empty cells as the selling point
Every agent runs a deterministic core. The neural miles are optional add-ons you pay for only when ambiguity exists.
"None" in a first-10% or last-10% cell isn't a gap — it's the discipline: we won't put an expensive model where plain code does the job perfectly.
One dial: how much neural do the edges need? Code always carries the middle.
