The infrastructure bottleneck
Data cleanup starts to unlock agentic
Most enterprises trying to “go agentic” are discovering a painful truth: the limiting factor isn’t models, it’s plumbing.
On paper, autonomous agents sound straightforward: wire an LLM into your tools, give it a goal, and let it execute. In reality, most organizations still lack the unified data environments and process documentation required for agents to reason effectively. The data estate is a patchwork of SaaS silos, legacy systems, and un-versioned spreadsheets. Processes live in people’s heads, half-written SOPs, or Slack threads.
We expect to see a major jump in enterprises prioritizing data organization as a 2026 goal. — Kit Colbert
Drop agents into that environment and they don’t become smart coworkers; they become very expensive interns, spending most of their “thinking time” just trying to interpret inconsistent inputs. Instead of compounding performance, they amplify noise.
The first missing piece is a usable data layer: an environment where customer, transaction, operational, and content data are cleaned, structured, and accessible in consistent schemas with clear provenance. That doesn’t mean boiling the ocean into a single warehouse, but it does mean deciding what “source of truth” actually means for core entities, and instrumenting the paths that matter. Until you do that, every autonomous workflow degenerates into special cases and brittle glue code.
The second barrier is orchestration. A single agent doing a narrow task in isolation is easy; coordinating multiple agents across departments is not. Real enterprise work cuts across finance, operations, support, sales, and compliance. That demands systems that can monitor, evaluate, and correct agent behavior in something close to real time. You need different agents looking at different data sources and perspectives, agents talking to each other and comparing their findings, and agents that specialize in quality control—spotting inconsistencies, challenging weak recommendations, and selecting the best response before anything hits a customer or a system of record. The real capability in 2026 isn’t logging what a single agent did; it’s orchestrating a mesh of agents that can cross-check, veto, and improve each other’s work.
Without an orchestration layer, organizations fall back to manual inspection, which defeats the point of autonomy. You end up with a human-in-the-loop for everything, not just the edge cases.
Then there’s governance, which will slow progress more than most vendors admit. You’ll need infrastructure guardrails that control what agents are allowed to do, such as blocking certain RAG or MCP calls, or constraining which systems they can touch. Couple that with real-time evals that watch what agents say and return, catching policy violations, sensitive data, or unsafe behaviour before it reaches a customer or a system of record.
As agents gain autonomy—touching money, customer accounts, or sensitive records—organizations will need frameworks for traceability, approval, and recovery when things go wrong. It’s not enough to log prompts and responses. You need:
Ironically, this kind of governance is easiest when you’ve already done the unglamorous work of process mapping and data alignment. If you don’t know how the human process works today, you won’t be able to explain or control the agentic one tomorrow.
So the next phase of “AI transformation” won’t be won by whoever plugs the latest model into their stack first. It will be won by the companies willing to do the infrastructural grind: documenting processes to a level a machine can follow, rationalizing their data environment, and investing in orchestration and governance as first-class products, not afterthoughts.
In other words: until you fix your infrastructure bottleneck, “agentic AI” is just a nicer UI on top of the same old chaos.
