Partnerfy Partnerfy
Türkçe English Deutsch
AI Marketing Automation

One campaign.
100,000 versions.
One per customer.

Old marketing automation was lists of "who opened, who didn't". The new one: a different subject line, a different image, a different send time, a different CTA per customer — plus a model that predicts what will happen in 30 days. The Partnerfy team installs the GPT-5 / Claude / Gemini layer into your CRM and delivers that leap in 6 weeks.

We don't sell "AI marketing"; we add intelligence to the stack you already have. Klaviyo, HubSpot, Iterable, Salesforce — whichever ESP you run, we sit on top.

AI readiness audit 6-week plan
Customer profile
EK

Elif K.

İstanbul · 24 mo. member

Last 5 actions

Page view14:02
Add to cart14:04
Compare14:07
Read reviews14:09
Cart abandon14:11

Purchase intent

0.84

Active segment

Comparator High intent Price-sensitive VIP candidate
AI
3 AI variants

Token stream

item_waiting_in_your_cart...
Purchase 0.84
Churn risk 0.12
Best channel E-mail
+34% CTR

Static segmentation is dead

"High-spending women 25-34" segments rot in 3 months.

The biggest fallacy of classic automation: the customer is static. But customers change hourly. Someone who compared this morning buys with a coupon by afternoon; in the evening, opens a friend's recommendation and goes to your competitor for 30 days. A static list walks 4 weeks behind behaviour. AI recomputes on every interaction.

01

Every customer is unique

1,000 customers = 1,000 behaviour curves. Squeezing them into "5 segments" throws away 95% of the information.

02

Behaviour shifts hourly

A customer is a different person in the morning, at noon, and in the evening. Context shifts, intent shifts, the right message shifts. AI captures it live.

03

Manual segments rot

Segments hand-built by marketing decay within 90 days; new ones form but no one notices. AI keeps refreshing.

04

Subject lines are guesswork

"Will this subject get clicks?" — humans guess; AI tests 50 variants and tells you. Open rate differs by 20-40%.

05

Best send time is hidden

Sending Tuesday 10:00 to the whole list is statistical convenience; in reality every customer has their own open hour.

06

Churn caught too late

Customers signal 30-60 days before leaving. Static reports miss the signals; AI alerts 6 weeks early.

AI core visualisation

Predict + generate + learn. Three layers of AI marketing.

Here is a real scenario: customer "Elif" arrives, behaviour is captured, the prediction model emits 4 metrics, the content engine writes three versions — and the system learns which one was opened to do better for the next customer. All in 200 ms.

Prediction panel

Four critical metrics per customer.

Probability of purchase 0.84
Churn risk (30 days) 0.12
Predicted CLV ₺3.840
Next best channel E-mail

200ms

Inference time

38

Active features

97%

Model AUC

Before AI vs after AI

Same campaign, same customer — two worlds.

Before

"Dear customer, don't miss our sale!"

Sent to entire list Tuesday 10:00. 3.2% open. 0 personal context.

Generic · Static · Lost
After

"Elif, we summarised reviews for the X in your cart."

Personal to Elif at 14:23 — synthesis of 38 signals. 22% open. 6.4% purchase.

Personal · Live · Winning

6.9×

Open rate

11×

Conversion

−47%

Unsubscribe

Who it's for

Anyone with customer count + data richness + communication volume.

01

E-commerce · 100k+ SKUs

Large-catalogue brands needing seasonal, cross-category, behaviour-driven recommendation engines.

02

SaaS · high churn

Companies with 3%+ monthly cancellation rate wanting expansion + churn prediction for healthy MRR.

03

Marketplaces

Two-sided dynamics; separate models for supply + demand sides, ranking + recommendation together.

04

Financial services

Eligibility + segment-based offers + risk scoring — personalised within the regulatory frame.

05

Publisher / media

Hundreds of content options per reader; an AI rec engine grows DAU 20-40%.

06

B2B long sales

6-18-month sales cycle; intent score + ranked account list + timing recommendations.

07

Agencies

Agencies wanting to sell AI-based services but lacking ML engineers — we operate as white-label.

08

Healthcare & clinics

Appointments, medication reminders, follow-up communication — personal channel, language and timing per patient.

10 capabilities, one panel

Prediction, generation, optimisation — under one operation.

Having each capability with a different vendor causes major fragmentation. With Partnerfy, 10 capabilities run on one backbone, share data, and feed one dashboard.

01

Predictive lead scoring

Model learning from past won / lost customers; returns 0-100 score + 3 reasons for every new lead.

02

Next-best-action engine

For each user, the decision "what should we do now": email? push? discount? phone? nothing?

03

AI-generated email copy

Model trained to your brand voice; 3 variants from a brief + approval flow + ongoing tone-drift control.

04

Dynamic subject lines

Same send, subject varies per recipient; AI picks based on the recipient's open history.

05

Send-time optimisation

Personal optimum hour for each user; personal time window instead of daily batch.

06

Churn prediction

List of customers leaving within 30/60/90 days + intervention suggestion; real-time alert.

07

Recommendation engines

Product, content, plan — collaborative + content-based hybrid; content-based fallback for cold-start.

08

AI segment discovery

Instead of manual "who is VIP"; clustering algorithms surface customer clusters.

09

Sentiment-driven workflows

Sentiment extracted from support messages; angry customers auto-escalate + route to humans.

10

AI multivariate testing

Bandit + Bayesian instead of classic A/B; 5-30 variants tested in parallel, winner dynamically distributed.

