By AI agents deployed in enterprise environments consistently fail at two things: finding the knowledge they need across fragmented, unstructured data sources, and operating within the access boundaries the organization requires. Most MCP Gateway architectures force agents to make a separate tool call for every source at runtime, flooding the context window, increasing token costs, and compounding the probability of error. The result is an agent that is slower, more expensive, and less accurate the more data sources you add.
In this interview on TFiR, Jean Lafleur, Co-Founder and Chief Operating Officer at Airbyte, walks through how Airbyte Agents addresses enterprise data access, semantic search over unstructured data, context store architecture, workspace-level governance, and the operational path from human-in-the-loop oversight to increasing agent autonomy.
Guest: Jean Lafleur, Co-Founder and Chief Operating Officer at Airbyte
Show: TFiR
Here is what every data engineer and AI platform team needs to know.
Technical Deep Dive
Q: What are the three core problems enterprises face when deploying AI agents against internal data?
Jean Lafleur, Co-Founder and Chief Operating Officer at Airbyte, identifies three distinct enterprise-level problems with AI agent deployments. The first is data sovereignty: enterprises need their data to remain within their own environment and cannot accept it leaving to a third-party control plane. The second is access limitation: agents are severely constrained if they cannot reach the data they need to reason over. The third is permissioning: organizations must know which employee or agent accessed which data, and any write action must carry a traceable identity for audit purposes.
“Who has access to this data? Because while an employee or an agent from that employee shouldn’t have access to some data, and also the right permissions so that if an agent changes a field somewhere in a tool, we need to know who has done that and on whose behalf.” — Jean Lafleur, Co-Founder and Chief Operating Officer, Airbyte
Q: How does Airbyte’s Flex hybrid deployment model solve data sovereignty for enterprises?
Airbyte’s Flex model splits the architecture so the control plane is hosted by Airbyte while the data plane remains entirely within the customer’s environment. Lafleur states that Airbyte never has access to the customer’s data, only the metadata required to manage and support data pipelines. This separation has removed the sovereignty objection for enterprise customers and applies equally to agent workflows as it does to traditional data replication.
“The control plane is hosted by Airbyte, but the data plane stays within your environment. We have never access to your data, only some of the metadata that helps us control and provide you support for your data pipelines.” — Jean Lafleur, Co-Founder and Chief Operating Officer, Airbyte
Q: What is wrong with the MCP Gateway architecture for multi-source AI agent queries?
MCP Gateway architectures require at least one round trip tool call per data source at runtime. A single query spanning Salesforce, HubSpot, and Zendesk, for example, generates multiple sequential tool calls, each returning a raw payload that floods the context window. The result is high latency, high token cost, and a context window degradation problem Lafleur describes as context rot: the more data sources that are consulted, the less effectively the agent can reason over its accumulated context.
“The raw payloads really flood the context window, and so your agent becomes less and less smart because the more context window you have context rot in there.” — Jean Lafleur, Co-Founder and Chief Operating Officer, Airbyte
Q: How does Airbyte’s context store architecture reduce tool calls and token usage for AI agents?
Airbyte’s context store prematerializes data context before any agent query is issued. When an agent submits a request, the context store has already performed entity resolution across all connected sources, so the agent receives a complete, reconciled answer in a single tool call rather than assembling one at runtime. Lafleur reports this architecture produces 40 percent fewer tool calls compared to MCP Gateway and uses ten times fewer context tokens compared to raw API approaches, and half the context tokens of MCP Gateway.
“We prematerialize the context. With our context store, we do all entity resolution, and you have a request, you just go to the context store. We know which entity, we have all the information, and it’s just one tool call at that point to give you a complete comprehensive answer.” — Jean Lafleur, Co-Founder and Chief Operating Officer, Airbyte
Q: What deployment options exist for the Airbyte context store?
