Marketing automation + AI – future of online sales

Marketing automation (MA) organises the “when and to whom” of the message you send, while AI refines the “what and why”, which means online sales can grow without adding more hours of manual work for the team.

In e-commerce, you automate contact across channels such as email, SMS, push and ads, while you use AI for prediction, recommendations, and content generation and variation. It is worth remembering that AI does not replace automation, but strengthens it, e.g. by predicting the optimal send time and matching the offer to a specific person. For this to make operational sense, you need consistent data and clearly defined KPIs, because “more communication” does not always translate into “more sales”. In the next sections, I show where combining MA+AI usually delivers the fastest results and how to approach it in practice.

Fundamentals of marketing automation and AI in online sales

Marketing automation and AI in online sales is a combination in which MA manages the timing and audience of communication, while AI uses data to select the most relevant content, offer and rationale.

In practice, rule-based automations (e.g. “if abandoned cart → send email after 2h”) always operate according to the same scheme and do not learn from results. ML models can match the time window and trigger (e.g. discount vs free delivery) to the probability of purchase, so communication pressure reaches only the right segments. If you are wondering whether AI replaces MA: no, AI increases the effectiveness of automation, but it does not remove the need to use it.

A modern setup is based on channel orchestration, where email/SMS/push do not compete with advertising, but split tasks between them across the funnel. For example, a user with high LTV may receive priority in SMS, while for a person with a low propensity to buy you activate cheaper push and retargeting in Meta/Google. At the same time, some decisions should work in real time (e.g. on-site recommendations, a personalised banner or a chatbot), while others can be calculated in batches (e.g. churn/LTV predictions once a day in BigQuery and sent to the MAP as segmentation attributes).

The most common implementation stack is store (Shopify, WooCommerce, Magento) + MAP platform (e.g. Klaviyo, Braze, HubSpot, Salesforce Marketing Cloud, ActiveCampaign), and at larger scale also a CDP and data warehouse (Segment/mParticle + BigQuery/Snowflake). For SMBs, a MAP and correct tracking are often enough, while a CDP and warehouse come into play when the number of sources grows and you need a unified customer profile. In planning, also account for real resources: basic automations (welcome, cart, post-purchase) can be implemented in 2–6 weeks, while predictive AI usually requires 6–12 weeks for data, testing and iterations.

LLMs (e.g. GPT-4.1, Claude, Gemini) support copy variant creation and consultants’ work, but without oversight they can hallucinate product information. In regulated products (e.g. supplements, medical) “guardrails” are essential: a knowledge base (RAG), a list of approved claims and content moderation. It is equally important to avoid the trap of “more flows = more sales”, because too many messages usually end with list fatigue and worse deliverability. When the number of unsubscribes and spam complaints rises, it is a sign to reduce pressure, introduce caps (e.g. 3 messages/7 days) and choose segments more precisely.

1. 01MA & AI synergyMA manages, AI personalises content
2. 02Rules vs learningStatic schemes versus adaptive matching
3. 03Intelligent pressureTriggers only to the right segments
4. 04Channel orchestrationChannels collaborate, not compete

AI increases the effectiveness of automation, it does not remove the need to use it.

Most common uses of AI and automation in e-commerce

The most common uses of AI and automation in e-commerce are scenarios that translate directly into revenue and retention, from recovering abandoned carts to recommendations and reactivation.

In practice, AI+MA delivers the highest return in areas such as abandoned carts, product recommendations, reactivation, post-purchase up-sell and service automation (chat). If you want a good starting point, cart + post-purchase is usually the quickest to close, because they have clear KPIs: revenue per recipient (RPR) and share of sales. These scenarios can also be tuned so you do not burn through discounts, for example by targeting offers only to people with a low propensity to buy. The result is a better match between the trigger and the customer, and a lower risk of training the list to wait for promotions.

• Abandoned cart: AI tests send windows and selects the trigger (reminder, social proof, discount) for the segment.
• Product recommendations: offer personalisation based on behaviour and preference data.
• Reactivation: AI detects a drop in activity and launches a campaign before the customer “disappears”.
• Post-purchase up-sell: automated suggestions for next steps and matched products after the transaction.
• Chat/support: automation of responses and qualification of questions that shorten the path to purchase.

