Agentic AI Has the Same Problem Penicillin Did: Deployment Capacity 

Justin Gillebo

August 19, 2026 • 5 min read

---

Afew weeks ago, an operations leader at a life sciences company walked me through a planning environment where hundreds of people maintained critical parameters across the network. Each was making reasonable local decisions based on what they could see, but taken together, those decisions produced poor system-wide outcomes. The issue was not bad intent or even bad execution. It was fragmented ownership of the inputs driving the system. 

That conversation reminded me of – stay with me – penicillin. Alexander Fleming identified penicillin’s antibacterial properties in 1928, but the breakthrough didn’t spread for more than a decade. While initially overlooked by the medical establishment, the primary obstacle was production. Penicillin could not yet be made reliably, cheaply, and at scale. It took a wartime effort of government, industry, and researchers treating manufacturing as urgently as discovery to turn a laboratory insight into something that saved lives at industrial scale. 

Thinking about this, I kept coming back to one idea: deployment capacity. It explains why some breakthroughs change the world slowly, but it also perfectly describes what is holding back agentic AI in the enterprise.  

The long, messy work of building the infrastructure is what allows an invention to operate in the real world. But right now, enterprise AI is much closer to a promising lab discovery than to a reliable industrial process. Changing this is about making the underlying decision system trustworthy enough to let models act. 

The Model Is Not the Bottleneck 

The agentic AI breakthrough has already happened. LLMs can now reason, retrieve, summarize, and in bounded cases even act. The frontier problem is no longer whether the models are impressive, but whether companies have built the systems, permissions, governance, and operating discipline required to deploy them inside real workflows. The constraint on ROI is shifting from technical novelty to deployment capacity, but our data and analysis show that 96% of companies are not responding fast enough, taking days to weeks to respond to changes in demand, supply, or cost.  

Across companies like AWS, OpenAI, Databricks, and Snowflake, the center of gravity has moved beyond generation alone toward grounding: connectors, semantic layers, retrieval, permissions, lineage, orchestration, and structured context. That reflects a simple reality: an enterprise agent does not just need a good model of language, it needs a usable map of the business. 

A customer service chatbot can survive a fuzzy answer, but agentic applications within supply chain mean that leaders must be ready to manage an inventory commitment or planning recommendation based on the right inputs, made inside the right workflow, and governed such that someone is willing to let the system act. 

That is why so much of the public discussion still misses the point. It overweights the interface and underweights the operating model. The hard enterprise problem is not whether AI can generate a response, but if a system can convert data into an approved, trustworthy course of action. 

The Operating Model Shift 

To create real ROI from agentic AI, the life sciences company we mentioned earlier first had to change its operating model. It moved from distributed parameter maintenance to centralized ownership and shifted from static master data to continuously reviewed decision inputs. It stopped treating governance as a control layer bolted on after the fact and started designing human-in-the-loop decision rights into the workflow from the outset. Governance was the unlock that made trustworthy automation possible. 

That is also why “data quality” is often too blunt a diagnosis to be useful. Most companies do not need to clean everything before they deploy AI. They need to make a limited set of inputs behind high-value decisions reliable, timely, governed, and fit for use. In supply chain, that usually means the operational data that drives action: planning parameters, lead times, sourcing rules, service thresholds, inventory policies, and the master-data fields that shape decisions upstream. 

For years, teams have compensated for weak inputs with human judgment, workarounds, and sheer effort. People know where the system is wrong, and they quietly correct it, but agents cannot do that safely. They do not have intuition, scar tissue, or permission to improvise their way around a bad parameter. If the underlying decision system is brittle, the agent will scale the brittleness. 

Selective Autonomy Is the Future  

The most credible near-term future is not ambient AI colleagues floating across the enterprise. It is selective autonomy: bounded workflows, explicit thresholds, trusted inputs, clear exception paths, and decision rights that are visible before the agent ever acts. The best early use cases will be the workflows where better speed, consistency, and judgment produce measurable gains in service, inventory, or cash. 

Agentic AI readiness starts with the data that drives action, not the data that fills dashboards. Before a company can trust an agent to act, it must make the underlying decision system worth trusting. 

In reality, this pushes operations leaders to identify the handful of decision inputs that actually matter. It forces clearer ownership, makes human-in-the-loop design a feature of the operating model rather than a temporary safeguard, and gives companies a practical way to create ROI without waiting for perfect data or perfect models.  

Start by picking one high-value workflow where speed and consistency matter: inventory positioning, planning exceptions, or supplier response. Then work backward from the decision: what inputs drive it, who owns them, how they are governed, when a human must approve, and what evidence is required to let the system act. This unglamorous work is the practical path to turning agentic AI from experimentation into measurable ROI.