Article

Nobody at the prompt

Nobody at the prompt

In the first of these pieces, Smartsheet's chief AI officer Drew Garner described the people who walked through the door when the company opened its platform to AI agents: project managers, business owners, a CFO who has never typed a number into a cell. Here he describes what happens next, when the people stop typing altogether and the agents keep working.

 


 

Somewhere on our platform this month, a request is going to end with "every Monday at eight, send it to these two people." From then on nobody will type anything. An agent, or a team of them, will wake on the schedule, read the sheets, compute the rollups, write the report and put it on two desks before the coffee is made. If a row changes that matters, another one will wake because of that. We are switching this on for ourselves first, and the question it answers is the one I get asked most: what does AI look like when the novelty of the chat window wears off?

It looks like automation with judgment. Our platform already runs a quarter of a billion standing automation rules, deterministic if-this-then-that that fires more than 100 million times a month. That machinery has been reliable for years and it does not think. Agents think and do not, on their own, run on rails. Put the agents on top of the rules, and you get both: a trigger that is exact, a response that can weigh what it finds.

Today, about a fifth of the agent requests to our platform arrive between midnight and eight in the morning. My bet is that this inverts. Four in five of the actions agents take will eventually happen while the person they work for is somewhere else, on a schedule, on a trigger, or because another agent has finished something. When that is true, three things a platform believes about itself stop being philosophy and start being the product.

The first is that the connection is the product, not a wrapper around it. Much of the market has shipped a thin layer over an existing API and kept the good parts for its own assistant: the clever features live inside the product, and what you get over the wire is a list of nouns and verbs. We went the other way. The tools, the agents behind them and the expertise served with every call all live on the open connection, and our own assistant is one more client of it. By design, the AI your teams already use is as capable on our platform as ours is. There is a simple way to test any vendor on this, and I would run it in the next meeting. Ask whether their own assistant can do something yours can't. If the answer is yes, you have bought a wrapper.

The second is that the unit of work is a team, not an agent. When someone describes the workspace they need, one agent does not build it; a team does, in parallel, the sheets, the dashboards, the reports and the rules that connect them, in one sitting. Sixteen thousand people in nearly 8,000 organisations had those teams build more than 40,000 artifacts in the first seven weeks. That capability is the next to move onto the open connection, so the assistant your people already use can dispatch the same team. Nobody has to become a power user of anything. The power user is the team, and it is hired by a sentence.

The third is that expertise has to be served, not carried. A finance chief's judgment about what the numbers mean lives in his head, and the next ten people on his team would carry 10 versions of it in their prompts. Most of the industry is solving this at the personal level, with memory features and custom instructions, which work for one person and fail for 11. Codify how the work is done once, serve it to every agent at the moment it acts, and the newest analyst gets the finance chief's answer. We tested this with one of the world's largest consulting firms: the same model on the same project data improvised a plausible dashboard when asked cold and rated the project as needing attention; given the firm's own methodology at the moment of work, it read every required source and flagged the project red, with hundreds of overdue tasks the improvised version had buried. Same model. Same data. The difference was the context delivered with the call, and when a correction lands in that layer, everyone's agent gets it the same day.

Now put those three together with the agent running at three in the morning. The connection carries the tools, the team and the methodology, and nobody is there to say "wait" . So the check cannot live in a person's diligence, which is where it lives today. It has to live in the platform: the agent proposes a change and shows exactly what will alter; someone other than the requester approves it, on their own schedule, not the agent's; the platform commits it and logs the whole chain. We are building exactly that, and until it is shipped the honest instruction to any customer is to switch on unattended writes only where the blast radius is one you would accept from a new hire on their first night shift.

The brief, then, for the people who make this safe. Treat the connection as the product and run the wrapper test on every vendor, including us. Decide which work may happen unattended and route it through approval rather than switching it off. Codify your methodology once, where the agents pick it up, rather than in 100 inboxes. Measure the work done, not the prompts typed, because the prompts are about to stop.

The Monday report will be on two desks at eight. Nobody will have asked for it that morning. That is the whole point.