In multichannel activities, orchestration is crucial, meaning deciding whether it is better to send push, SMS, email, or run retargeting, so that you do not spam and do not burn budget. For example, a customer with high LTV may have higher priority in a more expensive channel (SMS), while for a segment with a lower propensity to buy you activate cheaper touchpoints. The most common mistake is adding more automations without a strategy, which increases unsubscribe rates and damages deliverability. That is why, already at the planning stage, it is worth factoring in frequency limits and flow priorities in advance, before you implement further scenarios.

Channel architecture and communication orchestration

The architecture of channels and communication orchestration assumes that email, SMS, push, ads and onsite elements have assigned roles in the funnel and are synchronised with one another, rather than competing for the recipient’s attention. In practice, this comes down to choosing which channel should “play” at a given moment, for which segment and with what objective (acquisition, conversion, retention). For example, a user with high LTV may receive priority in SMS, while for someone with a low propensity to buy, cheaper push and retargeting in Meta/Google are triggered. This approach organises communication pressure and reduces duplicate identical stimuli.

Orchestration delivers the best results when real-time decisions are combined with batch analysis where immediacy is not required. Onsite decisions (recommendations on the page, a personalised banner, chatbot) must react in milliseconds, so infrastructure and cache are of key importance. By contrast, churn and LTV predictions can be calculated once a day in BigQuery and passed to MAP as segmentation attributes. This means automations do not have to “guess” whom to increase campaign intensity for and whom to reduce it for.

The key safeguard in orchestration is flow prioritisation and global frequency caps, so that the same people do not trigger multiple flows at once. In practice, exclusions and simple anti-spam rules are used, e.g. max 1 SMS/48h and blocking promotion of the same product after purchase for a defined period (e.g. 14 days). In addition, time and channel personalisation can be used: send-time optimisation selects the sending hour based on interaction history, and with small samples it is safer to rely on rules taking into account the time zone and activity. The result is a more predictable contact path and a lower risk of burning through the database.

1. 01Channel synchronisationCoherent operation without competition.
2. 02Contextual selectionThe right channel, segment, objective.
3. 03Value prioritisationHigher priority for LTV.
4. 04Real time + batchOnsite decisions and analysis.

Organises communication, reduces duplication and increases efficiency.

Platforms and integrations for effective automation

Platforms and integrations for effective automation are a set of tools that ensure a consistent customer profile, correct event tracking and the ability to run campaigns across owned, paid and onsite channels. Most often the core is the store platform (Shopify, WooCommerce, Magento) connected to a MAP (e.g. Klaviyo, Braze, HubSpot, Salesforce Marketing Cloud, ActiveCampaign), and at larger scale a CDP and data warehouse are added (Segment/mParticle + BigQuery/Snowflake). CDPs/data warehouses become particularly necessary when the number of sources grows and data needs to be unified for segmentation and models. In practice, the “sufficiency” of the stack depends on whether it is possible to reliably feed campaigns with behavioural, transaction and product data.

The foundation of AI and automation is a consistent event model and unambiguous identification of the customer and product across the entire stack. On the tracking side, AI expects events such as view_item, add_to_cart, begin_checkout, purchase, search and refund, as well as product attributes (category, margin, availability). A unique customer identifier (email/phone/user_id) and an unambiguous product_id remain crucial. In e-commerce, a large share of traffic consists of “guests”, so session stitching requires a well-thought-out strategy for collecting first-party identifiers and establishing an identifier hierarchy, otherwise it is easy to end up with duplicate profiles and distorted LTV. For greater flexibility in predictive analytics, GA4 gains the most after export to BigQuery, where you create features such as RFM, average basket value or time to purchase.

Integrations should be designed so that automations and algorithms do not optimise for a “false reality” of the data. Inconsistent order states often appear (cancelled, returns, partial returns) and there is a lack of margin information in the marketing tool, which makes it difficult to optimise for profit rather than revenue. With growing cookie restrictions, stability is improved by server-side tracking (GTM Server-Side) and sending events via CAPI (Meta) and Enhanced Conversions (Google), which usually reduces conversion discrepancies in ad dashboards. When you need quick process connections, Make/Zapier speed up implementation, but for real-time automations (e.g. basket, onsite personalisation) webhooks and native integrations or queues (Pub/Sub, SQS) perform better.

Personalisation and predictive segmentation in practice

Personalisation and predictive segmentation in practice comes down to using data to forecast customer intent and value, and then adapting the offer, channel and timing, rather than communicating “the same thing to everyone”. Most often you start with RFM segmentation (recency, frequency, monetary), because it quickly answers whom to reactivate and whom not to “bother” with overly frequent communication. In RFM, the “Champions” segment may receive early access to launches, while “Hibernating” gets only 1–2 campaigns per quarter with a strong value proposition, so as not to harm deliverability. Such logic organises priorities before you add more advanced models.

