Uvi Gives AI Agents Enterprise Context, but the Hard Part Is Keeping Decisions True

Uvi gives AI agents enterprise context through a newly funded system that models decisions, not just documents. The Tel Aviv startup says this missing layer causes agents to waste time, consume tokens, and act on fragmented evidence.

The company has raised a $7.5 million seed round led by CT Innovations, with Cardumen Capital, Manu VC, and 91 Ventures participating. Uvi plans to expand its United States team, product development, and market operations.

The announcement matters because connecting an agent to Slack, Jira, GitHub, and internal databases does not automatically create organizational understanding. Model Context Protocol, or MCP, can provide access to those systems. Retrieval software can find relevant files. Neither approach necessarily reconstructs why a decision happened or whether it still applies.

Uvi wants to own that reconstruction layer. Its product builds what the company calls a Decision Graph, which links decisions to owners, triggers, blockers, dependencies, and supporting evidence.

That puts Uvi against a common enterprise strategy: connect more sources, retrieve more text, and let the model infer the rest. Uvi argues that this approach repeatedly charges companies for rediscovering relationships their employees already understand.

The pitch is timely, but its hardest questions remain unanswered. Decision graphs must stay current, preserve permissions, represent contradictory evidence, and distinguish a valid decision from an abandoned one. If those controls fail, structured context can make an incorrect answer look more authoritative.

Uvi Gives AI Agents Enterprise Context Through a Decision Graph

Uvi is selling a persistent map of organizational reasoning rather than another interface for searching company documents.

According to the original enterprise context report, Uvi was founded in 2023 by CEO Gilad Salinger and CTO Dror Wolmer. The business initially operated under the name Naboo.

Its central argument starts with a familiar enterprise problem. Important knowledge rarely lives in one system or one carefully maintained document.

A product decision might begin in a Slack conversation. The implementation appears in a pull request, while Jira records the task as completed. A later incident produces a rollback, but the original ticket remains closed.

A conventional agent can retrieve each artifact. It still must determine which record reflects the current state and why the team changed direction.

Uvi says its Decision Graph represents decisions as first-class objects. Each decision can include an owner, triggering event, blockers, dependencies, status, and links to supporting material.

That differs from treating the organization as a collection of semantically similar passages. A vector index represents text according to mathematical similarity. It does not inherently represent responsibility, causation, approval, or the sequence behind a decision.

Uvi describes a two-tier architecture behind the product. The first tier is a Topic Graph, which indexes information across company systems and resolves connections among people, repositories, tickets, and services.

The second tier is the Decision Graph. It applies an organization’s private definitions to the relationships discovered below it.

For example, one company might define “shipped” as code merged into a production branch. Another might require full customer availability, completed monitoring, and formal approval.

Those definitions affect the answer to a basic operational question. An agent cannot reliably identify a blocked release unless it understands what release readiness means inside that company.

Uvi says specialists work with customers to encode those concepts and decision gates. The graph then becomes available to agents through GraphQL or an MCP server.

GraphQL supports structured queries from applications that already know which fields and relationships they need. MCP supplies a standardized connection through which an agent can express intent and request relevant context.

The company says the product can connect with systems including GitHub, GitLab, Jira, Linear, Slack, Confluence, Notion, Datadog, and Splunk. It also supports internal services that contain business-specific state.

The design targets questions that require several joins. A release agent might ask which approvals remain open, who owns them, and what event will clear each blocker.

An incident agent might ask why a feature flag changed and which customer commitment influenced that choice. Search can find mentions of the flag. A decision graph tries to reconstruct the chain that produced its current state.

That distinction explains why Uvi presents itself as infrastructure beneath existing agents. It is not asking enterprises to replace Claude, Gemini, GPT, or their internal agent interfaces.

Instead, Uvi wants companies to build their organizational context once and expose it to multiple models. The company’s claim is that models will change faster than the underlying meaning of company systems.

This is an appealing architectural boundary. It separates model selection from the work of defining ownership, operational vocabulary, and decision history.

It also creates a significant responsibility. The shared context layer becomes a dependency for every agent that relies on it.

