The Five-Click Gamble: Session AI Bets the Discount on What You Do Now

The most expensive coupon in ecommerce is not the one that fails. It is the one handed to a customer who was already reaching for a credit card. The sale appears in the ledger, the marketing dashboard glows, and ten or fifteen percent of the order quietly vanishes. Nobody complains. The customer certainly does not. A margin has been mugged in perfect silence.

Session AI has built its company around finding that silent mugging. Its software watches the small movements of a live visit - the pages opened, the cart edited, the hesitation, the pace - and sorts shoppers by what they seem likely to do next. A likely buyer can be left in peace or shown an upsell. A shopper on the fence can receive the 10 percent offer. Someone unlikely to buy might be invited to trade an email address for a future coupon. The point is not to personalize everything. It is to intervene only when the intervention may change the ending.

The stranger at the window

Traditional personalization likes a dossier: past orders, loyalty status, demographics, an email address. That works nicely for a known customer. The trouble is that most ecommerce visits are anonymous. Session AI's founders - Debjani Deb, Arnab Mukherjee, and Manish Malhotra - chose the harder, more fleeting object of study when they started ZineOne in 2014: not the person across time, but the session unfolding now.

That distinction sounds academic until one remembers how quickly a browser tab dies. A weekly audience segment is useless to the visitor leaving in 40 seconds. Session AI says its models can estimate intent in about five clicks and return a decision in under 50 milliseconds. It does not require personally identifiable information or third-party cookies for the core prediction. The raw material is behavior, which is both less intimate and more urgent.

The category has grown fashionable names. What began as real-time customer engagement became in-session marketing when ZineOne rebranded as Session AI in 2023. The current label is agentic decisioning. Beneath the wardrobe changes sits a consistent mechanism: set goals, predict an outcome, take an action, keep a control group, learn from the result. The company did not discover that retailers love AI. It discovered that retailers dislike giving away money.

What failed first was generosity

Blanket discounting is simple, measurable, and ruinously easy to repeat. One Session AI case study describes a multi-brand retailer whose sitewide promotions were diluting brands and eroding profit. That leakage changed the argument: restraint stopped looking timid and started looking measurable. Privacy changes also made the old supply of third-party identity signals less dependable. The response was almost comically restrained: withhold offers from shoppers predicted to buy, give a short-lived incentive to the influenceable group, and reserve future coupons for the least likely buyers. The company reported a 12-times first-year return and a 4 to 15 percent lift in revenue per visitor across use cases.

A footwear case study makes the same argument with louder numbers: a projected 20-times return, a 38 percent conversion lift in the on-the-fence segment, and $200,000 in incremental revenue during one week. These are vendor case studies, not laws of retail physics. Their proper use is not worship but imitation: define the segment, preserve a holdout, measure incrementality, and count margin as carefully as conversion.

A quarter-million-dollar traffic light

Session AI does not publish a friendly three-tier pricing page. It sells enterprise software. The clearest public number appears on AWS Marketplace: $250,000 for a 12-month contract, with the price tied to annual site traffic. Reporting, continuous model retraining, and deployment are included in that single listed dimension; contracts negotiated elsewhere may differ.

This cost explains the customer list: Sephora, Mattress Firm, New Balance, Men's Wearhouse, Balsam Brands, Tailored Brands, and other large operators. At that scale, a fractional conversion move can matter, but so can a fractional reduction in discount leakage. Session AI's homepage advertises examples such as a 35 percent reduction in promotions worth $40 million and a 10-times annual return for a $5 billion retailer. The sums are large because the sites are large.

The deployment is designed to avoid a replatforming melodrama. A retailer can use a website tag, SDK, plugin, cartridge, or server-to-server connection. The platform works with tag managers such as Google Tag Manager and Tealium, and it has Adobe Commerce and Adobe Campaign connectors. Business teams choose goals and guardrails - a margin floor, a discount cap, an inventory target, perhaps a rule never to disturb checkout. The agent then chooses from approved messages rather than improvising a carnival.

The grown-up feature is the control group

The flashy feature is prediction. The credible feature is abstention. Session AI says its testing system preserves control groups and measures each agent-driven variant against them. That matters because ecommerce attribution is a gifted flatterer. If a likely buyer sees a banner and purchases, the banner will happily take credit for gravity.

A control group answers the less glamorous question: what would have happened anyway? Session AI's agentic A/B testing shifts traffic among variants while retaining a baseline. The practice can be copied without buying the product. Stop counting every purchase after an offer as a victory. Keep a genuine holdout. Optimize for incremental profit, not the number of coupons clicked. Let “no offer” compete as a treatment.

Where the bet earns its odds

This is not software for every shop. Machine learning needs enough repeated behavior to learn, experiments need enough traffic to settle, and optimization needs several legitimate actions to choose among. A small merchant with a few thousand visits, one promotion, and thin gross margin may simply have purchased an expensive traffic light for an empty road.

The company itself has taken the long route to this narrower claim. It raised $2.5 million in 2017, another $2.5 million in an Omidyar-led Series A, $10 million in a Norwest-led Series B, and $28 million in a SignalFire-led Series C. Roughly $43 million later, it has patents around event sequences, selective intervention, and friction analysis; offices in San Jose and Mumbai; and a new chief executive, Sanjeev Kulkarni, with co-founder Debjani Deb serving as president.

There is a pleasing irony in the result. Marketing software has spent decades collecting more history about more people. Session AI's wager is that, at the moment of purchase, a few fresh gestures may beat a dusty biography. Five clicks are not a person. They may, however, be enough to keep a retailer from offering that person 10 percent off for no reason at all.