Process

From data audit to continuous learning: a 6-step path.

  1. 01

    1 · Data audit

    CRM, ESP, web, mobile, product, support — every data source mapped. Missing fields + unification strategy emerge. Usable feature list drafted.

  2. 02

    2 · Model selection

    Right model per task: GPT-5 for copy, gradient boosted trees for scoring, embedding models for search. We don't fall into the "one big model for everything" trap.

  3. 03

    3 · Training data prep

    12-24 months of historical data cleaned, labelled, train/test split. Data quality is 70% of the model.

  4. 04

    4 · Integration

    API layer on top of your ESP (Klaviyo, HubSpot, Iterable, Customer.io, Salesforce); two-way data flow. No disruption to existing flows.

  5. 05

    5 · Deploy + A/B test

    AI version runs alongside a control group for 30-60 days. Open, CTR, sales, churn — statistical lift validated.

  6. 06

    6 · Continuous learning loop

    Model retrained weekly; results piped to dashboard; drift alarms set; humans + AI co-review.

Tools we use

Foundation models + production-grade AI platforms.

OpenAI API Anthropic Claude Google Gemini Mutiny Persado Phrasee HubSpot AI Klaviyo AI Salesforce Einstein Iterable AI Segment Twilio BigQuery ML LangChain n8n

Client stories

Brands that moved from static to AI personalisation.

E-commerce +34% CTR

Lifestyle brand

AI subject lines + send-time optimisation: email CTR +34%; total channel revenue 2.1×.

B2B SaaS −19% churn

Vertical SaaS

Churn-prediction model + proactive intervention flow: 30-day churn −19%, MRR health restored.

Marketplace +48% ATC

Service marketplace

Hybrid recommendation engine: add-to-cart +48%, average order 1.8×.

Media +27% DAU

Publishing platform

AI content recommendation: DAU +27%, premium upgrade +14%.

Financial +54% cross-sell

Online broker

Next-best-action engine: avg products per customer 2.4 → 3.7; cross-sell +52%.

Education +24 pp NPS

Online academy

Personal learning path + timing: course completion 38% → 62%; NPS +18.

FAQ

The 8 questions most asked about AI marketing

When properly set up, the difference is measurable. Our clients average +15-35% open rate, +20-60% conversion; churn drops 15-25% with proactive intervention. But everything depends on "proper setup": without enough data, with bad labels, with the wrong model picked or no integration — zero benefit. AI isn't magic, it's an engineering discipline. Partnerfy's job is to ensure correct setup + prove the lift is real with A/B testing.
Customer data does not leave your systems. For foundation models (GPT-5, Claude, Gemini) we use "no-training" contracts — data we send is not used to train models. Sensitive fields (name, email, phone, payment) are hashed or tokenised. DPAs prepared under KVKK / GDPR / ISO 27001. For customers needing EU data residency we use EU-region APIs or local self-hosted models. Privacy is design-time, not bolt-on.
A different model per task. Claude is strong for long-form on-brand generation; Gemini is good for multilingual content; GPT-4o-mini is economical for very fast single-sentence generation. For scoring / ranking we don't use modern LLMs — gradient boosted trees or XGBoost are faster + cheaper + more explainable. Instead of "one model solves all", we pick the right tool per task; this is both cost and quality advantage.
Token usage is measured transparently with monthly caps. Typical distribution: light tasks (subject lines, segmentation) on small models ~$0.50-2 / 1k calls; medium tasks (email copy) ~$3-8 / 1k calls; heavy research tasks ~$15-40 / 1k calls. Typical token spend $300-3,000 / month. Plus platform subscriptions (Klaviyo AI, Salesforce Einstein etc.) where applicable. Budget alarms + cap enforcement always on.
For prediction-based models, 5,000+ active customers + 12 months of history give a satisfying signal. At 10,000+ the model becomes meaningful, at 50,000+ it is strong. AI also works for smaller lists — but on generation (copy, image, segmentation-help) side rather than prediction. For very small lists we set up classic automation properly first; we add the AI layer as data accumulates. Without data, AI is disappointment; we don't make empty promises.
Two-layer protection: first, model trained to brand voice + banned-word / claim / price filters; second, approval flow — high-value campaigns require human approval before send. Low-risk campaigns (transactional, simple reminders) flow directly; promotions or corporate claims go through approval queues. Plus a monthly tone-drift report: has the AI drifted from the brand's real voice over 30 days? If drift exists, fine-tune is refreshed. Full control stays with you.
Yes. Most modern ESPs — Klaviyo, HubSpot, Iterable, Customer.io, Salesforce Marketing Cloud, Braze, Mailchimp — expose open APIs. We place the AI layer in front of the ESP: the ESP asks AI "what should I send to which list", AI returns a prediction, ESP sends. Without disrupting your existing flows or reporting. For ESPs with native AI features (Klaviyo AI, Salesforce Einstein) we maximise what they offer; we fill gaps with custom models.
First real win (subject line + send-time optimisation) appears in 4-8 weeks in ESP reports: +10-25% open rate. Prediction-based models (churn, lead scoring) mature in 60-120 days. A typical client recovers project cost within 90 days, reports 2-4× ROI by 180 days, 5-10× ROI by 12 months. Slower in low-data segments, faster in high-data ones. We install measurement in the first 30 days; from there on it's clear monthly math.

Add a hundred thousand versions to one campaign.

A free 30-minute call to inspect your current marketing stack; we share the 3 highest-impact AI opportunities you can test in the first 60 days.

Book a call Contact us
results