LTV (CLV) prediction allows you to steer budgets and bids towards customers who will spend more than average within 90/180 days. In practice, you build a high-LTV list as a seed for Meta/Google and set a higher bid for the segment with predicted LTV > 500 zł. This means optimisation does not end with “cheap” acquisition, but supports cohort quality and long-term sales. This approach works best when the segments are then used consistently in campaigns and automations.

Propensity scoring answers the question “who will buy within the next 7 days?” and makes it possible to narrow discounts to the people who actually need them. A customer with a propensity of 0.85 may get recommendations without a discount, while a customer with 0.25 is more likely to receive a bundle or free delivery, because otherwise they will not come back. The greatest value of predictive models appears when, thanks to them, you reduce promotional pressure and do not train your database to wait for a discount. In practice, this means different incentives for different segments, instead of one mass promotion.

Product recommendations and onsite personalisation make sense when they take into account the store’s real constraints, namely availability, margin and exclusions (e.g. incompatible accessories). In stores with a large range, recommendations based on embeddings (e.g. Vertex AI, Amazon Personalize) perform better than simple “customers also bought”. AI can also reshuffle product order on the listing for a specific user, which can be a bigger uplift than email alone, because it works exactly at the moment of decision. This is complemented by personalisation of time and channel: send-time optimisation picks the hour based on interaction history, and with small samples it is more sensible to rely on rules that take time zone and activity into account.

1. 01Intent and value forecastMatching the offer and timing
2. 02Start with RFM segmentationQuick prioritisation of communication
3. 03LTV value predictionBudget steering by potential

Effective personalisation starts with data and ends with the right value proposition, avoiding communicating “the same thing to everyone”.

End-to-end sales automations

End-to-end sales automations are a set of connected paths from the first contact to retention, which consistently deliver revenue without manually guiding every customer. A welcome flow should immediately answer the question “who is this customer and what are they interested in?”, rather than being reduced only to an upfront discount. For example, a 3-step flow in Klaviyo can collect preferences (quiz/clicks), and AI then selects starter products on that basis, strengthening the first purchase and later recommendations. The best-performing paths are those that both increase conversion and improve the quality of data about customer preferences.

It is worth optimising abandoned cart and checkout flows so that you test send windows and incentives instead of clinging tightly to a fixed scheme. AI can check different timings (e.g. 30 min, 4h, 20h) and tailor the message to the segment: reminder, social proof or discount. If the goal is to limit discounts, you direct the promotional incentive to users with a low likelihood to buy or to those who abandoned the basket above a certain value (e.g. > 300 zł). As a result, automation works like a selective “boost” rather than a permanent promotion for everyone.

Browse abandonment targets people who viewed a product or category but did not add anything to the basket. This usually gives greater reach, but lower purchase intent than abandoned cart. A well-designed scenario shows the specific variants viewed (colour/size) and suggests 2–3 alternatives at a similar price, instead of sending generic bestsellers. It is worth relying on onsite intent signals such as filter usage (e.g. “size 38”) or comparing several products, because this can significantly shorten the path to purchase. For example, after intensive laptop comparisons, it is better to send an email with a comparison table, availability and delivery time than a standard newsletter.

Post-purchase should answer the customer’s question “what next?” and at the same time reduce the risk of returns thanks to instructions and relevant recommendations. Example: after buying a coffee machine, you send a series about setup, descaling and suggestions for a grinder and coffee, instead of promoting more coffee machines. In parallel, automations based on the life cycle can detect a drop in purchase frequency and launch a campaign before the customer “disappears”. If the average interval between purchases is 35 days, you set the reactivation trigger at e.g. 45–55 days, and adapt the content to the last purchase category.

Back-in-stock and price-drop build demand for a specific SKU because they directly answer the need “when will my size/colour be back?” and usually do not require aggressive discounts. For example, back-in-stock sent through two channels (email + push) with a time limit on basket reservation (e.g. 30 min) reduces the risk that the customer will again encounter out-of-stock. For higher-value baskets (high-ticket B2C or B2B), lead nurturing works well: lead scoring, email sequences and a task for sales when the lead reaches a threshold (e.g. 80 points) and visits the pricing page twice. This is supplemented by chatbots and assistants (Intercom Fin, Zendesk AI, Tidio, Drift), which answer availability, delivery time and compatibility immediately, and you measure the impact by post-chat conversion, average response time and the share of escalations to a human.

