Enterprise Agents at Scale
Deploying one agent is a technical problem. Deploying agents across an enterprise is an organisational problem.
What Changes at Scale
A single agent deployment is a contained system. Enterprise-scale deployment means:
- Multiple agents serving different teams with different needs
- Agents sharing infrastructure but needing different permissions
- Governance across teams that don't report to the same structure
- Compliance requirements that vary by geography and data type
- Talent that ranges from AI-native to AI-sceptical
None of these problems are solved by better prompt engineering.
The Three Layers That Break
Data layer: Enterprise data is messy, siloed, and often poorly documented. Agents need clean, governed data to work reliably. Most enterprises discover mid-deployment that their data isn't fit for purpose.
Identity and access layer: Who is allowed to ask the agent what? An HR agent should answer differently for managers vs individual contributors. Most access control systems weren't designed for dynamic, context-aware permissions.
Trust layer: Employees won't use tools they don't trust. Trust is built through consistency, explainability, and clear recourse when something goes wrong. Most enterprise AI deployments underinvest in all three.
What Actually Works
Start with a contained use case. One team, one workflow, measurable outcome. Build the discipline (evals, monitoring, feedback loops) before you scale.
Invest in the data foundation. An agent is only as good as the data it can access. Data quality work is unglamorous and essential.
Build the governance model before you need it. Once multiple agents are in production serving different teams, retroactive governance is nearly impossible. Define the rules when you deploy the first agent.
Measure outcomes, not outputs. An agent that generates 100 responses per day is not a success metric. An agent that reduces customer resolution time by 40% is.
The Talent Question
Nandan Nilekani's point about talent reskilling is the most underrated risk in enterprise AI. Junior developers learn by doing work that AI now does more efficiently. If you don't redesign the learning path, you break the knowledge transfer chain.
The fix: Mandate human review not just for quality control, but as a mechanism to force knowledge retention. Make it explicit that certain decisions require human reasoning — not because AI can't do it, but because the human needs to stay capable of doing it.
What's the most surprising challenge you've hit deploying agents at enterprise scale? Add it here.