Act I · Tension
Customers find products in four places at once
Search. Category listings. Recommendations. Increasingly, conversation too.
Conversation posts 58.11% click-through. We'll come back to this number at the end.
Search
Category listings
Recommendations
Agent
Four product discovery surfaces
Four discovery surfaces. One engine.
Search doesn't know what the customer clicks in recommendations. The category listing doesn't know what they searched for a moment ago. Each surface has its own engine, its own data and its own logic.
Search signals improve recommendations. Recommendation clicks change listing results. The model learns from every interaction on every surface.
TYPICAL STACK
Separate data. Separate logic. No signal carries over.
SYNERISE
Behavioral AI infrastructure
one engine learns from every interaction on every surface
One engine powers product discovery on every surface
Behavioral AI
Every decision grounded in behavior
Synerise predicts the next action of every person or object from real-time behavioral signals. It activates that action in any channel, human or agentic. Sub-millisecond inference, self-supervised foundation model. Open by design: MCP, A2A.
2.6B
decisions per day
65K
AI predictions per second
229B+
TerrariumDB queries per month
On proprietary systems
- Schema-free modeling: entities and relationships change without migrations
- Sub-millisecond inference · self-supervised foundation model
Architecture
Every path runs through the same governance layer
OPERATOR PERSONA
UX
Marketing operators, loyalty program managers, campaign teams
AGENT PERSONA
MCP + Skills
Claude Code, Cursor, ChatGPT — engineering and analytics teams
DEVOPS PERSONA
DX
Platform engineers, CI/CD pipelines, enterprise IT
Governance layer
Approval workflows · Audit log · Object versioning · Rollback
Synerise platform
Automations · Segmentations · Templates · Recommendations · Brickworks · Loyalty
The same approvals, audit trail and versioning on every path
Value
Revenue × productivity × interoperability
On one platform, each vector amplifies the next. Internal AI agents and MCP become the command center for your system. They take complex technical processes off your team's plate, automate the work and give you full control of the platform.
01 — Revenue
Customer Agent
- Agentic search across every channel
- Behavioral signals → conversion
- Ready for zero-click commerce
02 — Productivity
System Agent
- Agents that solve complex use cases
- They work with full context and operational knowledge
- Marketers focus on strategy, not integration
03 — Interoperability
MCP + Plugins
- Any agent reads from and writes to Synerise
- Developers deploy from Claude Code, Cursor and Codex
- No vendor lock-in when your stack changes
The third vector lets the first two scale beyond Synerise
Act II · Synerise Customer Agent
Every customer gets your best sales advisor
One agent that knows the customer, speaks in your brand's voice and sells based on current product and business data.
The agent's capabilities on one map
Personalization and context
Product discovery and sales
Conversation experience
Data, actions and integrations
Control, analytics and deployment
Act III · Voice
You define the persona. No code.
01 — Brand
Name and business description: products, channels, languages.
02 — Assistant
Motto, tone of voice, style guide and the agent's scope of work.
03 — Behavior
Word limit, engagement level, clarifying questions and instructions on when to ask them.
I'm not Agent Smith. Or an agent of S.H.I.E.L.D.
Give your agent its own persona. Same technology as every other client. Your own brand voice.
Charismatic? Sharp and energetic? Honest? Direct, funny, confident… anything goes. It's not a stiff bot or a hotline. A buddy who knows the gear better than anyone and is on the customer's side? Why not.
01 — Discovers
Understands the need and finds the right product in your catalog.
Picks up on the customer's words and finds what they meant, not what they typed.
02 — Helps decide
Clears up doubts, compares options and leads to the purchase.
Picks one option and talks about it in the first person, with an opinion. If the cheaper one wins, it picks the cheaper one.
03 — Takes care
Handles orders, deliveries and returns in the same chat.
Knows its limits: hands the conversation to the contact center and links to the right page.
04 — Inspires
Suggests styles and what goes with them. Works like a personal shopper.
