Before AI became a buzzword, the company I cofounded, Ivalua, was already laying the groundwork.
Years ago, we began investing in AI, cultivating internal innovation and exploring its potential. Our behind-the-scenes work soon evolved into a company-wide movement, which accelerated when our founder and CEO became our chief AI officer.
We trained our managers to build agents, encouraged everyone to experiment and innovate, and soon had 800 agents running across the company’s every function. That experiment forged the path we’re on now—one intelligent agent, governed by the same rules that govern the people using it.
The journey changed how we think about AI transformation, and what it means for how work gets done.
Our 3 phases of AI integration
The first phase was deliberately limited. A core team of fewer than 10 people from our R&D and product development teams explored tools, built the first prototypes, and figured out where AI could help and where it fell short.
The second phase was about education. We went department by department with a simple message: “AI is going to transform your business long before AI transforms you.” Take ownership of that. Don’t wait for someone to hand you a use case. Instead, find your own.
Shifting the framing turned passive observers into active participants. Our chief AI officer led sessions to set the tone, and a cross-department network of AI champions drove conversations and surfaced ideas from within their teams.
In the third phase, we trained managers to build agents—explaining how the platform worked, showing examples, highlighting early wins from the departments that had moved first. Then we encouraged people at all levels to try new ideas, build things, and propose changes.
The 800 agents that followed came from people who had the freedom to act within clearly defined boundaries. Those boundaries had a specific meaning in practice. For example, employees could build “collections”—curated sets of approved documents that gave an agent additional context for more complex requests. If an employee wanted their agent to draw on documents beyond their default access, executives would review and approve that collection before agent use.
Every department had an agent—sales, finance, legal, marketing, and customer success. In marketing alone, teams built agents for customer stories, product marketing content, and communications operations.
Elsewhere, one of the earliest agents answered company policy questions. Anyone could type a question and get an answer without knowing which document to open or which team to ask. It was a simple but successful use case.
What 800 agents taught us
Scale has a way of surfacing questions that small experiments don’t. When you have 800 agents running across a company, you start asking things you didn’t ask before. Who owns this agent? What data can it see? If it acts on something incorrectly, who answers for that?
Governance that works at 10 agents doesn’t automatically work at 800. We were in territory that required a new kind of thinking. The experiment proved a universal employee and manager appetite. But managing hundreds of agents adds nontrivial work to a company’s operations. Each agent carries its own access logic, data boundaries, and accountability questions. Multiply that across departments, and you’re managing a rapidly expanding management of AI.
The solution was simpler than we expected. We moved to a single agent, which can carry out activities across the organization, without worry about a given agent being designed in a silo. It’s the same agent for everyone, built into the platform, and accompanying each person like a highly capable colleague, but bound by the same permissions, knowledge, and access that person already has.
So what became of the 800? At our company of more than 1,000 employees, one particular team alone had built about 100 agents. Now, only one agent solves the team’s daily challenges. The team’s leader job ensures sure the agent has access to the right policies, procedures, and data.
(For complex work, Ivalua’s single agent creates temporary sub-agents on the fly, but users don’t need to worry about them.)
This structure solves the accountability question—who answers when something goes wrong. The single agent operates with an employee’s permissions and no one else’s. Each person on the team is accountable for what the agent does when they use it. There’s no ownership ambiguity, nor a new governance layer. The rules that already govern people govern the agent, too.
That’s the operational difference we sought. AI that travels with the person—and stops where they stop.
A transformational interface
Most enterprise software still asks people to navigate where to go, which module to open, which report to pull. In my experience, most people never fully learn the software they’re given, and that gap limits the system’s offerings.
When the interface becomes a chat where anyone can type what they need in plain language instead of learning where to click, that gap shrinks. A finance leader can ask a question and get a real answer from the procurement system. A new procurement manager can do meaningful work before learning the platform’s nuances.
The system also holds the organization’s expertise. Procurement policies, category strategies, the way your best people approach a negotiation or a sourcing event: Those details get captured as instructions the agent follows when necessary.
Our bottom-up experiment at Ivalua showed us that the most sophisticated AI strategy—even one built through real trial and error—won’t get you anywhere unless you put your people first.
Dan Amzallag is the cofounder and chief operating officer at Ivalua.