AI hasn't delivered the productivity gains most companies have expected.
Tom Davenport recently laid out ten reasons why, and Bob Sutton, who wrote The Friction Project, endorsed the argument in a LinkedIn post that's been making the rounds. Their core point: companies keep chasing AI at the individual level, when the real value lies in redesigning end-to-end processes. I agree. And I'd push it one step further.
The individual was the crawl step
For the last couple of years, most companies have run some version of the same mandate: use AI in your daily work. And it delivered. Faster emails, quicker drafts, a better first pass. That was real, but it was just the crawl step, the thing that got people like me comfortable with what this technology can do. If you'd told me four years ago what we'd be doing with MCP today, I'd have said no way. Impossible.
But a pile of faster drafts isn't a transformation. Optimize person by person and you get side projects, not the unlock at scale companies actually hoped for. In life and in organizations, you have to shape the forest before the individual trees can thrive. The payoff from AI lives at that level, the end-to-end process, not the individual to-do list.
The forest has been hard to see for a reason
Here's why the forest-level view stayed out of reach: you couldn't get an accurate picture of how work actually happens. What people say they do in a process is rarely what they actually do, and documentation goes stale the moment someone finds a better workaround. So leaders interview a handful of people, sift through the available docs, and design against a stated process that doesn't match reality. Then it breaks.
The stated process gets circumvented in other ways too. Budgets get pulled. Companies develop a kind of institutional attention deficit and move on to the next thing. And the gap between what a company hoped to get from its tech and what it actually got quietly widens.
What just changed
What's different now is our ability to capture how a workforce actually works, at low lift, across every application it touches. Not what one person does in a single app, but how whole teams move across their stack to get something done for the business. Gather that at pace and scale, read the patterns in it, and you can see the shape of the work for the first time.
And it isn't only behavior. You can account for the technology you've already bought. We commit to tools, then reality drifts from the plan and we lose track of what we're actually getting from them. Map how work truly happens against what your tech can really do, and you find value sitting unclaimed inside what you already own.
Process slop is what happens when you skip this
This is where "work slop," and its more specific cousin "process slop," come from. A company builds an agent off its corporate documentation and a handful of interviews. All of it is anecdotal. Meanwhile the work keeps changing, people find better ways or quietly route around what's broken, and none of it gets written down. The agent still does something. It just isn't the thing you needed, because it was built on how someone described the work six months ago, not how the work happens today.
Think of it like planning a party. Tell an AI to plan a party and it will. But if you never mention that it's outdoors, that it might rain, and that it's for six-year-olds, you'll get something lovely for forty-five-year-olds indoors. It's a party. It just doesn't resonate. Agents lose resonance the same way when they run on generic inputs instead of the nuanced reality of how work actually happens.
The nuance is the point
I've watched this land with customers. A team is copying and pasting the same information six, seven, eight times a day. Nobody flagged it. It's just the job. Those repeated steps are the employee experience no one measures. But the second they see it laid out at scale, they start working out a better way on the spot. That's the nuance an interview never surfaces, not because anyone is hiding it, but because it's too ordinary to mention. One paper cut is nothing. At scale, it's a real tax on the business.
Then you remove the thousand paper cuts
Ground truth gives you two things: how long work actually takes, and the real path people take to get it done. From there, you design the target state, the people, process, and technology it needs, and you close the loop. What's left is the death by a thousand paper cuts, the small inefficiencies no one bothers fixing because none of them alone seems worth the effort. Remove those, ideally through automation.
The return shows up on both sides of the ledger. More from the people already on your team, and more from the technology you already pay for, because you're finally designing around what each can actually do rather than what you assumed when you bought it.
Agents aren't a "play." They're something you tend.
One caution. I keep hearing agents described as a "play." Call something a play and it turns superficial, something you run once and move past. Agents aren't that. Used well, at the end-to-end level, they can 10x what a company does. But only at that level, which is exactly Davenport's point.
And end to end doesn't stop at launch. A tree needs the right conditions and steady care, not a one-time planting. Agents are the same. You discover where one can deliver a better outcome, create it, watch what it actually delivers, and adapt it as the ground shifts. Feeding real behavior into the model through MCP is what lets you make that call with confidence, and keep checking whether the agent is still doing what you thought. Because reality will keep moving, and that constant feed of what's actually happening is what keeps an agent alive.
The bottom line
AI pays off when it runs on real behavioral data, not assumptions. The barrier was never ambition. It was visibility, an honest picture of how the work actually happens. That's what Fullstory Workforce is built to give you. Shape the forest, and the trees will thrive.
