Article

The builders were already in the building

The builders were already in the building

When Smartsheet opened its platform to AI agents in March, the people who came through the door weren't developers. They were project managers, business owners and, in one case, a chief financial officer who has never typed a number into a cell. The heaviest users are already leaving the interface, and the rest will follow. That changes what IT is for, argues Drew Garner, the company's chief AI officer. 

 


 

The finance chief of a mid-sized manufacturer has not typed a number into a spreadsheet cell in nearly a year: one exception, in hundreds of thousands of actions. He was never that kind of user. What changed this summer is that an AI agent, working under his login, started doing it for him: 4,000 actions in the first month, 20,000 in the second, 55,000 in the third. In one six-week window it pushed 2.4 million cell-level changes into more than 700 project sheets, with zero deletions. A few thousand requests did it; each one now does the work of several hundred clicks.

A pattern-matching read of that activity guessed the person behind it was a procurement coordinator. It was the CFO, running the line items himself.

We opened our platform to AI agents in March, over an open connection any AI assistant can use. I expected developers to arrive first. They usually do. Launch week was under 1,000 people in a few hundred organizations; last week it was more than 20,000 in thousands of organizations, close to 2 million requests, most of them reads. Project manager is the most common job title among the heaviest builders, at nearly three times the next; then owners and chief executives of smaller firms. Two-thirds of the traffic comes through an ordinary chat assistant, the same window people use to draft emails and settle arguments. Nobody took a course. They typed.

The seats had been there for years. Of our heaviest agent users in the first three months, nearly nine in ten had never used the product or had held a license, mostly for over a year, without changing a thing in it; the median such seat had gone 681 days without a change before its first agent request. Some, like our CFO, were active for years without touching the data by hand. Either way, already in the building, paying for depth they couldn't reach.

Every mature enterprise platform has the same shape: 20 years of depth behind an inner door only specialists could open, while everyone else files a request and waits. The depth was never the problem. The door was, and it hasn't moved; what changed is who gets through it. The price of admission used to be expertise. Now an agent pays it for you.

On September 8 we let agents build automation rules, the specialist's corner of the product; in the first week 844 people in 709 organizations built 4,256 of them. Six weeks earlier we had opened creation itself, and 16,000 people in nearly 8,000 organizations had teams of agents build more than 40,000 sheets, dashboards and reports in the first seven weeks.

Every time a capability opens, the people who were waiting for exactly that walk in. Our CFO's favorite one computes live rollups over operational data, 40,000 requests in two months: not building the pivot table, asking for the answer.

The honest read, though: he is still the exception. In one recent 90-day window our users made roughly 580 million row edits by hand; agents and our own AI features made about a million. I read that ratio as runway, not verdict, because the people furthest along have already left the interface. When the 30 heaviest agent users at one pharmaceutical company moved their work to AI this spring, their weekly web activity fell by 44%. Every seat-utilization dashboard would flag them as churn risks; they were the company's most engaged users, and anyone measuring next year in logins will conclude AI failed at the precise moment it worked. For them, as for the finance chief, the screen is now where you look at what the agents built and approve it. That is where everyone is heading; the ratio only says how far there is to go. The last step, agents that work while nobody is at the prompt, is the next piece.

The fashionable version says the next wave belongs to the business, not to IT. Half right. The builders have changed; the people who make it safe to build have not, and they now have twenty thousand builders and no code to review.

An agent asked for an outcome in plain language has to know what "the status column" is and what changed last Tuesday. On a structured platform that context is data: column types, cell history, permissions. Without good data, AI can be missing things, and one guessed column is a dashboard wrong by lunchtime. Sovereignty is the same argument: region matters, but the version of this story that goes badly is the sheet pasted into a chatbot connected to nothing, with no permissions, no log and no idea what the model retains. Make the governed route the easy route.

 

Back to the CFO, because this is the part I would put in front of a security team. The month his agent use took oD, his own checking, opening the sheets and reading them, spiked to between two and five times his year-long baseline. The next month, as his reliance nearly tripled again, the checking fell below where it had started. He checked hard once, then stopped. Whether the agent earned that or he simply got busy, the telemetry cannot tell you. What it can tell you is that the controls that don't depend on attention held: every change is in the log, nothing was deleted, the permissions never widened. The one control that does depend on attention, a human reading the work, lapsed in the most numerate person in the building. That is not an argument against him. It is the argument for putting the check somewhere that doesn't get tired.

An agent that can write can write the wrong thing, at scale, at three in the morning. Today the controls are writes on or oD, plus a confirmation on the riskiest actions, and on-or-oD is not a governance model. Business users need what engineers have had for two decades: the agent proposes a change and shows exactly what will alter; someone other than the requester approves it; the platform commits it and logs the chain. We are building exactly that. It is not shipped yet, so put it to any vendor as a requirement, including us.

Customers are already building the controls themselves. At that pharmaceutical company, the AI-governance team asked an agent to build the registry that tracks and approves the company's other agents: draft, in review, approved, with a change log. Nobody wrote a line of code.

The brief for the next 12 months, then. Decide which assistants may connect, switch on writes deliberately, surface by surface, and hold that policy in the platform, not a training deck. Put the platform behind the AI your people already use rather than one integration per model. Measure what got built. Stop treating governance as the last step, and stop assuming the first people through the door will be technical. Ours weren't.

What separates the organisations that get value from AI from those that pilot it forever is not which model they rent; everyone rents the same ones. It is how much of their own depth their people can now reach, and how safely. We opened the door in March. Six months on, a CFO runs his company's line items himself, every day, and has still never typed one. He is not a curiosity. He is early.