The data layer agents run on.
A demo takes a week. Production takes two years, because agent projects die in the data underneath the model. Kleio is that layer: a memory you can trust, a catalog agents can reason over, and one orchestration path for every brand, market and channel.
The model is the easy part.
Every company building a conversational agent for a complex sale hits the same four walls. None of them is about the model.
Memory you cannot trust
A passing click overwrites what the buyer said outright, and nobody can say why. “Why did it say that?” is a GDPR and EU AI Act question.
A catalog the agent cannot reason over
Products × options × dates is a price surface of millions of rows, split across several systems. Offers stop matching the request.
Logic that forks per brand
Brands × markets × channels. Every variant becomes another copy of the prompt, the rules and the bugs.
State scattered across stores
Every copy of a buyer’s data is one more retention window, one more consistency bug, one more erasure to chase.
We built the layer underneath the conversation.
One platform, live in travel, real estate and automotive. Each layer answers one of the four walls.
Three stores, one orchestrator, one config system. Each brand is configuration, not code.
Keep what the buyer said. Compute what to believe.
Built for a buyer who says one thing on Monday and clicks on another on Wednesday. Structured and free-form memories, written live during the conversation, not at session end.
A message, a click, a CRM record. Appended to a log, never edited.
One value for one field, with its source and the exact words it came from.
Computed from claims: a winner, a margin, and a state, settled or contested.
Who is trusted more
A nightly CRM sync can never overwrite what the buyer said this afternoon. Close calls are not guessed: the assistant asks.
A buyer changes their mind
Turn 5: beach leads by 0.07, inside the 0.15 contested margin, so the assistant asks. Turn 6: the buyer confirms and the belief settles.
An agent sells only what it can find, price and explain.
Typed columns, brand-specific attributes, free text, geography and a price surface. No single search engine handles all five, so we built one joint system.
The agent returns offers that no longer satisfy the filter it was given, or quotes a price that has moved.
One joint system, one request, four steps
Search assumes the answer exists. The Product Store tells the agent what exists, what fits, and what it costs.
One turn, one path, one home for every byte.
What happens between the buyer’s message and the answer, and where each piece of state lives, for how long.
Where the state lives
Memory and products, working together.
The buyer rejects a first pick and says why. Beliefs are revised, the Product Store re-queries, and the pick that fits comes back, all in one conversation.
