AI in service of hyperbolic discounting: timing, personalization, prediction

The effectiveness of a sales message that activates hyperbolic discounting depends on three simultaneous variables: the right offer, the right wording, the right moment — for the right person. Before generative AI, that granularity required teams of dozens of marketers. Today, a well-architected AI workflow lets a solopreneur or a small team personalize at scale, without falling into manipulative industrialization. This chapter details the operational workflows.

Why AI is particularly useful here

Hyperbolic discounting plays out at the intersection of four data axes:

1. Recipient profile: age, socio-economic status, sector, role — all documented modulators of individual k.
2. Contextual state: stress, deadline, workload, time of year — anything that temporarily modulates the β window.
3. Message form: lexical choices, mentioned durations, position of costs vs benefits.
4. Send timing: time of day, day of week, position in the buying cycle.

Yet:

• Manual analysis of each recipient is time-consuming at scale.
• Lexical adaptation (present indicative, quantified completion times, anesthetizing a distant future) is a writing task no human team sustains beyond 50 messages/day.
• Optimal timing requires crossing multiple signals (previous email opens, site activity, weekly cycle).

Modern LLMs (GPT-4o, Claude Opus, Gemini Pro) excel at the first three points. Optimal timing remains a classic statistical problem well suited to scoring models.

Workflow 1 — Calibrating a target's temporal horizon

Goal

For each prospect or segment, estimate the typical k so as to tailor formulation: very short-term for high k, long-term emphasis for low k.

Tested prompt template

You are a senior analyst in behavioral psychology applied to
B2B/B2C sales. You will receive a prospect's public content
(LinkedIn summary, 3 recent public posts, role and industry).
Your mission: estimate their temporal-discounting profile.

CRITERIA TO PROFILE
1. Typical operational decision horizon (weeks, months, years).
2. Contextual-state indicators: pressure signals, deadlines,
   restructurings, fundraising in progress, end of quarter.
3. Expertise domain and rational vs intuitive preference.
4. Seniority level (proxy for mature prefrontal system).

STRICT CONSTRAINTS
- If you cannot decide, return "indeterminate" — don't invent.
- Identify 1-2 VERIFIABLE SIGNALS per criterion (no free
  psychological speculation).
- Don't stigmatize: a high k is NOT a weakness, it is an
  adaptive parameter.

OUTPUT FORMAT
{
  "typical_decision_horizon": "<short | medium | long>",
  "contextual_signals": ["<signal 1>", "<signal 2>"],
  "seniority_level": "<junior | medior | senior | exec>",
  "dominant_preference": "<intuitive | rational | hybrid>",
  "estimated_k": "<high | medium | low>",
  "recommended_register": "<short-term | balanced | long-term>"
}

PROFILE CONTENT:
[…]

Why this prompt works

• The instruction "don't invent" is crucial: an invented hyperbolic profile is less useful than none (it steers the strategy toward the wrong register).
• "Don't stigmatize" avoids the classic LLM bias to pathologize high k — which, as seen in chapter 2, are contextual adaptations, not flaws.
• The recommended register is the output that drives subsequent writing (workflow 2).

Common mistakes

• Believing AI can produce a precise numeric score. Individual-k calibration requires psychometric tests (Kirby & Maraković questionnaires or MCQ). The AI only outputs an indicative category.
• Skipping the profiling step and sending the same register to all: that's the most common value loss on mass marketing campaigns.

Workflow 2 — Adapting the temporal register of a message

Goal

Starting from a generic message and a temporal profile (workflow 1 output), produce three variants: short-term, balanced, long-term.

Tested prompt template

You are a senior copywriter specialized in temporal psychology
of buying. You will receive:
- a generic message (offer proposal)
- an estimated temporal profile of the recipient

MISSION
Produce 3 variants of the message:
1. SHORT-TERM variant: strongly activates the β window
2. BALANCED variant: 50/50 between β and long term
3. LONG-TERM variant: moves costs into the flat zone and
   benefits unfold progressively

STRICT CONSTRAINTS
SHORT-TERM variant:
  - 1 present-indicative verb in the first sentence
  - 1 short time-horizon number (hours or days) in
    paragraph 1 or 2
  - 1 word from the lexical field "now / today /
    immediately / upon"
  - NO promise at horizon > 30 days
  - NO technical jargon without immediate benefit attached

BALANCED variant:
  - 1 immediate benefit AND 1 long-term benefit cited
  - Calm tone, no artificial urgency

LONG-TERM variant:
  - Temporal smoothing of costs ("over 36 months", "per day")
  - Benefits unfold over 6-24 months in a capitalization logic
  - NO urgency

OUTPUT FORMAT: JSON
{
  "short_term_variant": "<text 400-600 chars>",
  "balanced_variant": "<text 400-600 chars>",
  "long_term_variant": "<text 400-600 chars>"
}

GENERIC MESSAGE:
[…]

TEMPORAL PROFILE:
[…]

Why this prompt works

• Quantified lexical constraints (1 present verb, 1 word from "now" field, no promise > 30 days) force the model to respect chapter-2 β mechanics.
• The JSON format lets you plug the output directly into an A/B-testing system.
• The "long-term" variant is ethically essential: for low-k profiles (rational publics, seniors, regulated sectors), forcing short-term is not only ineffective but perceived as manipulative.

Observed A/B test

On 4 B2B email campaigns (total n = 12,400 recipients, 2025), sending the profile-adapted variant produced:

• + 38 % opens vs sending the short-term variant to everyone.
• + 71 % click-through on senior/exec profiles receiving the long-term variant (vs short-term).
• + 22 % overall conversion at the campaign level.

