How Fully Homomorphic Encryption enables marketplaces to deliver search, recommendation, fraud detection and personalization — without ever decrypting what belongs to the consumer.
If you read only one thing from this eBook, read this.
The modern large e-commerce platform is, simultaneously, the largest behavioral-data collection operation in consumer markets and the operation that most depends on that data to function. Every click, every search, every product viewed and abandoned, every purchase completed, every order returned, every review posted — all of it is raw material for the algorithms that decide what you see on your next visit. The entire marketplace operation depends on processing this data at scale and in real time.
Over the last five years, three things have converged to create a strategic problem that nobody in the industry is comfortable discussing:
Conversion on funnels that require explicit consent has dropped 15–35% over the last 24 months. Users are exhausted.
LGPD, GDPR and the AI Act are demanding proof of minimization. Fines in the hundreds of millions have already been imposed in the sector.
Each billion AI calls per month costs millions. A proprietary model is the only path to economy at scale.
Those who outsource inference to OpenAI/Anthropic are handing data to parties who may become competitors.
FHE — Fully Homomorphic Encryption — unlocks an entirely new category of operation: search, recommendation, fraud and personalization without decrypting the user. It allows the marketplace to do everything it does today, with mathematical proof that the central server has never seen the consumer's individual history. And it enables anti-fraud consortia among competing marketplaces that are legally impossible today.
The next decade of digital retail will be defined by which marketplaces manage to deliver a personalized experience without persistent profiling — and by who does it first.
E-commerce has become — without realizing it — one of the largest behavioral-data collection operations in the world. Every feature users love is, on the other side, a new layer of regulatory risk.
In 2010, a typical e-commerce operated on two types of data: registration and purchase history. That was it. Algorithmic decisions were simple — "those who bought X also bought Y". There was no serious public debate about privacy in digital retail, not there, not anywhere. Most consumers had no idea what a third-party cookie was.
In 2025, the same e-commerce processes hundreds of data points per session. Mouse movement, time on page, scroll depth, device, geolocation, time of day, IP, browser fingerprint, search history from other sites (via integrated pixels), typing patterns, payment preferences, post-purchase behavior. Each of these points becomes input for increasingly sophisticated algorithms. The algorithm decides which product appears first, which shipping fee to charge, which coupon to offer, when to send the abandoned-cart email.
| Operation | What it is | Risk |
|---|---|---|
| Personalized search | Search results vary by user | Persistent profiling under LGPD |
| Recommendation | Suggestions based on longitudinal history | Special category when it includes health/orientation |
| Dynamic pricing | Prices varying by user | Algorithmic discrimination |
| Fraud detection | Model over behavioral patterns | Extensive data processing |
| Sales chat (live commerce) | Real-time conversation to close the sale | Intimate consumer content exposed |
A consumer who searches for "baby crib" does not understand, in any practical sense, that this search will enter a profile that lives for years on servers and will be cross-referenced with dozens of other searches to form a detailed picture of who they are. They believe they are running a search. They are also contributing to a persistent behavioral profile.
The question for the marketplace board is not whether the current architecture is sustainable. It is how long until the first public decision reshapes the sector.
Three continents, three converging regulations, and a consumer fatigue that no one has yet priced in.
LGPD does not classify behavioral data as a special category — but it requires a robust legal basis for processing, and the Brazilian Data Protection Authority (ANPD) has been interpreting "legitimate interest" increasingly narrowly. In 2025 the authority opened an investigation against a large marketplace for using search data for cross-sell without adequate consent. It will be the first of many.
GDPR turned the cookie consent banner into a permanent operating cost. More importantly: national Data Protection Authorities regularly fine sites that collect more than they need. The TCF (Transparency and Consent Framework) that underpins European behavioral advertising has already been partially invalidated by the Belgian authority. The industry operates in an expanding gray zone.
The AI Act, phased in through 2027, classifies AI systems that assess consumers for credit, pricing or ranking as high-risk. Marketplaces that do dynamic pricing, fraud risk scoring, or seller ranking are in the crosshairs. Estimated compliance cost per system: six to seven figures.
The Illinois Biometric Information Privacy Act (BIPA) generated billion-dollar settlements against Facebook and Google for facial recognition use. Marketplaces with biometric features (AR try-on, image search) are in the crosshairs of the next wave.
This is the least discussed and most important part for marketplaces. Conversion on funnels requiring explicit consent has dropped 15–35% over the past 24 months on most measured platforms. Users are tired. Every extra "I accept" click the marketplace asks for is a lost conversion. The traditional solution — asking for more consent — no longer scales.
