YESWARD

NEXT BEST ACTION

Next best action — including the action of doing nothing.

Yesward ranks candidate actions for a live web session by expected incremental gross profit, with silence in the ranking rather than outside it. On considered purchases, silence usually wins.

DEFINITION

Next best action is a decisioning pattern: given what is known about a person right now, choose the single action from a defined set that is most likely to produce the outcome you want. It came out of banking and telecoms, where the set was offers, calls and letters.

Ported to a website, the pattern keeps one assumption it should have dropped: that an action will be taken. The set almost never contains an entry for “leave this person alone”, so the system cannot choose it, cannot measure it, and cannot learn that it was the right answer.

A DISTINCTION WORTH KEEPING

Next best action vs next best offer.

Next best offerNext best actionYesward's version
The set of optionsProducts, bundles, prices, discountsAny intervention: message, offer, call, content, journey changeSeven levels of response, of which the first is silence
Implied goalSell something nowAdvance the relationshipRemove whatever is blocking a buyer who already wants to proceed
Failure modeDiscounting people who would have paid full priceActing on everyone because the model always returns a best actionBeing too cautious — which the permanent holdout also measures
“Do nothing” isAbsentA suppression rule, applied after the choiceA candidate action with its own expected value, chosen ~9 times in 10

The difference between a suppression rule and a candidate action is measurement. A suppressed session is missing data; a chosen silence is a data point.

THE ZEROTH ACTION

Sometimes the next best action is no action. It should have to win that on the numbers.

THE OLD QUESTION

Which action should we take?

Answerable, and answered well by several vendors. It just starts one question too late.

THE EARLIER QUESTION

Should we take an action at all?

A buyer working carefully through a €1,800 decision is not idle. Interrupting them has a cost that belongs in the calculation.

THEN, ONLY THEN

What is the smallest intervention likely to beat silence economically?

Not the most persuasive one. The cheapest one that resolves the actual blocker.

Most CRO and personalisation software is structurally incentivised to do something — a system that renders nothing looks broken in a dashboard built to count impressions. Ours counts decisions instead, which makes silence visible, comparable and improvable.

THE RANKING

Rank actions in profit, not in propensity.

A propensity score tells you someone might convert. It does not tell you whether your message helped, or what it cost you to send. So the score we rank on is an expected value in money, and every candidate — including silence — is scored the same way.

p(resolve)How likely this response is to remove the friction we think we see — learned per friction type, per journey step.
margin at stakeContribution margin of the outcome in play, taken from your server-side data — not order value.
trust costA price on the interruption itself, rising with the level of the response. This is the term the category is missing.
delivery costModel spend, human time at level 6, and any margin given away if discounts are switched on at all.
baselineWhat this session is worth if we do nothing. Silence is the incumbent, and it does not have to justify itself.

ONE SESSION, FIVE CANDIDATES

Commitment friction · confidence 78% · plans page · €1,240 margin at stake

L0 · Silence

The incumbent. No trust spent, no cost.

€0.00

L1 · Passive adaptation

Term comparison appears in page flow.

+€2.10

L2 · Micro-nudge

One corner line about switching terms later.

+€3.40

L3 · One question CHOSEN

“How long do you expect to stay at this address?”

+€7.80

L4 · AI adviser

Above the founding ceiling. Trust cost exceeds the gain.

−€1.20

Then the assignment gate runs: a stable hash sends about one in five eligible sessions to the holdout, where the chosen action is recorded but not rendered.

ILLUSTRATIVE DETERMINISTIC DATA · NO LIFT CLAIMED BEFORE EVIDENCE

THE ACTION SPACE

Seven candidates, ordered by what they cost the visitor.

The full methodology →

0

No action

A ranked candidate, the default, and the incumbent every other level must beat.

CHOSEN 91.1% OF THE TIME

1

Passive adaptation

The page quietly gains the missing information.

TRUST COST ·

2

Micro-nudge

One contextual line, dismissible forever.

TRUST COST · ·

3

One question

A single question whose answer selects the resolution.

FOUNDING CEILING

4

AI adviser

A conversation that resolves the blocker. It never sells.

OFF IN PILOTS

5

Agent action

Checks a date, saves a quote, holds a slot.

OFF IN PILOTS

6

Human handoff

Where a person is cheaper than a lost sale.

HUMAN, DECLARED

A bounded action space is what makes the ranking auditable. Seven named levels, each with the friction types it may address, the copy it may not use, and a trust cost that has to be earned by evidence.

MEASUREMENT

How to measure the incremental value of next best action.

Three failure modes account for most overstated NBA results. All three are structural, and all three are fixable before launch rather than after the board asks.

The measurement maths →

01

Counting the willing

Attributing every conversion that followed an action to that action.

Fix: randomise after the eligibility gates, so the holdout contains sessions that would have been treated.

02

The vanishing control

A holdout that exists during the pilot and disappears at rollout.

Fix: a permanent holdout with a server-side floor. Re-randomising requires two approvals and is logged.

03

The invisible cost

Reporting response rate while dismissals and discounts sit elsewhere.

Fix: one screen. Lift, cost of speaking and margin given away, together, or none of them.

04

And the useful test

Does the system ever conclude that an action lost money?

Ours retires an arm after 14 days below holdout, and the decision log keeps the record of it having been tried.

FREQUENTLY ASKED

Next best action, briefly.

What is next best action?

A decisioning pattern that picks the single action most likely to produce a desired outcome for a specific person right now, from a defined set of possible actions.

What is the difference between next best action and next best offer?

Next best offer chooses a product or price. Next best action chooses any intervention, including non-commercial ones. Neither traditionally includes doing nothing as a scored option.

Can “no action” really be the next best action?

On considered purchases it usually is. Most sessions contain no resolvable friction, and interrupting a buyer mid-deliberation has a real cost. On illustrative pilot traffic silence is chosen for 91.1% of sessions.

Is next best action the same as real-time decisioning?

Real-time decisioning is the capability — deciding within the session. Next best action is the pattern the decision follows. Yesward is both, applied narrowly to considered-purchase journeys.

Do I need a CDP first?

No. Decisions are made from in-session behaviour, so nothing depends on a unified profile. If you have a CDP, margin and outcome data from it make the economics sharper.

Where does the action get delivered?

In your own page, in your own brand tokens, with an AI mark and one dismissal that holds for the session. The engine decides; your site renders. See the seven surfaces.

Find out how often your best next action is nothing at all.

Shadow mode ranks every candidate for two weeks and renders none of them.

Review one leaking journey