Customer referrals

Satisfied customers exist. Referrals happen by chance.

Make a good experience easy to recommend. Test the path from a relevant invitation to a new customer’s first purchase.

Challenge and test scope

These are test directions, not growth promises. Agree one primary metric, costs and decision criteria for the chosen challenge.

Who it is for
For a business with satisfied customers whose referrals are incidental or lost. Identify confirmed value and make the offer easy to explain to someone else.
Data needed
Pseudonymous referrer ID, confirmed value event, contact permission, invitation date, referral ID, new customer ID, source, enquiry, payment, refund, reward and costs.
AI and people
AI drafts an invitation and a short message the customer may forward voluntarily. Staff check relevance. The new customer contacts the business themselves; the team records source and outcome.
What the company keeps
Trigger map, messages, new-customer journey and attribution rules. Define reward conditions and duplicate protection in advance.

Three AI scenarios. People remain accountable.

Value moments

AI prepares
Permitted feedback → possible invitation moments with a quote or source ID
A person checks
Confirms the actual benefit and that no problem remains unresolved

Referral invitation · starting point

AI prepares
Confirmed benefit + offer → a personal draft and a message to forward
A person checks
Checks suitability and sends the invitation

Referral records assistant

AI prepares
Source records → disputed matches, duplicates and reconciliation questions
A person checks
Confirms attribution and entitlement to a reward

AI must not invent prices or terms, send important messages without agreed oversight, change agreements or execute payments.

What to measure and compare

01

New referral purchases

Confirmed first purchases by new customers with a permitted referral source; also report per member of the initial referrer cohort.

02

Referral funnel

Invitation → voluntary referral → new enquiry → first payment.

03

Contribution and cost

New-purchase contribution minus rewards, handling, services and additional costs.

04

Guardrail measures

Complaints, duplicates, self-referrals, cancellations, returns and unsupported rewards.

From GRI findings to a test

GRI is an author-developed analytical framework, not independent certification, an investment rating or a growth guarantee.

  1. GRI

    Understand the constraint

    Find an observable loss and test possible explanations against the facts.

  2. GTM

    Connect the change to the customer

    Define the segment, value, offer, channel, funnel step, economics and accountability.

  3. AI

    Build an assistant for the task

    Specify the inputs, expected output, limits and employee’s role.

  4. TEST

    Build a basis for a decision

    Agree a metric and comparison method, check quality, include costs and decide the next step.

Explore the methodology

The first 21-day cycle

The shared Bootcamp structure. Data, actions and the metric are tailored to the chosen lever.

00

Before the start

Discuss the challenge, available examples, team involvement and scope. Agree the work before the programme begins.

  1. Days 1–701

    Choose and design

    Review the workflow and data, choose a constraint and one lever, define the GTM action and experiment rules.

    Constraint map · GTM on one page · experiment brief
    Day-by-day calendar
    1. Day 1

      Session: map the customer journey and locate the loss

      Process map and challenge statement
    2. Day 2

      Collect examples; check IDs and dates

      Data dictionary and missing-data list
    3. Day 3

      Review relevant GRI areas and test two or three explanations

      Evidence-backed constraint map
    4. Day 4

      Session: choose one lever, segment and GTM action

      GTM on one page
    5. Day 5

      Describe the usual process, a non-AI improvement and an AI variant

      Technology role and comparison variants
    6. Day 6

      Define cohort, metric, observation window, costs and stopping rules

      Experiment brief v1
    7. Day 7

      Checkpoint: agree the solution scope

      Build / revise / prepare data
  2. Days 8–1402

    Build and test

    Prepare context, build the AI workflow, test typical cases and exceptions, and define human handover rules.

    Tested workflow version · error log · limited-test approval
    Day-by-day calendar
    1. Day 8

      Session: create a context base and first prompt

      AI scenario v0.1
    2. Day 9

      Run typical cases and verify source facts

      Error and correction log
    3. Day 10

      Test missing data, conflicts, duplicates and refusal

      Exception set and human handover rules
    4. Day 11

      Session: independent quality review on held-out examples

      Readiness assessment and critical defects
    5. Day 12

      Run in shadow mode without contacting customers

      Time measurement and staff usefulness review
    6. Day 13

      Prepare the log, group assignment and instructions

      Launch pack with prompt version
    7. Day 14

      Checkpoint and owner decision

      Limited-test approval or correction plan
  3. Days 15–2103

    Observe and decide

    Check execution, compare available outcomes, account for costs and limitations, and prepare team handover.

