There is a pattern playing out in boardrooms and operations meetings around the world right now. A team runs an AI pilot. It looks promising. They try to scale it, and it falls apart. MIT's research initiative studied over 300 AI deployments and found that 95% of enterprise AI pilots fail to deliver any measurable financial return. That is not a technology problem in the way most people assume.
The models, meaning the actual AI brains powering these tools, are genuinely capable now. The breakdown happens at the data layer. When an AI agent is asked to help with something, say, identifying which customers are most likely to leave, it needs to pull together account history, recent support tickets, payment records, and product usage signals. In most companies, each of those lives in a completely separate system. The agent has to make five or six separate queries, wait for each one, and try to piece together a coherent picture. That chain of calls is where things go wrong.
Researchers studying production AI failures found that an agent operating at 85% accuracy per step only successfully completes a ten-step task about 20% of the time. Each additional step multiplies the chances of failure. So even a fairly reliable agent becomes wildly unreliable when it has to juggle multiple data sources in sequence.
This is the problem Airbyte is targeting with its new Agents product. Rather than letting agents go hunting through your live systems at the moment they need data, Airbyte pre-pulls and organizes data from your key business tools, things like Salesforce, Zendesk, Slack, and HubSpot, into a single searchable index. The agent queries that index first to understand the situation, then only goes back to the live system when it needs to take an action.
The practical difference is significant. Airbyte's own early benchmarks show agents using this approach need up to 80% fewer processing steps compared to querying each system directly. On the widely used Zendesk connector alone, the token reduction reaches 90%. Fewer steps mean lower costs, faster responses, and critically, fewer places for the AI to get confused and generate a wrong answer.
What makes this launch notable is the write capability. Many tools in this space only let agents read data. Airbyte's connectors let agents actually update records, create support tickets, and post messages back into source systems. That is the difference between an AI that can describe a problem and one that can actually resolve it. It is also where real governance risk begins, since an agent writing to your CRM or ERP can cause damage at speed. Airbyte uses permission controls tied to individual users, so the agent can only touch what the person asking it to act is already allowed to touch. That approach is sensible but still early, and enterprises in regulated industries like insurance, finance, and healthcare will need to think carefully before letting any automated tool write to systems of record.
The broader story here is about where enterprise AI investment needs to go next. Companies have spent the past two years focused on choosing the right AI model. That race is largely over since the leading models are all capable enough. The next two years are about data plumbing: getting business data clean, connected, and available to AI tools in the right form at the right time. The companies that solve that, whether through tools like Airbyte or through internal investment in data infrastructure, are the ones whose AI pilots will actually survive contact with reality.
Gartner predicts more than 40% of agentic AI projects will be cancelled before 2028, not because the AI failed, but because the organizational and data infrastructure around them was not ready. The agents that do survive will be the ones that know where to find information before they start asking questions.