By Andrew Tan
Why 85% of Enterprises Aren't Ready for AI
Fivetran's 2026 benchmark found 85% of enterprises running agentic AI on infrastructure that isn't ready. The model isn't the problem. The data pipelines feeding it are.
Here's a pattern showing up in an increasing number of places:
A company spends six months evaluating LLMs: They run benchmarks, negotiate contracts, build a proof of concept. The model looks great. Then they deploy it to production and the agents start hallucinating in ways that have nothing to do with the model at all.
The AI is fine, but the data feeding it isn't.
This is what Fivetran's Enterprise Data Infrastructure Benchmark Report 2026 is actually measuring when it says 85% of enterprises aren't ready for agentic AI. Not that their models are wrong, but that their pipelines are wrong.
What "not ready" actually means
The phrase "AI-ready infrastructure" gets used to sell a lot of things. It usually means something vague about cloud scalability. That's not what the Fivetran report is measuring.
The three concrete problems they identified:
Stale data. The agent is reasoning over yesterday's state. For a customer service agent, that means it doesn't know a refund was already issued. For a fraud detection agent, it's working from patterns that are 18 hours old. Batch pipelines running hourly or nightly can't support agents that need to act on what's happening now.
No schema enforcement. Source systems change constantly. Columns get renamed, types get widened, new fields appear. If your pipeline doesn't enforce schema contracts downstream, an agent can receive malformed data that looks valid — and confidently act on it.
No observability. Most pipeline monitoring tells you whether a job ran. It doesn't tell you whether the data is correct. An agent can be consuming data from a pipeline that's technically running but silently dropping 20% of events. You won't know until someone notices the agent is acting strangely, and by then the damage is done.
These aren't new problems. They're the same data quality issues that have plagued analytics teams for years. What changed is the blast radius. A data quality problem in a BI dashboard is a bad chart. A data quality problem in an agentic system is an autonomous decision made on bad information.
The stack mismatch
The AI-ready data stack problem is a deep mismatch in architecture assumptions.
Most enterprise data stacks were built around batch processing. Nightly ETL jobs. Daily warehouse refreshes. Dashboards that update every morning. The entire system was optimized for throughput over freshness.
AI agents have different requirements. They need data that's current, not just accurate. They need to act on what's happening in the last few seconds or minutes, not the last twelve hours. And when they're wrong, they need the system to catch it — not just log it and move on.
The companies that are ready — the 15% — didn't necessarily replace their entire stack. Most of them shifted to real-time where it matters and kept batch where it makes sense. The customer order stream runs in real-time. The quarterly cost accounting still runs nightly. The difference is intentionality: they made explicit decisions about which data needs freshness and built pipelines accordingly.
The three properties that actually matter
There's a lot of noise about what "AI-ready" infrastructure requires. Usually it's a vendor's checklist that conveniently maps to their product. So here's the version that maps to the actual failure modes:
Freshness. Not just "real-time for everything" — that's expensive and often unnecessary. But for the data your agents actually act on, you need to know the lag and have guarantees about it. If your customer data is 4 minutes old, fine — as long as your agent knows that and acts on it. What breaks agents is when they assume data is current and it isn't.
Consistency. Agents often combine data from multiple sources to make a decision. If those sources are running on different schedules or at different freshness levels, you get subtle inconsistencies that are hard to debug. A customer lookup says the account is active; the transaction stream hasn't caught up and shows it as pending closure. The agent makes a decision that's correct for each source individually but wrong for the combined state.
Observability at the event level. Pipeline monitoring tells you about jobs. Agent reliability requires monitoring individual events. Are events being processed or dropped? Is the schema matching what the agent expects? Are there bursts that overwhelm downstream consumers? This is a different class of monitoring than most teams have built.
What this means for the build vs. buy decision
The AI wave is surfacing a bill that's been accumulating in a lot of engineering orgs: the cost of treating data infrastructure as a solved problem.
Teams that built custom batch pipelines two years ago and called it done are now facing a hard choice: retrofit real-time onto a system that wasn't designed for it, or rebuild. Neither option is cheap. Retrofitting tends to produce the two-pipeline problem — a batch system for history, a streaming system for real-time, two codebases doing roughly the same thing with slightly different logic and perpetually diverging results.
The teams handling this well are the ones who don't have to make that choice. When batch and streaming run as the same pipelines with the same tooling, switching a workflow from hourly to real-time is a configuration change, not a rewrite. The observability, the schema enforcement, the failure handling — it comes with the platform, not as a custom build on top.
That's what layline.io is built around. Not just real-time for its own sake, but the ability to make explicit, intentional decisions about data freshness across the stack — without maintaining two separate systems to get there.
The question to sit with
The 85% number is striking. But the more interesting question is: how many of those enterprises know they're in that 85%?
The teams that are in trouble aren't usually the ones with obviously broken pipelines. They're the ones with pipelines that appear to work — jobs running, dashboards loading, no active incidents — but with silent reliability issues that only show up when an AI agent starts making consequential decisions on the data.
If your agents are acting strangely and you've already ruled out the model, look at the data.
Andrew Tan is a serial entrepreneur and founder of layline.io, building enterprise data processing infrastructure that handles both batch and real-time workloads at scale.