    Report · instructions · next-step decision
    Day-by-day calendar
    1. Day 15

      Session: check the first live events and stopping procedure

      Working limited test
    2. Day 16

      Review deviations and protocol adherence

      Event and error log
    3. Day 17

      Reconcile customers, orders, payments, costs and time

      Observed-results table
    4. Day 18

      Review group comparability and alternative explanations

      Limitations of the conclusion
    5. Day 19

      Session: calculate measures and economics

      Draft report and decision
    6. Day 20

      Prepare instructions, materials and continuation plan

      Handover pack
    7. Day 21

      Review: continue, change, stop or extend

      Approved report and 30/60/90 plan

A long sales or repeat-purchase cycle may require more than 21 days to assess a monetary effect. In that case, the result is tested execution and a plan for continued observation.

Team involvement · proposed format

Proposed format: six 90-minute working sessions and three 30-minute checkpoints. Indicative time for the experiment owner: 7–9 hours per week plus 2–3 hours of preparation. This is not an approved package; scope and involvement are agreed separately.

What the team keeps

Trigger map, messages, new-customer journey and attribution rules. Define reward conditions and duplicate protection in advance.

  1. 01Constraint map
  2. 02GTM plan
  3. 03Experiment brief
  4. 04AI workflow or prototype within the agreed scope
  5. 05Log and calculations
  6. 06Staff instructions
  7. 07Final decision and 30/60/90 plan

A 90-day plan is a document, not 90 days of included implementation and support. Production integration, data migration, licences and ongoing support require separate agreement.

Prompts, research, cases and books

Research and cases inform hypotheses; they do not guarantee the same result in your company. External companies are not Bootcamp clients or partners. Not every case concerns generative AI.

Materials: 10 / 10

AI prompt05

Find a confirmed value moment

RAFT working template · English Bootcamp adaptationRead material

Using [confirmed value events], [feedback] and [permission], identify an appropriate moment to invite a referral.

R — You are a careful business workflow analyst for this specific task.

A — Using [confirmed value events], [feedback] and [permission], identify an appropriate moment to invite a referral. Do not turn neutral feedback into praise or request other people’s contacts without consent.

F — Evidence and source IDs; observations; missing facts; draft action; human review checklist; metric, costs and limits.

T — Clear, specific and without growth promises.

Before starting, ask the clarifying questions needed to understand the context, objective and limits. Do not proceed until the inputs and rules are clear.

Limits of the findingUse permitted or anonymised data. A person verifies facts and approves consequential actions.

Bootcamp practiceReplace bracketed inputs with approved context. Test typical cases, exceptions and stopping rules before operational use.

AI prompt05

Prepare a referral invitation

RAFT working template · English Bootcamp adaptationRead material

Using [approved offer] and [referral rules], draft an invitation and an optional short forwarding message.

R — You are a careful business workflow analyst for this specific task.

A — Using [approved offer] and [referral rules], draft an invitation and an optional short forwarding message. Do not impersonate a customer or invent a review. The new customer must choose to make contact.

F — Evidence and source IDs; observations; missing facts; draft action; human review checklist; metric, costs and limits.

T — Clear, specific and without growth promises.

Before starting, ask the clarifying questions needed to understand the context, objective and limits. Do not proceed until the inputs and rules are clear.

Limits of the findingUse permitted or anonymised data. A person verifies facts and approves consequential actions.

Bootcamp practiceReplace bracketed inputs with approved context. Test typical cases, exceptions and stopping rules before operational use.

AI prompt05

Reconcile referral outcomes

RAFT working template · English Bootcamp adaptationRead material

Using [referral IDs], [new customers], [payments], [returns], [rewards] and [costs], reconcile first purchases.

R — You are a careful business workflow analyst for this specific task.

A — Using [referral IDs], [new customers], [payments], [returns], [rewards] and [costs], reconcile first purchases. Identify duplicates, self-referrals and unconfirmed sources. Keep forwarded links separate from paid outcomes.

F — Evidence and source IDs; observations; missing facts; draft action; human review checklist; metric, costs and limits.

T — Clear, specific and without growth promises.

Before starting, ask the clarifying questions needed to understand the context, objective and limits. Do not proceed until the inputs and rules are clear.

Limits of the findingUse permitted or anonymised data. A person verifies facts and approves consequential actions.

Bootcamp practiceReplace bracketed inputs with approved context. Test typical cases, exceptions and stopping rules before operational use.

Research05

Referred customers can differ in value

[R09] Schmitt, P.; Skiera, B.; Van den Bulte, C. (2011). Referral Programs and Customer Value.Read material

A study followed roughly 10,000 German bank customers for almost three years; referred customers had at least 16% higher average value than matched customers, with segment differences.

Limits of the findingThis is a context-specific comparison, not random channel assignment or a universal 16% improvement. Referrals are not necessarily free.

Bootcamp practiceTrack confirmed first purchases, contribution, rewards and handling costs.

Original source
Research05

Sharing content and buying have different causes

[R10] Berger, J.; Milkman, K. L. (2012). What Makes Online Content Viral?Read material

Berger and Milkman study New York Times content and experiments, linking sharing to activation and other content features.

Limits of the findingSharing does not establish payment, trust or high-quality demand. Emotional pressure is not proposed as a practice.