The context store can be deployed in two configurations. It can run as a standalone service, either cloud-hosted or within the customer’s own environment, or it can run as a layer on top of an existing data warehouse. Lafleur notes that many enterprises have made significant investments in their warehouse as a source of truth for structured data, so the context store is designed to be compatible with and complementary to that existing infrastructure rather than replacing it.
“We needed to be comfortable with that. Companies want to leverage their warehouse. They spent a lot of time, a lot of investment in their warehouse as a source of truth for structured data.” — Jean Lafleur, Co-Founder and Chief Operating Officer, Airbyte
Q: How does semantic search in the Airbyte context store differ from traditional keyword search for agent use cases?
Traditional keyword search matches exact wording, which fails when agents phrase queries in natural language or when the relevant documents use different terminology. The Airbyte context store indexes unstructured documents such as call transcripts, meeting notes, and support tickets, and enables agents to query them semantically. An agent asking about pricing concerns across customer conversations will retrieve topically relevant documents regardless of exact phrasing, covering all indexed sources in a single pass.
“Semantic search is really the biggest unlock for agents in terms of finding information on unstructured data.” — Jean Lafleur, Co-Founder and Chief Operating Officer, Airbyte
Q: How does the context store handle unstructured data that does not live in a data warehouse?
Unstructured documents such as call transcripts, Linear tickets, and meeting notes are ingested and indexed directly into the context store. Because the context store is also compatible with the data warehouse layer, agents get a unified view: structured data from the warehouse and unstructured documents from the index are both accessible through the same natural language interface. This design choice was made specifically because unstructured data typically does not live in the warehouse and would otherwise remain invisible to agents.
“What we’ve seen is that companies could just put those unstructured documents as transcripts that we index on the context store, and the context store is still compatible and covers your data warehouse for your structured data. At that point your agent can see everything.” — Jean Lafleur, Co-Founder and Chief Operating Officer, Airbyte
Q: What governance controls does Airbyte provide to prevent agents from accessing data they should not reach?
The primary governance mechanism is workspace isolation. Organizations can create multiple context stores within separate workspaces, ensuring that sensitive data sets such as employee records are never mixed with context stores used by go-to-market teams. Access to each workspace is controlled at the employee level. A second layer applies through agent connectors, which handle fetch, search, and write operations directly against APIs and can enforce permissioning at the point of action rather than at the context store level.
“Each workspace gives their own access to which employee. So that’s one side of it. And we can implement permissioning at the agent connector level when you fetch, search, and write.” — Jean Lafleur, Co-Founder and Chief Operating Officer, Airbyte
Q: How should enterprises balance AI agent autonomy against the need for human oversight?
Lafleur recommends starting with a human always in the loop and progressively removing that requirement as confidence is established for specific task types. At Airbyte, the AI Factory used for connector maintenance begins with every AI-generated pull request reviewed by a human engineer. As the system demonstrates consistent accuracy for particular issue categories, those categories can graduate to autonomous execution. This incremental confidence model is what Lafleur observes across enterprise AI deployments broadly.
“At the beginning you always want a human in the loop. And once you start building confidence, then you can remove that human in the loop. But most likely you will start with a human in the loop.” — Jean Lafleur, Co-Founder and Chief Operating Officer, Airbyte
Q: How is Airbyte merging its data replication product with Airbyte Agents, and how large is the user base?
Airbyte Agents was initially developed separately from the core data replication platform. However, enterprise adoption showed significant overlap in audience: the same teams managing data pipelines also want to enable agents over that data. The current product direction is to merge both surfaces so existing customers can query their pipelines and data directly, create context stores on top of their warehouse, and enable agent connectors without switching products. Lafleur states Airbyte serves between 28,000 and 30,000 companies across open source, cloud, and Flex deployments.
“All the customers that were using Airbyte to move data can chat with the data pipelines, but also chat with their data. So that’s the main focus right now, to really merge those two and expand the Airbyte Agents capabilities to a lot more users.” — Jean Lafleur, Co-Founder and Chief Operating Officer, Airbyte
Q: How did Airbyte’s open source origins shape its enterprise deployment model?
