AI DECISIONING · REAL-TIME CONVERSION OPTIMISATION
AI decisioning for considered purchases.
Yesward detects why buyers hesitate, decides whether intervention is worthwhile, chooses the smallest useful response, and measures the incremental gross profit it creates.
Real-time AI conversion optimisation for high-consideration digital journeys.
SIGNALS
HYPOTHESIS
DECISION
L3 · one question, then get out of the way
How long do you expect to stay at this address?
RECEIPT WRITTEN EITHER WAY · MOST SESSIONS: THIS PANEL DECIDES TO DO NOTHING.
THE CATEGORY
What AI decisioning means here.
Yesward is AI decisioning software for considered purchases. It reads behavioural signals in a live session, forms a scored hypothesis about which friction is blocking the buyer, decides whether an intervention is economically worthwhile at all, chooses the smallest response likely to resolve it, and measures the incremental gross profit created against a permanent holdout.
A considered purchase is a purchase people research before they commit — roughly the €300–€3,000 band, where the journey runs comparison → configuration → price → checkout or application, and where hesitation is normal rather than exceptional.
“Some visitors need help. Most need you to get out of the way.”
THE POSITION, STILL. NOW THE SECOND THING WE SAY — NOT THE FIRST.
SEARCH CATEGORY
AI conversion optimisation · real-time personalisation
The language buyers already type. It exists to acquire the visit.
PRODUCT CATEGORY
AI decisioning
An established category with an established buyer. It exists to explain the shape of the product.
OWNED POSITION
AI decisioning for considered purchases
Narrow enough to win, large enough to matter. It exists to be remembered.
DIFFERENTIATION
Friction diagnosis + intervention restraint + profit optimisation + causal measurement
No single ingredient is the moat. The combination — and the discipline it enforces — is.
Deeper: AI decisioning · considered purchases · the Intervention Ladder
THE EARLIER QUESTION
Everyone else asks which. We ask whether.
A PERSONALISATION PLATFORM ASKS
Which experience should we show this visitor?
Assumes an experience should be shown.
A NEXT-BEST-ACTION SYSTEM ASKS
Which action should we take?
Assumes an action should be taken.
YESWARD ASKS, IN THIS ORDER
Should we take an action at all?
And if yes — what is the smallest intervention likely to beat silence economically?
Doing nothing is a treatment. It gets randomised, recorded and measured like every other one. Why no action is an action →
THE PROBLEM
Your funnel can show where buyers stop. It cannot tell you whether speaking will help.
Blanket prompts optimise visibility. Chat optimises conversation. Experimentation optimises variants. Yesward optimises the governed decision to intervene — and the incremental economic value created when it does.
€300–3,000
Telecom first, then adjacent
High-consideration service journeys — subscription and configuration purchases where friction is costly and intent is observable.
13 frictions
The blocker is knowable
Term anxiety, unclear total cost, install uncertainty, compatibility doubt. Answerable questions, left unanswered.
1 contract
Restraint is affordable here
One recovered contract pays for months of platform. That is what buys the discipline to stay silent.
THE MECHANISM
Five steps, and one of them is a receipt.
signal → hypothesis → decision → response → receipt
01 · OBSERVE
Signals, not identity
Dwell, toggles, comparisons, FAQ topics, stalls.
PRIVACY-SAFE EVENT STREAM
02 · INFER
A hypothesis, with evidence
Provisional, scored, never a fact.
03 · DECIDE
Evidence, eligibility, trust cost, economics
04 · ACT
The smallest permitted response
In the customer's brand, dismissible once and forever.
TREATED · 11.1% OF SESSIONS
HELD OUT OR LEFT ALONE · 88.9%
Silence is the other branch, not a failure
Weak evidence means nothing renders. A permanent 20% share of eligible sessions is withheld on purpose, so the comparison stays honest.
05 · PROVE
Both branches write a receipt
What was seen, what was decided, which arm it landed in, and what happened next.
THE LADDER
Seven levels. The first one is doing nothing.
Escalation costs trust, so it is earned by evidence. Every surface is dismissible once and forever, and never references the visitor's behaviour back at them.
0
Silence
The default outcome. No signal strong enough, no blocker the site can fix — nothing renders.
1
Passive adaptation
The page quietly gains the missing information.
2.5%
2
Micro-nudge
One contextual line, in the corner, dismissible forever.
2.1%
3
One question
A single question whose answer selects the resolution.
1.6%
4
AI adviser
A conversation that resolves the blocker. It never sells.