More Connectors Do Not Explain Why a Decision Happened

The pressure falls on teams that have treated data access as the same problem as organizational understanding.

Anthropic introduced Model Context Protocol in November 2024 as an open standard connecting AI systems with tools and data sources. MCP reduced the need to design a separate integration pattern for every assistant and application.

That standardization solved an important interface problem. An agent can discover available tools, request information, and invoke supported actions through a common protocol.

MCP does not decide what a company’s internal terms mean. It also does not determine whether a closed ticket was superseded by a rollback discussed elsewhere.

Uvi’s argument begins where the connector ends. Access gives the agent more evidence, while the reasoning layer attempts to organize that evidence around operational intent.

This difference becomes important when an agent moves from summarization to action. A search assistant can return several documents and let a person resolve their inconsistencies.

An autonomous agent must choose which record to trust before changing a configuration, escalating a blocker, or updating a customer workflow. That choice requires both context and authority.

Traditional retrieval-augmented generation, or RAG, retrieves material related to a query and supplies it to a model. It works well when the answer exists inside a document that the retrieval system can identify.

Enterprise tasks often depend on relationships that no document states directly. The answer emerges from a ticket, a code change, a conversation, telemetry, and the company’s definition of completion.

Microsoft’s GraphRAG project illustrates the industry’s broader move beyond isolated text chunks. GraphRAG combines extracted text, network analysis, prompting, and summarization to build richer representations of a dataset.

Microsoft also supports graph-based context providers that retrieve connected subgraphs instead of isolated passages. Its Neo4j integration combines vector, full-text, hybrid search, and graph traversal.

Those developments validate part of Uvi’s premise. Relationships can provide information that similarity search alone misses.

However, they also show that Uvi is entering an active technical field rather than creating an uncontested category. Enterprises can combine graph databases, agent frameworks, search platforms, and custom data pipelines.

Uvi must therefore prove that decision modeling delivers more value than a well-designed GraphRAG system. It must also show that its managed layer costs less to operate than an internal alternative.

The competitive divide is not simply Uvi against RAG. Most serious enterprise systems will continue using several retrieval methods.

Document search remains useful for policies, reports, and specifications. Vector retrieval remains helpful when users do not know the relevant vocabulary. Conversation memory can preserve preferences and recent interactions.

A decision graph addresses a different need. It should identify the choices, dependencies, and responsibilities that determine what an agent can safely do next.

The likely architecture combines these layers. An agent could use search to locate supporting material, a graph to navigate dependencies, and source systems to verify live status.

This combination places pressure on enterprise search vendors and internal platform teams. Returning relevant links no longer satisfies every agent workflow.

Search products will need richer relationship models, stronger permission enforcement, and outputs designed for machines taking actions. Internal teams face the same demand while maintaining custom connectors and data pipelines.

Uvi’s opportunity rests on making that complexity reusable. Its challenge is proving that a specialized vendor can understand each customer’s private operating language accurately enough.

The Mechanism Depends on Live Joins, Intent, and Permissions

Uvi’s technical advantage will depend less on graph terminology and more on whether its relationships remain current and correctly scoped.

The product’s proposed workflow begins by indexing information across operational systems. It identifies explicit links, such as a Jira ticket referenced inside a pull request.

It also tries to resolve implicit relationships. These can include different names for the same service, employees represented by several accounts, or internal phrases that describe one release process.

Entity resolution means mapping those references to consistent identities. Without it, the graph can treat one person, repository, or feature as several unrelated objects.

The next step introduces company-specific meaning. Uvi says a specialist works with the customer to define concepts such as owner, blocker, deploy, approval, and shipped.

This human involvement is notable. The system does not assume that a general model can infer every operating rule from raw communications.

The approach resembles the work required to build a reliable knowledge base. Useful organizational memory depends on definitions, provenance, maintenance, and access boundaries.

Uvi then makes the resulting graph queryable. An agent can request a structured chain containing decisions, owners, statuses, dependencies, and supporting evidence.

That output differs from a ranked list of documents. It gives the agent a smaller operational object that can guide its next step.

Consider a release that appears complete in a project tracker. A code repository shows the corresponding pull request merged, but monitoring data records an error increase after deployment.