Airbyte launched as an open source project and became the open source standard for data connectors rapidly. The self-hosted nature of open source shaped the technical culture toward local and on-premises deployment, which directly informed the design of the Flex hybrid model. The open source community has contributed more than 35,000 custom connectors built by users for their own needs, layered on top of the more than 600 connectors Airbyte itself maintains.
“Being open source and self-hosted, that’s how we developed that Flex offer. And at that point our goal is that wherever your data is, you can have access to it.” — Jean Lafleur, Co-Founder and Chief Operating Officer, Airbyte
Resources and Documentation
- Airbyte, open source data movement platform with hybrid deployment, 600-plus connectors, and Airbyte Agents for enterprise AI data access
- Airbyte Documentation, official documentation covering connectors, deployment options, Flex hybrid model, and context store configuration
- Airbyte on GitHub, open source repository for the core Airbyte platform and community connector contributions
***
👇 Click to Read Full Raw Transcript
Swapnil Bhartiya: Today everybody is on the AI bandwagon. Enterprises are moving from let’s experiment with AI to we need this feature, we need this in production. And that’s where friction starts to emerge because now you are dealing with actual enterprise data, actual problems. And when it comes to agents, in most of the cases they cannot even find the knowledge they need, which is buried in wikis or Slack or a lot of other unstructured data that companies have. And when they do get access to this knowledge, there is no way to control what they can access and what they can do with that knowledge. It’s like giving someone the keys to the entire building without knowing which doors they will open or what they will take from those rooms. They need guardrails. An Airbyte just solved both problems at once with their latest agents platform update. They have built semantic search and fine grained governance into the platform so agents can find what they need without breaking your security policies and governance practices. And today we have with us Jean Leffner, Co founder and CEO at AirByte. First of all, John, it’s great to have you on the show.
Jean Lafleur: Thank you Satya, great to be here.
Swapnil Bhartiya: It’s my pleasure. As I was talking about this problem, access to data and then what you can get, what you can do with that data. Let’s start with that problem area. When it comes to enterprises, AI in production, what kind of friction? You see, at one hand, agents do need access to knowledge, otherwise they won’t be able to perform tasks at the same time. Now they also have right access, they can do whatever they want. So you also need to govern what they can or cannot do. So let’s talk about the problem area, then we’ll talk about what airbyte is doing.
Jean Lafleur: At the enterprise level we see three main problems. The first one is sovereignty, like data sovereignty, a lot of the tools, the data is within the enterprise environment and they want to have the data stay there. So that’s one problem. The second one is yes, agents are very limited if they don’t have access to the data. And as you mentioned, the third one is really permissioning. And permissioning meaning, okay, who has access to this data? Because while an employee or an agent from that employee shouldn’t have access to some data and also the right permissions so that if an agent changes a field somewhere in a tool, we need to know who has done that and whose behalf. And those three problems are pretty complicated. And Airbyte agents, that’s what we are trying to solve.
Swapnil Bhartiya: Initially when I was trying to introduce, I Focus on two problems. One was access to data and second is what they can do with the data. And you also talk about sovereignty, which is actually becoming very important because unless you are using everything local LLMs, then it’s not a problem. Then you have to just focus on which box. But if you’re using API, if you’re a frontier model or open source model, then you do lose. So talk about that, that aspect Also when we talk about AirByte and of course your air agents platform, how do you tackle these problems? And then we can talk about how the latest update is actually making things better.
Jean Lafleur: What we have is a hybrid deployment solution that we name Flex. And that means the control plane is hosted by airbyte, but the data plane stays within your environment. So we have never access to your data, only some of the metadata that helps us control and provide you support for your data pipelines. And that’s the way we’ve seen no enterprises having any issue with this. That’s a big unlock for them and same for agents, agent workflows. So that’s how AirByte solves the problem by having that hybrid solution. We provide the best, meaning you have the data plane stays in your environment. And also we can provide the best experience because we can provide you support.