FHE offers a different path: an architecture where data is never "processed" in the legal sense, because it remains encrypted. This drastically reduces the need for robust consent, restoring conversion and simplifying UX.
| Risk | Probability 5 years | Impact |
|---|---|---|
| LGPD/GDPR fine for weak legal basis | High | 2–4% of global revenue |
| BIPA class action over a biometric feature | Medium | USD 100M–650M |
| Feature blocked by the AI Act | High in the EU | Loss of regional market |
| Consent fatigue → drop in conversion | Certain | Silent loss, hard to measure |
| Reputational crisis after a breach | Medium | 12–24 months of impact |
No mathematics. What the board needs to understand.
A transparent vault. You see that something is inside, you do not see what it is. You manipulate the contents from outside — add, multiply, compare, run entire recommendation models — without ever opening it. You return it sealed. This is FHE.
All cryptography your marketplace uses today protects data in transit (TLS) and at rest (AES). The third state — in use, during processing — has always required plaintext. That is the instant when the search engine accesses the user's history in plaintext. It is where the recommendation algorithm runs. It is where the fraud system analyzes patterns. FHE eliminates the third state.
| Technology | Promises | Fails |
|---|---|---|
| De-identification | "We removed identifiers" | Trivial re-identification via linkage |
| TEE | "The chip isolates" | Trusts the manufacturer; vendor lock-in |
| Federated Learning | "Data stays on the phone" | Gradients leak data |
| Differential Privacy | "We added noise" | Poor for individual personalization |
| FHE | "Server never sees in plaintext" | High computational cost — but decreasing |
FHE is expensive for billion-call volumes — that is the honest argument against it. The answer is a hybrid architecture: use FHE in the sensitive core (persistent profile, deep personalized recommendation, collaborative fraud) and keep the rest of the operation on traditional architecture. For a Brazilian top-10 marketplace, the total investment is below 0.2% of the IT budget — comparable to the cost of a single payment platform migration.
Search is the #1 direct-ROI use case in any marketplace. Every percentage point of improvement in search conversion is worth millions. Today the improvement depends on personalizing results per user — which requires a persistent plaintext profile.
Under FHE: the user profile remains encrypted, the search engine computes rankings over the encrypted profile, and returns personalized results without the server seeing the history. Conversion improves without persistent profiling.
"Why are you seeing this" — personalized recommendation is the core asset of any marketplace. Today it requires plaintext history. Under FHE, the algorithm runs over encrypted history, and the recommendation appears to the user without the central server ever seeing what they bought.
This is the case where FHE unlocks something unique. Fraud — chargebacks, mule accounts, fake sellers, review purchasing, ranking manipulation — is a sector-wide problem. Each marketplace fights it in isolation, with limited effectiveness because fraudsters move between platforms. Effective fighting requires cooperation between competing marketplaces. Today impossible.
Under FHE with PSI (Private Set Intersection): marketplaces encrypt lists of tax IDs, IPs and suspicious patterns, discover only the intersection — those who appear in three or more platforms with fraud patterns. Without revealing entire databases. It is a sector-wide defense that literally does not exist today.
A top marketplace has 100M+ SKUs. The catalog needs descriptions, titles, bullets, alt-text, SEO, in multiple languages. Today this requires sending SKUs to an external LLM (expensive, commoditized, data-leaking). Under FHE on a proprietary model: descriptions are generated over encrypted data; the LLM operator (even if it is a vendor) never sees the proprietary catalog.
80% of customer service contacts are repetitive: "where is my order", "how do I return", "my boleto expired". The human agent needs to see the history. Under FHE, the chatbot operates on encrypted history, returning answers without persisting the interaction. The human agent sees only what is needed during the call and archives nothing afterwards.
Marketplaces use sales data to help sellers sell better — price suggestions, listing improvements, demand prediction. That data is the marketplace's IP that sellers cannot see in detail. Under FHE, suggestions can be delivered to the seller without the seller ever accessing the models or other sellers' data.
Virtual makeup, virtual clothing, virtual glasses. Every try-on involves 3D facial mapping processed on a server. Under FHE, the mapping happens locally on the phone and virtual pigments are applied under encryption. The marketplace receives aggregated metrics ("X users tried lipstick Y") without storing the image.
Dynamic pricing is legally sensitive because it can become algorithmic discrimination. Under FHE, the pricing algorithm runs over encrypted user data, returning a price without the central system having seen the justifying data. This offers a robust defense against discrimination claims.