Offers accessories and complementary categories, one click away from the answer.
Same agent, any brand voice
You hear the persona in every sentence
1 — Says no in its own voice
“Discounts like that would be great, but I'm here to help you find products, not set prices.” No quoting the terms and conditions. Then it closes with a question.
2 — One question, then results
A gift? One question about what they like. No grilling on occasion, budget or style.
3 — Picks up on the customer's words
The customer types “beauty,” and the agent builds the whole answer on that word.
4 — Sells the effect
“A healthy glow people notice,” “like a day at the spa.” Not a single spec.
Product discovery and sales
Sells like your best sales advisor
Product recommendations
The full range of Synerise recommendation models as agent tools, with your boosting, filters and business logic.
“Something like this, but cheaper” triggers a similar-products campaign.
Quick filters
The agent builds filters around the current results, and the customer narrows the list in one click.
After “show me laptops”: RAM, screen size, budget.
Complementary categories
Below the results, the agent adds buttons for accessory categories.
A camera? Lenses, tripods, memory cards.
Add to cart
Every product in the chat has an add-to-cart button. The customer doesn't have to type anything.
One click on the recommended TV and it's in the cart.
Clarifying questions
You decide whether, and how often, the agent asks customers about their needs.
A broad question? First: “For gaming or everyday use?”
Follow-up questions
The agent ends its summary with a question that leads to the next step.
“Want me to compare the top two?”
One question, then everything is one click away
1 — Clarifying question
Asks about what splits the catalog. Two clicks, no typing.
2 — Add to cart
Every product has its own button. The conversation ends in the cart, not with a link.
3 — Quick filters
It builds the filters itself: categories, top brands, price cap.
4 — Complementary categories
Then whatever goes with the chosen product, one more click away.
Two clicks instead of typing
Similar products, each with a reason
1 — Three picks, three reasons
Longest battery life, best value for money, a balanced model with PC support. The campaign picks the products; the agent says what each one is for.
2 — Names the trade-off
Marshall models look great and last a long time, but have no ANC. The agent says plainly that they make sense if noise canceling isn't a priority.
3 — Follow-up question
It ends with a question that decides the next step: ANC or the longest battery life?
Compares and delivers an honest verdict
1 — Same price, different job
2,599 zł vs. 2,599 zł. Price doesn't settle it, so the agent explains the real difference: who each watch is for.
2 — Specs as arguments
One is for training, the other for the outdoors. The 1.5-inch screen and the pulse oximeter come up once, as arguments in the conversation.
3 — “My honest take”
It speaks in the first person, names a winner for each lifestyle and asks which one fits the customer.
Act V · Personalization and context
Knows who it's talking to and where in the store
Profile context
The agent recognizes the Synerise profile behind the conversation. So it personalizes the results too: ranking, comparisons and answers.
Two customers ask the same question and get two different lists.
Shopping context
From the first message, the agent serves the customer in their context: what they bought, what's in their cart, what brought them back.
“I see the GARMIN Varia in your cart,” before the customer types a single word.
Page context
The conversation follows the page. Product, category or cart: the topic changes with where the customer is.
On a kettle page, it opens the conversation with that kettle.
Personalization covers ranking, comparisons and answers
Profile context, page context and shopping context in one
Viewed product
What they're looking at right now.
Hi, Aneta
In the cart
Purchase started, not finished.
Hi, Aneta
Recent order
What they already bought, and when.
Hi, Sebastian
Every opener comes from what the customer is doing right now
Context includes orders too
It checks the orders on the account and says what it found. It doesn't make up a delivery date or send the customer off to support.
Orders on the account
Read live from your systems.
No orders? It says so
It doesn't make things up when data is missing.
It checks and reports what it found. It doesn't invent delivery dates.
The page decides how the conversation starts
In the mobile app, the agent starts with a recommendation carousel.
Electronics — product page
It starts with the product the customer has on screen.