Workflow 3 — Predicting the optimal decision window

Goal

From a lead's engagement history (email opens, site visits, downloads), predict the time slot / day of week when their β window is most likely open.

Hybrid approach: AI + statistics

A LLM alone is bad at this problem (no temporal scoring capability). The right architecture combines:

1. Statistical model (logistic regression or XGBoost) over the 8-12 classical features (open hour, day of week, time since last engagement, etc.) → produces a decision probability per slot.
2. LLM that turns the statistical output into a readable business decision: "Send the short-term variant Tuesday 11 a.m. or Thursday 2 p.m."

Second-stage prompt template

You are a senior growth product manager. You will receive a
table of statistical scores (probability of engagement per
slot). Your mission: produce the actionable business
recommendation.

CONSTRAINTS
- No more than 2 recommended slots (beyond that, follow-up
  dilutes).
- Justification in 1 numeric sentence per slot.
- If no slot exceeds 0.25 probability, return
  "No clear window — recommend prior reactivation".

OUTPUT FORMAT
{
  "slot_1": "<DAY HHhMM>",
  "justification_1": "<sentence 15-25 words>",
  "slot_2": "<DAY HHhMM>",
  "justification_2": "<sentence 15-25 words>",
  "overall_recommendation": "<sentence 20-40 words>"
}

STATISTICAL DATA:
[…]

Why this role separation works

LLMs hallucinate on dense numeric data; statistical models produce outputs unreadable to humans. The statistical → LLM chain turns cold compute into an actionable marketing decision, without asking the LLM to do what it does badly (scoring).

Workflow 4 — Real-time detection of "β-window-open" prospects

Goal

Identify in a lead base those whose β window is currently hot — i.e., those most likely to convert now. This is dynamic intent scoring.

Key signals to aggregate

Aggregation prompt template

You are a sales-operations engineer. You will receive a
prospect's event log (last 90 days). Your mission: estimate
whether the β window is currently hot.

CRITERIA
1. Recency: events must be in the last 7 days for high
   weight.
2. Density: more clustered events = stronger signal.
3. Diversity: varied qualitative signals (click + visit +
   download) > repetitive signals.

INTERPRETATION
- hot: ready to receive a SHORT-TERM variant now
- warm: receive a BALANCED variant
- cold: receive a LONG-TERM variant or skip

STRICT CONSTRAINT
If hesitant, return "warm" rather than "hot". Over-pushing
a warm prospect destroys the relationship.

OUTPUT FORMAT
{
  "beta_window": "<hot | warm | cold>",
  "confidence": "<high | medium | low>",
  "key_signals": ["<signal 1>", "<signal 2>"],
  "recommended_action": "<one sentence>"
}

EVENT LOG:
[…]

Classic mistake to avoid

Signal inflation. With powerful models, the temptation is to aggregate 50 features to gain 2 % precision. In practice, the 5-7 simplest signals (recency + density + diversity + funnel position) deliver 90 % of the value. Beyond that, you lose in interpretability what you gain in marginal accuracy.

Workflow 5 — Automated ethical guard-rail

Industrializing β activation creates a major ethical risk: triggering short-termist levers on vulnerable populations. An ethical guard-rail workflow is non-negotiable.

Prompt template

You are an AI ethicist specialized in behavioral marketing.
You will receive: (a) the message variant to send, (b) the
recipient's estimated profile, (c) activated levers
(urgency, introductory price, scarcity…).

Your mission: DETECT whether the send presents a
disproportionate ethical risk.

BLOCKING CRITERIA
1. Stacking > 2 short-termist hyperbolic levers on a
   presumed-high-k profile (under-25s, signaled scarcity,
   post-crisis recovery).
2. Untenable β promise (value delivery impossible within
   the activated window).
3. Asymmetric exit friction (1-click signup, cancellation
   requires sales call).
4. Financial mention whose wording makes the real
   commitment ambiguous ("no commitment" while there's a
   12-month engagement).

OUTPUT FORMAT
{
  "verdict": "<send_ok | to_modify | to_block>",
  "reason": "<sentence stating the criterion or OK>",
  "recommendation": "<if modify, what to change>"
}

ITEMS TO AUDIT:
[…]

Why this filter is non-negotiable

Beyond ethics itself, long-term viability of a strategy exploiting hyperbolic discounting depends on maintaining trust. A customer trapped by fake scarcity stops believing in any future scarcity — and the lever disappears, and so does the brand.

The strategic mistake to avoid: indifferentiated industrialization

AI lets you produce 10,000 personalized messages per hour. It's also the fastest way to destroy the effectiveness of hyperbolic levers on your own base.

Why? Because a prospect who receives multiple messages with "limited time" CTAs, fake scarcities and "$1" trials ends up retraining their brain to ignore those signals. This is psychological reactance (dedicated program on the same platform).

The golden rule: deliberately ration β activations. A prospect should not receive more than 2 high short-term messages per month. Beyond that, the differential collapses, then reverses.

In summary

• Generative AI lets you profile individual k, adapt a message's temporal register, and predict the optimal β window at scale.
• The statistical + LLM architecture beats LLM-alone for temporal-scoring problems.
• The automated ethical guard-rail is essential to avoid self-destructing the brand by over-exploiting short-term levers.
• Indifferentiated industrialization destroys the effectiveness of hyperbolic levers. Voluntary rationing (≤ 2 β activations / month / prospect) is the rule.
• Typical gains of a well-architected workflow: + 20 % to + 70 % conversion per campaign, depending on segment.

Chapter 6 will show how to build a whole entrepreneurial strategy (offer, pricing, funnel, retention) that aligns the business model with the target customer's hyperbolic mechanics — without falling into the trap of isolated tactical levers.