Swapnil Bhartiya: Is it possible for you to walk our audience? Because you know, the fact is that stating one thing is easy, but actually walking through that, this is how it works, this is how we ensure. I mean you talked about sovereignty, but let’s also talk about giving access to data and also what data they can access and then what they can do with the data. The governance model as well.
Jean Lafleur: I think it goes back to the history of airbyte. We started as an open source project, we became the open source standard very, very fast. And open source is self hosted. And with the open source we managed to, you know, expand the number of connectors we were building. We’ve got more than 600 and actually our users have built more than 35,000 custom connectors for their own needs. And that is locally or also cloud hosted. And so being open source and Safari, that’s how we developed that Flex offer. And at that point our goal is that wherever your data is, if it’s cloud hosted and it’s okay for you to be cloud hosted, go for it, or if it’s something sensitive in your environment, you can have access to it. And with the agents. The difference is you have heard, I’m sure about MCP gateway where agents have access to the tools through an MCP Gateway. And the issue with the MCP gateway is that for every source you have at least a round trip. So let’s say you’re asking what is the latest with that account? And that account lives in Salesforce, Hotspot, Zendesk, in many tools. At that point, you will have to have many tool calls for each source. Then you reconcile all of that during runtime. And that’s very long, very costly in terms of tokens, context token. And you also, it’s usually plausibly wrong in the sense that by doing all of that, you’re prone to making errors. In addition to that, the raw payloads really flood the context window. And so at that point your agent becomes less and less smart because the more context window you have context rot in there. And what we solve with airbyte agents is that we prematerialize the context. So with our context store, that can live by itself or live on top of your warehouse. At that point we do all entity resolution. And so you have a request, you just that request, go to the context store. We know which entity, we have all the information and it’s just one tool call at that point to give you a complete comprehensive answer. And at that point it’s faster, a lot less context tokens use. So that’s pretty important for agents when you have multiple tasks.
Swapnil Bhartiya: Can you also walk us through your context store? What makes it unique when it comes to AirBud agents?
Jean Lafleur: Our context store can live so outside of your warehouse by itself. And it can be either cloud hosted. If it doesn’t matter to you, it can be hosted in your environment. And it is dedicated to how agents interact with data. Meaning that. So we do entity resolution. You can create different workspaces with different context stores if you want to have to limit access to some context and data. And it works as natural language. There’s a natural language interface for the agents to interact with the data. And that means. And so that’s one thing. The second thing is also we build a context source that it’s compatible with the data warehouse, your data warehouse. So it would be then a layer on top of it. And what we do in the context store, we also index all the unstructured document uncharted data, so the transcript, whatever tickets you have in linear and so that you can do semantic search. And what is interesting with semantic search is that agents think like this. You want something about, okay, what’s the latest on the pricing? They look at all the gong calls with the talks about pricing. Nice. Especially that wording and what happens is that where usually your unstructured documents is not on your data warehouse. So what we’ve seen is that companies could just put those unstructured documents as transcripts that we index on the Context Store and the Context Store is still compatible and covers your data warehouse for your structured data. And at that point your agent can see everything. And that’s how we built the Context Store this way because we understood that some companies want to leverage their warehouse. They spent a lot of time, a lot of investment in their warehouse and as a source of truth for like for the structured data. So we needed to be comfortable with that.
Swapnil Bhartiya: Jean, is it possible for you to share some real example of how semantic search helps customers?
Jean Lafleur: Semantic search is actually how agents think they use natural language questions. It could be, you know, what customer concerns were discussed at last week’s planning meeting where we discussed pricing strategy across documents or Milinos, what are customer conversations where there were pricing concerns and so they look at the topic and find the topics across every conversation that has been dexed or every meeting notes all of that. It’s very different from just search where you just look at the exact wording, how it’s being used. Semantic search is really the biggest unlock for agents in terms of finding information on unstructured data. And so yeah, there’s so many examples, engineering issues, login failures, any topic, any teams could have some.