Real-time conversation between seller and buyer is the fastest-growing channel. Each conversation contains intimate information: motivation, context, preferences. Today, it sits on the marketplace server. Under FHE, the conversation can happen with end-to-end encryption, and the marketplace receives only aggregated metrics (close rate, average time).
| Component | Investment |
|---|---|
| Founding team (crypto + ML eng + PM + legal) | USD 4M – 7M / year |
| Licenses and tooling | USD 300k – 1M |
| Compute infrastructure | USD 1.5M – 4M |
| Strategic consulting | USD 800k – 2M |
| Regulatory study | USD 300k – 800k |
| Integration with core platform | USD 1.5M – 4M |
| Total year 1 | USD 8M – 17M |
| Item | Estimate |
|---|---|
| Compute | USD 2M – 6M |
| Maintenance team | USD 4M – 7M |
| Audit | USD 400k – 1M |
| Stabilized annual opex | USD 6.4M – 14M |
For a Brazilian top-10 marketplace with revenue above USD 5B, this represents between 0.15% and 0.3% of revenue.
Conversion dropped 15–35% in funnels with robust consent flows. For a USD 5B marketplace with a current conversion rate of 2–4%, recovering one percentage point is worth USD 50–150M per year.
Estimated sector fraud: 1–3% of revenue. For a USD 5B marketplace that is USD 50–150M of annual exposure. Capture via PSI: USD 15–50M per year.
Inference outsourced to OpenAI/Anthropic at billion-call scale costs millions per month. A proprietary model under FHE captures that saving: USD 30–100M per year over three years.
Expected exposure to fines: USD 50–200M NPV. Hedge: USD 20–80M of insurance value.
"The marketplace that respects your data" as a premium narrative. Hard to quantify but measurable in retention and LTV.
For any top-10 marketplace, FHE is the digital transformation investment with the highest return asymmetry available in 2026.
Marketplaces are the industry that loses the most when trust collapses, and the first to feel it when consumers get tired. Every new feature that asks for consent adds a bit of friction. At some point, the friction becomes systemic drag — and competitors that offer the same value with less drag win.
Focus on direct communication with Gen Z. Public privacy manifesto. Works best for premium marketplaces or D2C brand-first.
Focus on building an FHE consortium among marketplaces. Captures the sector-organizer role. Works for the top 3.
Focus on a proprietary AI model without dependence on OpenAI/Anthropic. Inference sovereignty as a differentiator. Works for large marketplaces that want to become technical platforms.
The scenario to spell out: what happens if none of the large Brazilian marketplaces structurally adopts FHE? Answer: they will remain dependent on OpenAI/Anthropic/Google, paying growing margins, delivering data to parties that may become direct competitors. In five years, the gap will be structural and expensive to reverse.
Hire a founding crypto engineer. Identify three use cases. Align with legal and the DPO.
Build one end-to-end case. Recommendation: fraud detection under FHE OR encrypted personalized recommendation for a segment.
Launch to a user segment. Measure conversion, NPS, latency. Compare against a control group.
Campaign. Manifesto. Possibly the first anti-fraud consortium with other marketplaces.
High probability. Mitigation: hybrid architecture. FHE only in the sensitive core. The rest stays on traditional architecture.
Mitigation: partnership with specialized consulting.
The growth team depends on aggressive personalization. It will resist. Mitigation: show that FHE preserves personalization.
Mitigation: speed.
FHE must report to the CDO/CMO, not the CIO.
Announcing before a working product generates backlash.
Who custodies the user's key? UX is half of the project.
The industry you lead grew on an old promise: that retail can be cheaper, more convenient, more personalized when intermediated by software. That the consumer who prefers to buy online instead of in the physical store gains time, price and variety. That the marketplace is the infrastructure that makes this convenience possible.
That promise has lasted two decades. It survived bubbles, crises, new technologies. It endured because it was — and largely still is — true. Consumers who buy online instead of in the physical store do so for real, practical reasons.
But over the last fifteen years, without anyone decreeing it, the relationship between marketplace and consumer has changed in nature. The consumer is no longer someone who enters the site, searches for the product, buys, and leaves. They have become a continuous source of behavioral data feeding ever more sophisticated algorithms. The aggregate result is an operation that extracts value from consumers in ways they would never consciously consent to.
It is possible to walk this back without losing the benefits. FHE allows you to continue offering personalization, recommendation and fraud detection — without ever seeing the individual consumer.
In three years, some marketplace will lead. The question is whether it will be yours, or the one you will have to look at as a reference.
There is a window. It is narrow. It is real. The rest is courage.
Computation over encrypted data.
Private Set Intersection — discovering the intersection of datasets without revealing the rest.
FHE schemes.
Vector representations of products and queries that underpin semantic search.
Relevant regulations.
FHE libraries.
| Vendor | Focus |
|---|---|
| Zama | Concrete, TFHE, focus on developer experience |
| Duality | OpenFHE, focus on finance and analytics |
| Inpher | FHE+MPC for finance |
| Tune Insight | Lattigo |
| Stickybit | Brazilian technical boutique |
The Catalog That Does Not Spy
Strategic eBook for senior management at e-commerce platforms and marketplaces.
Volume I · Edition 2026 · Confidential distribution.
Set in Iowan Old Style and SF Pro.
— end —