1.7%
5
Agent action
Checks a date, saves a quote, holds a slot — declared actions only.
0.7%
6
Human handoff
For deal sizes where a person is cheaper than a lost sale.
0.3%
TO SCALE — the same seven levels drawn honestly. Restraint is not a footnote; it is the distribution.
Founding pilots are permitted silence through one question. Levels 4 to 6 are built and inspectable, and stay switched off until low-trust-cost responses are proven safe and valuable.
REASONS TO BELIEVE
Restraint is the product, not a limitation.
Five claims, each tied to the product evidence that has to exist before we make it.
01
Decide before engaging
Intervention is earned, not automatic.
eligibility gates · confidence · suppression · recorded decision
02
Minimum effective help
The least intrusive response likely to restore progress.
a bounded ladder, from silence to one question
03
Economic optimisation
Profitable progress — not clicks or conversation volume.
margin-adjusted reporting · explicit response cost
04
Built-in incrementality
Every live policy preserves a counterfactual.
stable holdout assignment · receipts · attribution
05
Visitor safeguards
Commercial value and trust are evaluated together.
dismissal suppression · frustration guard · global stop
PROOF
Built so the number survives the CFO.
Every eligible session is randomised before treatment — a share is always withheld. What we report is the difference between the two groups, in contribution margin.
ELIGIBLE POPULATION · 14 DAYS
95% CI €6.8K–€15.7K · ROI 5.9× ON TOTAL COST · ILLUSTRATIVE DATA
GUARDRAILS
Copy rules live in code, not in a prompt.
“I noticed you've looked at pricing three times…”
Behaviour is never referenced back.
The engine acts on what it sees; it never tells the visitor it was watching.
“Hi, I'm Sarah from customer success 👋”
Always self-identifies as AI.
EU AI Act Art. 50 by construction: every surface carries the AI mark. Humans are only claimed at Level 6, where they're real.
“Only 2 slots left at this price!”
No manufactured urgency or scarcity.
Forbidden vocabularies are enforced at render time. A treatment that can't say it, won't.
“Here's 20% off if you buy right now.”
Discounts only by explicit policy.
Off by default. Price friction is resolved with information first; margin is a last resort you switch on.
THE PILOT
Six-week Shadow-to-Signal Pilot.
One journey, one outcome, one decision at the end. It begins silently, contains the risk, and closes with a go/stop recommendation grounded in evidence.
WEEK 0 · CONTRACT
Scope and verify
Select one journey and one outcome. Verify consent, the data contract and the economic metric. One SDK snippet, minimal payloads, fail-open.
DECISION: IS IT MEASURABLE AND SAFE TO TEST?
WEEKS 1–2 · SHADOW
Decide, render nothing
Signals, would-be decisions, suppression and data quality are all recorded while the site stays exactly as it is.
DECISION: ARE THE ELIGIBLE MOMENTS CREDIBLE?
WEEKS 3–6 · CONTROLLED
Approved responses only
Low-trust-cost responses go live behind a permanent 20% holdout, capped at one question.
DECISION: VALUE INSIDE THE GUARDRAILS?
CLOSEOUT · SIGNAL
Go, revise, or stop
The incremental outcome report, its data caveats, the visitor-response record, and a recommendation you can act on either way.
DECISION: CONTINUE, REVISE, OR STOP?
Fixed fee, quoted after qualification and credited to your first annual subscription if you continue. Stopping is a supported outcome — if the evidence is weak or visitor response deteriorates, stopping is the default.
WHERE TO GO NEXT
Start from the words you already use.
Six entry points into the same argument. Whichever vocabulary you arrived with, the page you land on names it plainly before it names us.
/product/ai-decisioning/
AI decisioning
What a decisioning engine is, what it must be able to prove, and the decision we add before all the others.
/product/ai-conversion-optimisation/
AI conversion rate optimisation
Three generations of CRO, why more interventions stopped working, and what to ask a vendor.
/product/real-time-personalisation/
Real-time personalisation
Session signals instead of stored profiles — anonymous visitors, no cross-site identifier, EU-first.
/decisioning/next-best-action/
Next best action
Ranking actions by expected incremental profit — with “no action” in the ranking, not outside it.
/considered-purchases/
Considered purchases
The pillar: what makes a purchase considered, the five friction families, and the four verticals we start in.
/compare/
Compare
AdaptiveCX, Wisepops, Pathmonk and Optimizely — including where each is the better choice.
Find the leakage before changing the journey.
Start with two weeks in shadow mode. Yesward records what it would have done while your website remains unchanged.