Swapnil Bhartiya: One of the problem is also challenges, also cost. Can you talk about if there is any impact on cost that your customers are seeing with these new capabilities?
Jean Lafleur: Yes, because they use a lot less tokens. That’s actually pretty significant depending whether they use we can compare to raw API calls and or MCB Gateway. But for instance for the MCB gateway we see 40% less tool groups like fewer tool groups in terms of token usage compared to raw APIs we had 10x. So with 10% in terms of context tokens used compared to raw APIs and compared to MCP Gateway, we are twice less context tokens used at that point. And so those are pretty significant in terms of cost but also in terms of context routing where your agent becomes less smart while the context window is being filled.
Swapnil Bhartiya: I also want to talk a bit about the security part as well, which is what kind of controls are you providing so agents don’t overstep their bounds when it comes to accessing data. The irony is that agents will hallucinate no matter how many guardrails put they will do whatever. We have seen so many stories, but given that what are you building? What kind of guardrails are you putting in place?
Jean Lafleur: The first guardrail is workspaces where you can have several contact stores within your company and so that you can make sure for instance your employee data is is never part of the mix of other context store used by the go to market teams, for instance, that’s the first thing. So each workspace gives their own access to which employee. So that’s one side of it. The second is at Ambu we have both the context store, but we also have what we call agent connectors that can fetch, search and write. And when they do that they just go to the API and do the course. And we can implement permissioning at that level, but it won’t be at the context store level, it will be when you fetch search and write. And we can. Something we considering for instance is if we could do that, we could check the permissioning before you have access to context store. But right now what we’re using today is really the workspace.
Swapnil Bhartiya: And you also talk about of course human oversight. How are you thinking about the balance between autonomy that we want to give to agents? Because that’s the whole point of automation. At the same time keep humanly loop. Human oversight is also there because this is a complex as we are giving them more access. More right access. How do you balance that human versus autonomy?
Jean Lafleur: I would say let’s take an example. At airbyte for instance, a lot of our connector maintenance issues right now is done through AI and we’ve built AI factory for that. It’s taking a long time because maintaining pipelines data plan is hard. But what we do is at the beginning this AI would submit a pr but there’ll be a human in the loop to check the PR always. Now then with time we build confidence, we make it better. We see that for certain issues it always works perfectly. At that point we can remove the human in the loop. And I think that’s pretty representative of what is happening in every enterprises. At the beginning you always want a human in the loop. And once you start building confidence, then you can remove that human in the loop loop. But most likely you will start with a human in the loop.
Swapnil Bhartiya: How organizations are embracing, adopting of course AirByte agents, the whole context tool. Just, just, just give us, you know how airbyte is growing in this space.
Jean Lafleur: So right now airbyte agents is, has been separate from airbyte. The historical of airbyte that was focused on data replication. And what we see is that at the beginning we’re not sure that they would be the same audience, but in a lot of enterprises it’s the same audience. So one effort, whether we’re doing right now, is actually merging the two products so that all the users, all the customers that were using Ambide to move data at that point, can chat with the data pipelines, but also chat with their data. Because with the agent connectors you can fetch search right directly, but also create a context store on top of their warehouse or outside of their warehouse so that they can enable the agents. So that’s the main focus right now, to really merge those two and to expand, I’d say the abate agents capabilities to a lot more users. Because in the end I think we got more than 20, between 28 and 30,000 companies in the end using AirByte overall between open source, cloud and Flex.
Swapnil Bhartiya: John, thank you so much for breaking down how AirByte is bringing semantic search and governance together, making it easier for teams to actually use AI freely without worrying about what access it should have or what data it should get. So thank you for that. And also those who are watching developer teams, please go ahead and check airbuy.com to learn more about how they are building these technologies. Once again, Jean, thank you so much and I look forward to chatting with you again.
Jean Lafleur: Thanks Swapnil. Looking forward to the next time.
