As Airtable's CCO, I spend a lot of time with customers deploying agents. The ones who are succeeding all took the same, unglamorous step first. They built the system before they deployed the agent. I know, that's not the advice anyone wants to hear right now. New models ship every few days, each measurably smarter than the last. And the pull to just point Opus 5 at your business and say “go” is almost irresistible. But a recent VentureBeat survey found that 57% of enterprises have traced a wrong agent answer back to missing or inconsistent business context. Just because the intelligence is there, doesn’t mean the context follows. Most agents default to reading whatever they can find — email, docs, Slack, transcripts — and end up reconstructing your business from a dozen unreconciled descriptions. When they work inside a structured, relational system your team already operates in, the difference is more than just accuracy — it’s immediacy. The answer is tailored to your business, and the work is already in motion before anyone sits down in the morning. That's the shift we're building toward at Airtable, and like always, our strategy starts with what our customers are already doing. Here's the full picture, including how four teams are putting agents to work today: https://lnkd.in/gEys293Y
This is the boring part nobody wants to buy, but its where the leverage actually comes from. An agent sitting on top of fragmented docs, stale CRM fields and contradictory Slack threads isnt intelligent, its just very fast at inheriting your mess. Build the operating system first, then give the agent permission to act inside it.
This tracks with what I've seen building agents at the enterprise level. The gap is never model intelligence, it's whether the agent has a reliable source of truth to pull from. I've watched the same failure mode happen with Copilot rollouts: give it messy, unreconciled inputs and it produces confident nonsense. Structure the data first and the "agent" part becomes almost easy.
Emily's question is the one that decides the whole project. In my own agent stack, three things earned their place: an immutable raw layer nothing can edit, a single index the agents read first, and a log of past decisions they check answers against. Source of truth turned out to be the easy part — the decision log is what stopped the confident wrong answers.
This is my line of thinking, but one issue I have with it is that if I make this context exist and robust in my relational data, what is new that agentic AI can bring to my database? I can compile it through automations, formulas etc. agentic AI can of course bring its own evaluated responses and interpretation if prompted but it’s then operating outside predictable processes which can be unsafe for compliance. I’d be interested to know more case studies, as I do love the concept… ripping data off documents is the best use I have had out of it so far.
What are the must haves in the "system"? I think this is what trips people up, knowing what the foundation needs to include. I've been encouraging a source of truth, but even that doesn't feel specific enough.
This is the part of enterprise AI that often gets underestimated. The model may be capable, but if the underlying business data, workflows, ownership, and relationships are fragmented, the agent is still operating on an incomplete version of reality. The strongest AI implementations I’m seeing start with the same foundation: structured data, clearly defined processes, consistent taxonomy, and governance. Once that exists, agents can move beyond simply answering questions and participate in the workflow. “Build the system before you deploy the agent” is a very important principle for where enterprise AI is heading.
Airtable doesn’t get enough credit as a data-cleaning tool. It makes it easy to organize, standardize, and scrub thousands of records at once. What’s a single row with dozens of columns in Excel becomes a record with clean relationships across multiple tables in Airtable
"What does this agent NEED to know to serve it's purpose well?" Probably a question we should ask about new hires as much as we do our agents but it's proven to be a critical step for enterprise agents in production, in my experience.
Building the structured data foundation before deploying ai agents is essential for reliability. Relying on disorganized sources often leads to those common hallucinations.
This is something we spend so much time communicating with clients on - agents need context and structure to be successful. If the process doesn't exist and data isn't cleanly available, then the agents are just doing expensive guessing!