A later conversation says the rollout stopped at partial availability. The decision graph should connect those events and prevent an agent from describing the release as complete.

Intent also changes the required context. Reviewing the release architecture requires design documents, dependency information, and historical tradeoffs.

Diagnosing a failed rollout requires recent telemetry, rollback details, changed code, and current ownership. Both queries mention the same release, but they need different evidence.

Uvi says its layer narrows the graph according to the task. If that works, the agent receives less irrelevant material and spends fewer model turns exploring disconnected sources.

The company reports substantial performance gains on its website. Uvi says it won 97 of 100 head-to-head queries against an MCP-enabled GPT system in a Global-e benchmark.

It also claims twice the speed of RAG baselines and an eightfold reduction in query cost compared with a do-it-yourself agent using the same model. These are vendor-reported results, not independent benchmarks.

The published material does not provide enough methodological detail to generalize them across companies. Workload selection, graph preparation, answer grading, source complexity, and baseline configuration can materially affect the outcome.

Still, the benchmark points toward the correct unit of measurement. Enterprises should compare systems using correct answers, latency, total model calls, and human verification time.

A smaller token count means little if employees must investigate every answer. High retrieval accuracy also means little when an agent misses the current decision state.

Security is the other central mechanism. Uvi says it can run on-premises or inside a customer-controlled virtual private cloud.

The company also says permissions are enforced during retrieval. An agent should only receive information available to the person behind the request.

That control matters because graphs create indirect paths to sensitive facts. A user might lack access to a confidential document but infer its contents through connected nodes.

A secure deployment must protect the underlying material and the relationships derived from it. Even the existence of a hidden project, owner, or dependency can disclose sensitive information.

Source-level permissions can also conflict. A Slack channel, repository, and project tracker might assign different rights to the same person.

The reasoning layer must calculate the valid intersection without leaking restricted context. It must repeat that calculation whenever roles, group membership, or source permissions change.

Uvi’s enterprise case therefore depends on four capabilities working together: fresh ingestion, accurate entity resolution, task-aware selection, and permission-preserving retrieval. Weakness in any one layer can undermine the final answer.

A Decision Graph Can Preserve Errors as Easily as Knowledge

Uvi’s most important risk is that structured context can turn uncertain organizational memory into confident machine instructions.

Corporate decisions are rarely clean records. Teams disagree, exceptions accumulate, and temporary workarounds survive beyond their intended period.

A product manager might approve a release in one channel while a security lead blocks it elsewhere. The final outcome may depend on an undocumented conversation between executives.

A graph must represent that uncertainty instead of forcing every artifact into one definitive state. Otherwise, the output will look coherent while hiding a real conflict.

The company also says its system reconstructs reasoning that was never explicitly recorded. This is useful, but it introduces an inference boundary that buyers should examine carefully.

A relationship supported by a ticket reference is different from a model’s interpretation of several conversations. Both might appear as edges in a graph, yet they carry different evidentiary weight.

Enterprises need to know which links came directly from source records. They also need confidence scores, timestamps, and review paths for inferred relationships.

Decision validity changes over time. An emergency exception can expire, a service can move between teams, and an approval policy can be replaced.

The graph must retain history without confusing an old rule with the current one. It should also explain why the current state superseded earlier evidence.

Fresh ingestion alone does not solve this problem. A new message can be current while remaining unofficial, speculative, or incorrect.

Authority matters alongside recency. The system must distinguish an accountable decision-maker from a participant discussing possible outcomes.

These problems make evaluation difficult. A benchmark can test whether the system returns the expected answer, but the expected answer may itself be disputed.

Customers should include contradictory cases, revoked decisions, confidential relationships, and deliberately stale records in their evaluations. Easy questions about well-documented projects will not reveal the system’s failure modes.

They should also test how Uvi handles missing evidence. A safe system should expose uncertainty rather than invent the undocumented reason behind a decision.

Human review remains necessary for consequential actions. A graph can reduce exploration, but it should not eliminate approval controls for production changes, financial commitments, or security decisions.

This is where a searchable engineering base remains valuable. People need access to the underlying artifacts when they review an agent’s proposed action.