Bootcamp practiceGive customers useful, truthful explanations and measure purchases separately from views and forwards.

Original source
International case05

Airbnb · convenient, two-sided referrals

[C09] Bosinoff, J. / Airbnb Engineering (27.03.2014). Hacking Word-of-Mouth: Making Referrals Work for Airbnb.Read material

Airbnb Engineering’s March 2014 account describes redesigning referrals for web and mobile, with rewards after a referred guest’s first trip.

Limits of the findingHistorical company self-report about that programme, not all Airbnb bookings or generative AI.

Bootcamp practiceLink rewards to a confirmed event. A sent link is not a completed referral purchase.

Original source
International case05

Dropbox · a product-linked reward

[C10] Dropbox Help (обновлено 30.09.2026). How to refer friends to Dropbox.Read material

Dropbox Help describes storage rewards for referrals, with plan-specific limits and qualifying actions.

Limits of the findingThe help page establishes mechanics, not a causal contribution to business growth.

Bootcamp practiceChoose a useful reward, cost it and check new-customer validity, duplicates and abuse.

Original source
Quote05

We all want to be liked.

Jonah Berger · Contagious · publisher excerptRead material

A discussion excerpt for the referral route. Use the source’s own context, not as a promise of programme results.

Limits of the findingAn expert quote is not causal evidence or a universal business rule.

Bootcamp practiceUse the question to sharpen your own hypothesis, metric and decision criteria.

We all want to be liked.
Original source
Book05

Jonah Berger · Contagious: Why Things Catch On

[B09] Berger, J. (2013). Contagious: Why Things Catch On. Simon & Schuster.Read material

Understand why people share information. Prepare useful messages without fabricating experience or emotion.

Limits of the findingThe AI exercise is an original Bootcamp adaptation, not content attributed to the book. The programme uses publisher descriptions and open excerpts, not a claimed analysis of every full text.

Bootcamp practiceUnderstand why people share information. Prepare useful messages without fabricating experience or emotion.

Original source
Book05

Fred Reichheld, Rob Markey · The Ultimate Question 2.0

[B10] Reichheld, F.; Markey, R. (2011). The Ultimate Question 2.0: How Net Promoter Companies Thrive in a Customer-Driven World.Read material

Use customer feedback to improve service. Willingness to recommend and NPS are not actual referral purchases.

Limits of the findingThe AI exercise is an original Bootcamp adaptation, not content attributed to the book. The programme uses publisher descriptions and open excerpts, not a claimed analysis of every full text.

Bootcamp practiceUse customer feedback to improve service. Willingness to recommend and NPS are not actual referral purchases.

Original source

Working templates

Selected from AIStart360 Growth Labs, edition 2.0, 26 October 2026. Each item retains its provenance, limitations and links to original publications.

Articles and practical guides

3

How to validate demand before building

Start with recent customer situations, existing alternatives and the cost of doing nothing. Then test a concrete offer through an application, pilot or another meaningful commitment. Positive opinions about an idea are not demand.

More from the library

External research and cases inform hypotheses; they do not promise the same outcome. Check current conditions and relevance to the specific challenge.

Books

Competing Against Luck

Clayton Christensen et al.

Frames demand around the progress customers hire a solution to make.

Books

Continuous Discovery Habits

Teresa Torres

Connects desired outcomes, opportunities, solutions and assumption tests in a weekly cadence.

Books

Product-Led Growth

Wes Bush

A playbook for self-serve value, time-to-value and product-led acquisition.

Books

Obviously Awesome

April Dunford

A step-by-step method for positioning around competitive alternatives, differentiated value and the customers who care most.

Articles

The Racecar Growth Framework

Brian Balfour

Separates growth loops from one-off tactics and balances acquisition, retention and monetization around a durable engine.

Full library

Related leadership decisions

The challenge first. Then the terms.

Check whether the challenge fits, and discuss examples and team involvement. Then agree location, language, format, scope, fees and participation terms.

Design a referral journey

Before we speak

Do we need a technical specialist?

A manual workflow usually needs a process owner and permitted data. Integration with operational systems requires technical participation and a separately agreed scope.

Can we cover all six levers?

A company works on one lever per cycle. The others inform the next step; they are not six parallel projects.

What if we have little data?

Establish minimum record-keeping and test on permitted examples first. Never replace missing facts with invented data.

What if the hypothesis is not supported?

Document quality, costs and limitations. A justified stop helps avoid expanding an uneconomic solution.

Are software and implementation included?

Licences, integration development and ongoing support are included only when explicitly agreed in the scope.

When should we expect sales?

It depends on the workflow and customer cycle. The 21 days cover a limited scenario; long-term effects are assessed separately.

How is data used?

Agree permitted materials, anonymisation, access and external services before starting. The first-contact form does not need customer databases or documents.

Discuss your challenge