METHODOLOGY · PUBLISHED IN FULL
The Yesward Intervention Ladder
Seven levels of response to a hesitating buyer, ordered by how much trust each one spends. Level zero is silence, and it is the default. Escalation is earned by evidence, capped by policy, and reversed automatically when it stops paying.
Diagnose the friction first. Decide whether intervention is justified. Use the lowest-cost intervention capable of resolving it. Then prove whether doing so beat silence economically.
Written down because a ladder that only exists in a slide gets climbed. This page is the version the engine implements, including the parts that constrain us.
THE SEVEN LEVELS
Every level has a trigger, a form, and a prohibition.
L0 SILENCE · L1 PASSIVE · L2 NUDGE · L3 ASK
L4 ADVISE · L5 ACT · L6 HAND OVER
| Level | Fires when | What renders | Never | Trust cost |
|---|---|---|---|---|
L0 Silence91.1% OF SESSIONS | Evidence below threshold, no resolvable friction, buyer progressing normally, prior dismissal, consent absent, or the expected value of every other level fails to clear zero. | Nothing. The page is exactly as your team built it — and a receipt is still written. | Never treated as a failure, a fallback, or an untracked outcome. | none |
L1 Passive adaptation2.5% | A specific piece of information is missing and the friction is informational: total cost, a promise, a distinction between two options. | A block in normal page flow. No overlay, no animation, nothing to dismiss because nothing was imposed. | Never moves or hides content the buyer was already reading. | · |
L2 Micro-nudge2.1% | The information exists on the page but the buyer has not found it, and the stall is long enough that pointing at it is worth the interruption. | One line in a corner card, in your brand tokens, carrying the AI mark and a single ✕. | Never blocks the viewport, never re-appears after a dismissal, never mentions the visitor's behaviour. | · · |
L3 One question1.6% · CEILING | The right answer depends on something only the buyer knows, and one question resolves it — commitment horizon, room size, usage, travel dates. | A single question with two or three answers, each mapped to a concrete resolution. Answering it must visibly change something. | Never asks for personal data, never asks a second question, never asks something the site could infer. | · · · |
L4 AI adviser1.7% · OFF IN PILOTS | The blocker is genuinely conversational — several interacting constraints that no single question can untangle. | A right-docked panel that answers questions about the product and the decision, declared as AI, with the same one-dismissal rule. | Never sells, never claims to be human, never invents a promise that is not in the approved claim library. | · · · · |
L5 Agent action0.7% · OFF IN PILOTS | The blocker is a task, not a question: check an install date, hold a slot, save a configured quote, verify coverage at an address. | The action, plus an action receipt showing exactly what was done on the buyer's behalf and how to undo it. | Never takes an action outside the declared list, never commits money, never acts without saying that it did. | · · · · · |
L6 Human handoff0.3% · OFF IN PILOTS | Deal size and evidence both high enough that a person costs less than the lost sale, and a person is actually available now. | A named human, marked HUMAN rather than AI, with the context already passed across so the buyer does not repeat themselves. | Never promised when nobody is there. Never used as a euphemism for a queue. | · · · · · · |
DRAWN TO SCALE ON ILLUSTRATIVE PILOT TRAFFIC — RESTRAINT IS NOT A FOOTNOTE, IT IS THE DISTRIBUTION.
SURFACE ANATOMY
Four things every surface above L1 must carry.
Not guidelines. Structural requirements enforced at render time, which means a surface that cannot satisfy them does not ship.
01 · AI SELF-ID
It says what it is
Every surface carries the AI mark. EU AI Act Article 50 by construction rather than by policy. A human is only claimed at L6, where one is really there.
02 · ONE DISMISSAL
✕ once, gone for the session
One dismissal suppresses every surface for that visit — not just the one that was closed. No second attempt, no different wrapper, no reappearance after a scroll.
03 · THE PROMISE IN WRITING
Closed stays closed
The surface states the rule it is bound by, on itself, where the buyer can read it. A promise you have to look for is not a promise.