Vendor lock-in creates another concern. Uvi argues that companies should own one context layer that works across models.

That portability is valuable only if customers can export graph data, definitions, provenance, and evaluation records. Switching models is easy compared with rebuilding years of encoded operating knowledge.

The deployment model also needs scrutiny. Running inside a customer environment reduces some data exposure, but it does not automatically guarantee secure behavior.

Buyers should examine administrative access, connector credentials, logging, update processes, encryption, and incident response. They should also verify permission behavior across every integrated source.

Uvi’s benchmark claims deserve the same disciplined review. The reported results indicate potential, but they do not establish performance across different organizations or workflows.

A customer with consistent ticketing practices presents a different environment from one that relies on private messages and informal approvals. Graph quality will reflect those operating habits.

This limitation does not invalidate the product. It clarifies what Uvi is actually selling.

The company is not merely installing a database. It is helping an enterprise formalize how decisions, ownership, evidence, and permissions connect.

That process can produce value even before an agent acts. It can also expose how much essential knowledge remains informal or inaccessible.

The risk is treating the resulting graph as objective truth. It is better understood as a maintained operational model with explicit sources and unresolved uncertainty.

Three Signals Will Show Whether Uvi’s Context Layer Works

Uvi now needs evidence that its decision model stays accurate in production, transfers across agents, and survives enterprise security review.

The first signal is independent evaluation. Uvi has published customer examples and vendor-run comparisons, including its reported Global-e result.

The next step is a reproducible benchmark with documented questions, baselines, grading rules, latency, model usage, and graph preparation. External validation would make the efficiency claims easier to assess.

The benchmark should test more than retrieval. It should include questions requiring several systems, changing states, conflicting evidence, and permission boundaries.

A meaningful result would strengthen Uvi’s claim that structured decisions outperform connector-driven exploration. Weak results would suggest that custom graph preparation contributes more than the general architecture.

The second signal is sustained customer adoption beyond controlled pilots. Uvi identifies Melio and Global-e as production users and displays statements from their executives.

Future case studies should explain which agents use the graph, which actions they support, and how often employees override their conclusions. They should also report maintenance work required after deployment.

Adoption across multiple agent interfaces would support Uvi’s model-agnostic position. It would show that one context layer can serve coding assistants, workflow agents, and internal applications.

Limited use within a single search experience would weaken that claim. It could indicate that the graph performs well for knowledge discovery but not for operational action.

The third signal is how larger platforms respond. Microsoft, Anthropic, cloud vendors, graph database companies, and enterprise search providers already occupy adjacent layers.

They can combine connectors, graph retrieval, model hosting, permissions, and agent frameworks. Some customers will prefer that integrated stack over another specialized vendor.

Uvi must establish a durable advantage in decision modeling, deployment speed, or operational accuracy. Connector count alone will not protect its position.

The company’s $7.5 million seed round gives it resources to pursue that proof. It does not settle whether the reasoning layer will become a distinct enterprise software category.

Buyers should watch for clearer documentation about graph updates, inferred relationships, access controls, and data portability. Those details matter more than a polished demonstration.

They should also ask how the system behaves when evidence conflicts. A useful answer must show sources, preserve uncertainty, and identify who can resolve the disagreement.

For developers, Uvi’s arrival reinforces a broader architectural lesson. Tool access, retrieval, memory, and organizational context solve related but separate problems.

MCP can connect an agent to company systems. RAG can find relevant text. Graph retrieval can expose linked entities. A maintained decision model can add ownership, causation, and operational state.

No single layer guarantees a correct action. Production agents need all these capabilities, plus evaluations and human controls.

For enterprise buyers, the immediate question is practical. Identify one workflow where employees repeatedly reconstruct the same decision chain across several systems.

Measure the current model calls, delay, error rate, and review effort. Then test whether a maintained graph improves the complete workflow rather than a curated answer set.

Uvi gives AI agents enterprise context by converting scattered records into a structured account of how work moves. Its future depends on whether that account remains trustworthy after the demo ends.

Before letting an agent act on organizational memory, ask three questions: Can it show the evidence, can it respect every permission, and can it admit uncertainty?