Amazon just commodified the AI-commerce stack. Differentiation moves up the stack — to brand, to trust, to decision quality. NERVA is the layer that sits between the assistant and the customer's actual interest, scoring every decision across five states: COMMIT, HOLD, WAIT, CONSULT, TOXIC.
Three categories of decision live outside ASA's optimization function. Each is a structural gap a retailer at sub-enterprise scale cannot fill alone — and one Amazon is structurally disincentivized to close.
Models, infra, and learnings from Amazon's own retail business — customizable to your catalog and brand voice. Real, hard to replicate, available now.
Operator decisions about the AI — thresholds, upsell aggression, confidence floors. Customer decision quality — is the shopper right 90 days later? Brand-objective drift — invisible without a governance layer.
The position is structurally available only to challenger brands — and to the infrastructure that serves them. The category will be defined by whoever ships first.
Proceed. The kernel doesn't slow down good decisions.
Returned with the missing evidence stated. Re-score when satisfied.
Queued with a defined release condition. Price cycle, season, data window.
Escalated above the ops team with explicit sign-off required. Default for one-way doors.
Blocked with a written rationale. The rationale becomes a precedent.
Entropy measures whether the decision-maker can articulate what they want without naming a brand. Stakes weight financial, reputational, and relational exposure. Evidence weight only counts independent sources — a single opinion restated three times still counts as one.
Any high-stakes irreversible decision escalates to CONSULT or TOXIC by default — even when evidence weight is high.
Hit-rate of decisions that followed the kernel vs. decisions that overrode it, measured across the 90-day window. No number is claimed before the pilot runs. Tracked openly and reported at day 90 — including, if it comes to it, a negative result. The willingness to be falsified is the product.
Cart hesitation, repetitive or contradictory search queries, rapid category-switching within a session. The shopper cannot yet articulate what they want without naming a brand.
Order value, custom-order or non-returnable categories, sizing-dependent fit, and downstream commitment chains (subscriptions, large gifts, bundled items).
Review density and independence, sizing/spec clarity, third-party validation, and presence of comparable purchase history for the same shopper.
Configuration A is the lowest-risk deployment — a governance layer over ASA configuration changes, operator-facing only. Configuration B is the higher-narrative deployment — a decision panel surfaced inside the customer's conversational session. The pilot recommends sequencing A first, then B.
Before any change to recommendation thresholds, pricing rules, conversational flow, or upsell triggers ships to ASA, the change is scored by the NERVA kernel and routed by state. COMMIT decisions flow normally. Everything else gets a guardrail — not a checkpoint.
Configuration A deploys against a single category or workflow. NERVA scores every ASA config change before it ships.
Operator lift score, deployment regret rate, and CONSULT/TOXIC precision reviewed against pre-committed targets.
Configuration B deploys on a single category or shopper segment. Return rate at 30/60/90 days is the primary metric.
measured vs. matched control period
measured vs. concurrent control segment
No retroactive metric changes. If these conditions are not met, the layer is not adopted — and the pilot ends cleanly. The willingness to be wrong is the product.
Decisions that followed NERVA outperform decisions that overrode it. Effect size honestly stated against sample size; no claim of significance the data can't support.
Of changes shipped under NERVA, materially fewer get rolled back within 30 days than were rolled back in the equivalent prior window.
The customer panel can suppress some sessions — that's intended. But total revenue net of returns must improve, not collapse. Single-digit conversion dip is the ceiling.