AI in content creation and the shopping experience

AI in content creation and the shopping experience speeds up the production of materials and personalises contact, but it requires oversight so as not to mislead the customer. An LLM can prepare variants of email subjects, preheaders and CTAs for different segments, and then you keep the best versions based on clicks and revenue per send. In abandoned cart campaigns, you can develop separate versions of communication for the “price sensitive” and “premium” segments, instead of one message for everyone. To make AI safe for sales, impose style, length and a list of prohibited phrases, and implement quality control before publishing.

AI can also support product descriptions and SEO, provided the content remains aligned with the parameters and the product feed. A proven practice is to create descriptions based on structured attributes (e.g. material, dimensions, compatibility) and then validate them with rules so they do not diverge from the data in Merchant Center. In dynamic advertising, DCO helps by combining headlines, visuals and CTA to surface the best-performing combinations without manually preparing hundreds of variants. The condition remains a consistent catalogue and high-quality input assets, otherwise the algorithm may amplify incorrect or unavailable variants.

In the store itself, AI increases conversion mainly by matching the experience to intent, rather than simply adding more messages. Personalised landing pages can change the hero layout, sections and social proof depending on the traffic source (e.g. one layout for comparison sites, another for remarketing), so the user reaches the right filters and arguments more quickly. An AI-powered internal search (e.g. Algolia, Elasticsearch, Vertex AI Search) improves relevance through synonyms, typo correction and conversion-based ranking, which shortens the path “from query to product”. In turn, moderation and review analysis help identify recurring issues (e.g. “small sizing”) and limit the exposure of products with a growing share of negative reviews in recommendations and campaigns.

Generative graphics and video make sense as a tool for preparing creative variants, but they should not change product features, because that increases the risk of returns. AI can also act as a shopping “agent” assistant that gathers preferences and builds a basket, provided this is connected to the catalogue and stock levels (available size, delivery time). Communication stability is ensured by a prompt library, brand voice, a list of claims and a dictionary of banned phrases (brand safety), especially in sensitive sectors. This allows AI to create content within permitted boundaries and not generate “random” promises.

Optimising advertising campaigns and budgets with AI

Optimising advertising campaigns and budgets with AI comes down to algorithms receiving better signals and conversion values, while you assess their real impact through experiments and incrementality analysis. In Google Ads (Performance Max, Smart Bidding), it is crucial to provide correct signals and conversion values, and when the goal is profit optimisation, it is worth setting values adjusted for margin or using value rules for high-LTV segments. In Meta Advantage+, the best results come from feeding the algorithm first-party signals, e.g. via Conversions API with purchase events together with value and currency and the “VIP” tag. This approach reinforces learning on the most desired transactions instead of relying solely on “random” conversions.

Attribution and testing are essential because the last-click model underestimates the role of the upper funnel and can lead to misguided budget cuts in prospecting. In practice, it is worth combining data-driven attribution (GA4) with incrementality tests (geo-holdout), and at larger scale also with MMM (e.g. Robyn, Meridian). Holdout tests (e.g. 10% of the group without campaigns) show the real uplift, rather than just the ROAS “assigned” by the platform. If you do not verify incrementality, AI may optimise for platform metrics rather than additional sales.

• Budget across channels: manage it through LTV:CAC and payback (e.g. target LTV:CAC ≥ 3 and payback ≤ 60 days), and raise bids for high-LTV and lower them for low-LTV.
• Feed for algorithms: refine titles and attributes, because systems “read” the feed, and with imprecise data you can easily lose match quality.
• Rules and scripts: implement automatic pauses and alerts (e.g. when CPA rises by 30% week on week), so you can stop budget burn in time.
• Creative testing: with a large number of variants, consider a multi-armed bandit (MAB), and leave classical A/B testing for single hypotheses that require a hard statistical conclusion.

The results of campaigns are increasingly influenced by the quality of the feed and creatives, because they determine intent matching and delivery stability. Feed optimisation (titles, attributes, product segments) improves query matching and traffic quality, and in catalogue campaigns it reduces situations where the algorithm promotes variants based on incorrect data. In demand forecasting, models (e.g. Prophet, Vertex AI Forecast) help predict what may run out, so you reduce exposure to SKUs at risk of stock depletion and shift budget to substitutes. This closes the “marketing ↔ stock” loop, thanks to which automation does not sell what you are unable to deliver.