04 · HOST TOKENS ONLY
Your brand, not ours
Type, colour, radius and spacing come from your design system. Nothing on your site should look like it was rendered by a vendor.
FORBIDDEN AT RENDER TIME
Enforced in the engine, not in a prompt. A treatment that cannot say it, will not say it. The full guardrail list →
TWO SAFEGUARDS THAT OVERRIDE EVERYTHING
Frustration guard · always on
Rage clicks, rapid reversals and repeated dismissals return the session to silence. Not a switch an operator can turn off.
Global stop · one control
Clamps every decision to silence across every journey, while still writing a receipt for each one. Silence stays a logged decision, never a gap in the record.
THE FOUNDING CEILING
Levels 4 to 6 are built, inspectable, and switched off.
Founding pilots run with a ceiling of L3: silence through one question. The higher levels exist in the product and you can read their specifications above — they stay off until low-trust-cost responses have been proven both safe and valuable on your own traffic.
The ceiling is a policy value the engine clamps to, not a feature flag in a slide. In the console, levels above it render dashed and inactive — visible, never hidden, and never quietly raised.
See the policy controls ↗LADDER CEILING · CURRENT POLICY
CEILING L3 · ABOVE IT: BUILT, INSPECTABLE, CLAMPED TO SILENCE
THE VOCABULARY
Twelve terms we use precisely, defined once.
Five friction families, four cost terms, three measurement terms. If a vendor cannot tell you which of these their product addresses, it probably addresses whichever one you name.
FRICTION
Choice friction
Two or more options are equally viable and nothing available to the buyer separates them. Resolved by a distinction, worsened by a third option.
FRICTION
Price friction
The total cost, or its shape over time, is unclear. Distinct from unwillingness to pay — and usually resolved with information rather than margin.
FRICTION
Commitment friction
Doubt about what is being locked in: term length, cancellation, reversibility. The dominant friction in subscription and service purchases.
FRICTION
Trust friction
Doubt about the company, the promise, or what happens when something goes wrong. Resolved by the specific promise, never by badges.
FRICTION
Coverage friction
A hard constraint rather than a doubt: eligibility, availability, dates, address, compatibility. Sometimes the honest answer is no, said early.
COST
Trust cost
What a response spends in attention and goodwill, independent of whether it works. Rises with the ladder level, and is priced into the decision.
COST
Cost of speaking
The measured version of trust cost: dismissals, rage clicks, bounce delta and suppressed sessions, reported next to the lift rather than beneath it.
COST
Intervention cost
Everything an action costs to deliver: model spend, human time, and any margin given away. Netted out before anything is called lift.
TREATMENT
Silence treatment
Doing nothing, treated as a ranked candidate action with an expected value and a receipt — not as an absence of decision.
MEASUREMENT
Qualified outcome rate
Share of eligible sessions reaching a genuinely valuable outcome. Not clicks, not conversations opened, not messages seen.
MEASUREMENT
Restraint rate
Share of sessions deliberately left alone. Reported as a headline figure, because a system that never abstains cannot be measured against one that does.
MEASUREMENT
Permanent holdout
A stable-hash share of eligible sessions withheld for as long as the policy runs, so the counterfactual never goes stale. Floor 5%, enforced server-side.
USING THIS
Take the vocabulary. We would rather the argument spread than stay ours.
Use the ladder in your own CRO reviews, name the frictions in your own analytics, and hold your current vendors to the cost-of-speaking test. Attribution appreciated, not required.
Related: when no action wins · next best action · considered purchases
HOW TO CITE
Yesward. “The Yesward Intervention Ladder.”
yesward.io/company/methodology/
Accessed [date].
Version 1.0. Changes to level definitions, guardrails or ceiling defaults will be dated on this page rather than edited silently.
Run the ladder against your own traffic, at level zero.
Shadow mode decides every session and renders nothing. You see the distribution before you agree to anything.