Enterprises are increasingly trusting AI agents to answer questions that used to require a person who understood the business, not just the database. That trust assumes the agent has been given context that's accurate and consistent, and much of the market still can't guarantee that it has.
That gap is what Credible Data is built to close. Its platform runs on Malloy, an open-source, semantic modeling language, and stores a customer's semantic model as portable code rather than proprietary metadata. The company disclosed a $10 million seed round on Monday from Gradient, SignalFire and K5 Global, along with angel investors including Fivetran co-founder Taylor Brown and G2 co-founder Godard Abel.
The scale of the problem shows up in the numbers. In a VB Pulse survey of 101 qualified enterprises conducted in June, 57% traced a confidently wrong agent answer back to missing or inconsistent business context in the past six months, and 31% said it happened more than once. Only 25% run a governed context layer in production today, while 34% are still building one and 41% haven't started.
CEO and founder Kyle Nesbit met Malloy creator and Looker founder Lloyd Tabb while working at Google following its acquisition of Looker, as Tabb began building the language he has described as his career-defining project. Nesbit described the problem that eventually led him to leave Google and start Credible.
"It's the same pain point people have had for 30 years, the lack of governed data analysis," Nesbit told VentureBeat in an exclusive interview. "Now with AI, it's the same problem, but orders of magnitude more chaos and pain."
How Credible actually works #
Everything Credible builds runs on Malloy.
Malloy is an open-source semantic modeling language, but unlike most competing products, including Looker's own LookML, it is also a full query language. Nesbit argues that combination is what sets it apart.
"Malloy can do anything SQL can do," Nesbit explained. "It's a full semantic model language plus query language, designed so anything you can express in SQL you can express in Malloy."
Malloy was built at Google as an open-source language, MIT-licensed from the start. It has since been handed to the newly formed Malloy Foundation, a nonprofit Nesbit said Credible will be the first corporate sponsor of. The result is a semantic layer whose governance sits outside Credible's own commercial interests, structured so no single vendor controls the language customers build on.
Credible's platform is what the company has built on top of that foundation, broken into three parts.
Data context. This is the semantic model itself, the definitions, relationships and types that make raw data usable by both people and agents.
Workflow context. Nesbit calls these skills, the instructions for how an agent should actually use that data once it has it. Credible has open-sourced its default skill set and works with customers to extend it for their own workflows.
Portability. This runs through Malloy's dual role as both a modeling language and a query language. The company built an open-source server called the Malloy publisher to compile and serve Malloy models, and customers can run a lightweight version of it locally alongside a coding agent to build and test data models before deploying them.
"Once you've developed that model and data app locally, it's really just a package, an artifact, that you can push to Credible to serve with enterprise-grade controls, or bundle and serve yourself."
A build-versus-buy call from an early customer #
Peter Nummerdor, senior VP of product management at ad tech company VideoAmp, ran into this problem before Credible existed as an option. In 2023, VideoAmp built its own natural-language analytics chatbot on AWS Bedrock. The team never brought it to market. Data quality issues made the bot's answers unreliable, and the prompt and context engineering needed to keep it from hallucinating consumed effort that produced nothing reusable once the project ended.
That experience shaped how VideoAmp evaluated Credible when the option came up.
"We really felt that this was somewhere to buy, after the lessons learned earlier, to let us focus on our core business."
VideoAmp uses Credible to build semantic models of its customer-facing data. Credible sits between raw storage and the context window, so natural language queries find the right metrics and dimensions. It also powers a retrieval layer that returns labeled data into VideoAmp's agents. Nummerdor was direct about why that infrastructure work matters more than it looks like it should.
"You can build anything quickly that mostly looks good. Joe Schmoe competitor can give Claude a csv of data and build a pretty looking dashboard."
The competitive market for context #
Credible joins a growing list of different approaches to tackle the challenge of AI context.
Snowflake recently launched its Horizon Context and Cortex Sense offerings, which provide context defined by Snowflake and its customers. Amazon has its AWS Context service, a knowledge graph that learns from how agents use it rather than from manual curation. Database vendors including Pinecone, Couchbase, Redis and even Oracle all have context layers that are generally tightly coupled with the underlying database.
Credible differentiates with its open-source Malloy technology at the foundation, which offers the promise of data portability and expertise in the business intelligence space from Looker.
That openness however also might leave the company somewhat vulnerable, as it doesn't have a particularly deep competitive moat as any organization or hyperscaler could build with the open-source technology.
"We have distributed systems and machine learning experience you don't see concentrated in an organization like ours," Nesbit said. "But we're at the early innings of this whole AI revolution, and staying on the cutting edge is what's going to set us apart once the technology curve flattens out."
What this means for enterprises #
The semantic context layer is becoming a piece of production infrastructure data teams have to understand and manage. Credible's approach is one answer to a question every team evaluating this category now has to face directly: how much control to hand a vendor over the layer that defines what their data means.
Ownership is now a live procurement question. For years, the semantic layer sat inside whatever data tool a company happened to run and few people asked what happened to it on a platform switch. AI agents change the stakes, since that layer now drives decisions instead of dashboards, which makes portability a question worth asking a vendor directly rather than assuming.
The category is still unsettled enough that no answer is safe yet. Teams choosing an architecture now are choosing before the market has told them which approach wins.
A polished demo and a production-grade context layer look identical until scrutiny hits. Anyone can wire an LLM to a spreadsheet and produce something that looks finished. What actually separates the two only shows up once queries get complex and edge cases start generating support tickets, the exact lesson VideoAmp learned the expensive way before it looked at Credible.
"It is only when the rubber meets the road and you really analyze quality that your data stack matters," Nummerdor said.