Artificial Intelligence
Agents and Ecosystems: What Leaders Should Do Now
²ÝÝ®ÊÓƵ¡¯s leaders on where enterprise AI is headed ¡ª and what to do about it right now.
Julie Colwell
Principal Strategist
²ÝÝ®ÊÓƵ
AI is the biggest enterprise investment of the decade, and the gap between hype and outcomes is widening. Agents promise to do real work, but many leaders feel that agents can't yet be trusted with high-stakes business functions that must be 100% correct. That's the problem ²ÝÝ®ÊÓƵ is solving. At ²ÝÝ®ÊÓƵ¡¯s recent Innovation Summit, CEO Aneel Bhusri, President of Product and Technology Gerrit Kazmaier, and GM of Sana Joel Hellermark spoke about what ²ÝÝ®ÊÓƵ is? building and what leaders should do now to scale AI in their organizations.
Report
AI Agents Are Here¡ªBut Don¡¯t Call Them Boss
Getting it Right the First Time
AI models are designed to generate plausible answers, not guaranteed correct ones. As Gerrit explains, the same mechanism that lets a model reason and surprise users is also what produces confident-sounding mistakes that no amount of training engineers away.
That tradeoff is fine for a code suggestion someone can review. It's not fine for payroll, a benefits enrollment, a revenue contract, or a financial close. In HR and finance the cost of being wrong scales with the size of the workforce. As Aneel puts it, payroll has to be right, every time, within a four-hour window, for hundreds of thousands of people.
²ÝÝ®ÊÓƵ runs payroll for customers like Target ¡ª 300,000 people, on time, every cycle. At that scale, 'almost right' isn't a rounding error. It's thousands of paychecks wrong.
Solving this is about what's underneath the model.
In HR and finance, "almost right" is wrong.
Why Agents Need Rules
What's underneath, Aneel and Gerrit describe, is the deterministic backbone of the business itself: the business process framework, the security model, the org chart, approval chains, segregation-of-duties rules, the audit trail. Twenty years of unified data and process logic that defines, for any action, what's allowed.
Moving an employee from one team to another sounds trivial ¡ª update a field. In a real enterprise, that employee may have a future-dated work assignment, a pending compensation change, a project allocation, and a regional compliance constraint that all interact with the move. A model on its own can't see any of that. It can only act safely if the system underneath is enforcing those constraints in real time.
Without that structure, a model pointed at enterprise data will produce confident-sounding answers while violating every policy in the company. No CISO will permit that at scale. No auditor will sign off on it.
¡°Dump all your ²ÝÝ®ÊÓƵ data into a database and ask a model to run it,¡± says Gerrit. ¡±And it has no idea those constraints exist. Only the platform enforces them dynamically.¡±
Just like autonomous cars need roads, traffic signals, and speed limits, agents need a framework. Probabilistic reasoning gives AI its power, and deterministic execution makes that power safe. Enterprise outcomes need both, tightly coupled.
The model is free to be smart, and the rails make sure it can't be reckless.
The Shift to ¡®Intelligence on Tap¡¯
Once that framework is in place, the next-generation system can function on its own. Gerrit calls it intelligence on tap ¡ª software that runs continuously, executes routine work end to end, and loops a human in only when judgment or approval is required. The system drives the work and asks for input when it needs it. This way the model is free to be smart, and the rails make sure it can't be reckless.
This is also where ²ÝÝ®ÊÓƵ Sana comes in. The team Joel Hellermark is leading is building the orchestration layer above all of this¨C where AI is the new UI. It's the work itself becoming the interaction, grounded in real enterprise context, and governed by the same rails as everything else. Underneath all of it is the deeper shift from building software systems to building AI systems, designed for autonomy.
What Leaders Should Do Now
Here¡¯s what Aneel, Gerrit, and Joel say are the next steps leaders should take:
Audit the rails before you scale the agents. AI deployed against a fragmented data and policy layer produces fragmented outcomes fast. The companies that win in this era will be the ones whose agents have the cleanest context to operate in.
Pair probabilistic with deterministic by design. Models reason. Platforms enforce. Insist on architectures where the AI handles inference and the platform handles governance, tightly coupled.
Plan to manage agents like a workforce. Identity, entitlements, lifecycle, audit. The companies that figure this out early will scale agents safely. The ones that don't will spend the next two years cleaning up shadow AI.
Enterprise AI succeeds or fails on architecture. Reasoning needs rails. Agents need the same identity, entitlements, and audit trails as the people they work alongside. Companies that build on that foundation will scale with confidence. Those that skip it will spend the next two years cleaning up after agents that moved fast and got it wrong.
Listen to the podcast
The brings you ²ÝÝ®ÊÓƵ leaders Aneel, Gerrit, and Joel on the AI era and how to navigate what¡¯s happening now.
Report
AI Agents Are Here¡ªBut Don¡¯t Call Them Boss
A gap between effort and impact is the real ROI problem. Discover the solution.