Fashion — product page
Same mechanism, different assortment: from the sweatshirt on screen to category quick replies.
Fashion — entry from search
The agent appears in search suggestions, next to recently viewed products.
Electronics and fashion · the same mechanism in two stores
Act VI · Data, actions and integrations
The catalog is only part of it. The agent runs on your ontology.
Synerise data
Orders · Points · Vouchers
External systems
connected via MCP
Documents
FAQs, policies, manuals
Your ontology
objects · relations · actions
Agent — every new tool is a new skill
Synerise data as tools
“Where's my order?” Status read live.
Brickworks connects systems
Complaints, warranties and in-store availability as objects in the ontology.
External MCP integrations
Checks consultant availability before offering a handoff.
Document knowledge base
Returns questions, answered from your own PDF. Coming this quarter.
Selling products is only one of its jobs
Under the hood
The language model runs the conversation. The answers come from your data.
Semantic caching, custom nodes and the reranker decide what reaches the model at all. The LLM gets ready-made context from platform data instead of a raw query.
Underneath, Synerise models, product attributes, search, recommendations and Brickworks do the work. The language model writes the sentences. The content comes from the platform.
Request with user prompt
Synerise Customer Agent Service
Semantic Caching Layer
Custom Nodes (e.g. fast intro node)
Search & Reranker node
Tools
LLM Gateway
LLM Model (via Azure Foundry)
Grounded in Synerise capabilities
Synerise ML/AI models (SLMs, classifiers, …)
Items Attributes
Items Search
Recommendations
Brickworks
Embedding models
Under the hood: the agent service and Synerise capabilities
Data, actions and integrations
The agent's tools in one place
01 — Recommendations
Similar, personalized, bestsellers, cross-sell. Your campaigns work as tools.
02 — Brickworks schemas
Order list and status from your systems.
03 — MCP server and actions
Consultant availability, conversation handoff, links for complaints, installments and promotions.
You connect the tools. The agent decides which one to use.
Agent
picks the tool on every call
Search
searches your catalog
Recommendations
similar and complementary product campaigns
Brickworks
order status from your systems
MCP + custom action
checks availability, then hands off
No routing rules to maintain
The agent reads the request and picks a tool. In a single answer, it can combine several: search, then recommendations, then a filter. The customer doesn't have to leave the conversation.
The shelf is still yours
You decide which campaigns, models, filters and boosting are available. The agent chooses from them and never builds its own ranking.
No decision tree, no routing rules to maintain
Act VI · Ontology
Systems outside Synerise plug in as tools
check_chat_availability asks the live chat system whether anyone is online. No one is, so the agent says so plainly and doesn't promise a handoff that won't happen.
Configure once
Tool name, when the agent should use it, URL and credentials. No connector to build, no waiting for a release.
What the customer gets
The customer gets an honest answer, not empty promises.
Example: checking consultant availability
Hands off the conversation to your processes
You define tools that return structured actions to your UI. You handle them your way.
Handoff to a consultant
A complaint isn't a shopping question. The agent shares the form, and when the customer asks for a person, it returns an action your UI already handles.
Straight to the right page
Travel deals, installments, warranty, returns. For questions outside the catalog, the agent answers with a link to the right page and doesn't guess.
Handoff to a consultant and links to the right pages
Conversation history
Customer and business see the same history
What the customer sees
A history tab in the widget. Yesterday's searches are one click away, and the conversation picks up where it left off.
What the business sees
Every conversation in the workspace, searchable and viewable in full. Titles generated from each thread show what customers really come for.
Customer widget and workspace
Act VII · Deployments
Live in production
Not a demo. Not a pilot. Right now, the agent is talking to customers in stores.
Customers click on over half of all responses
58.11%
Click-through rate
Product clicks as a share of all responses
5.44%
Conversion
Purchased products as a share of those shown
99.16%
Error-free responses
Status “ok”
Based on results from production deployments