Most AI projects I have watched over the last three years failed on a question nobody thought to ask, because it looked too obvious to say out loud: why are we still doing this ourselves? Instead, almost every AI meeting opens with a different question. How do we do this faster? It sounds reasonable. It is the start of nearly every disappointed AI initiative I know.
I told myself the faster story too. At SlopeLift we spent 2023 and 2024 building workflows, testing them, tearing them down. Some caught fire, some missed completely. After enough demolished workflows, I realised the framing was the problem. The technology worked fine. Accelerate a bad process and you get more of the wrong output, faster. That sounds obvious the moment you say it. In practice you only see it once the pile of output is in front of you and someone asks what to do with it.
The shredder before the tool
The more useful order reverses the question. Before a process gets a tool, it goes on the bench with one question: does it belong in the shredder? Most organizations keep a to-do list of tasks that need doing and shop for tools to speed them up. The Anti-To-Do List looks for the opposite: tasks that should leave the desk entirely, because a person should not be touching them at all.
A faster task is still your task. A handed-off task leaves your desk, and only then does capacity open up for the work where differentiation actually happens.
Three categories
Three categories show up at almost every company we talk to. The first is the work nobody wants. Data cleanup, formatting, the repetitive reporting steps qualified people feel are beneath them and therefore do halfheartedly or not at all. This is the easiest place to start, because nobody defends their claim to fixing umlauts in a CSV. We had a weekly data cleanup before reporting that reliably ate two hours; an agent now does it in four minutes, and nobody mourned the two hours. The saving is visible the next day, and it converts skeptics faster than any slide about agentic workflows.
The second is the knowledge gaps. Quick research in an unfamiliar industry, competitive scans, briefing prep. This work disguises itself as waiting: someone stalls because they would first have to look at ten competitors, and looking at ten competitors is tedious, so it slips. An agent runs that sprint in minutes, and the human steps back in where the material has to be interpreted. One caution: cleanup is forgiving, research is not. A wrong number in a summary can sound convincing and walk straight into a client document. Hand off the gathering, keep the judgment.
The third is the blank-page block, the jump from zero to something: first drafts, concept scaffolding, the empty page at the start of every creative process. This is where resistance is highest, because a first draft feels like authorship. The value, though, sits in the judgment that follows the draft: what stays, what goes, what lands with the client and what misses. An agent that fills the blank page dissolves the block and leaves you the authorship.
What the research says, and where it warns
The Anti-To-Do logic has an empirical backbone that rarely gets cited here. In 2023 a team led by Fabrizio Dell’Acqua at Harvard Business School watched 758 consultants work with GPT-4. Inside the band of tasks the model handles well, the gains were large: more tasks completed, faster, at higher quality. Outside that band, on what the researchers named the jagged frontier, the model was confidently wrong, and the people who trusted it there did measurably worse.
The lesson maps onto the list precisely. Hand the commodity work over fully, the three categories above, so your attention is free for the work where you have to stay in control. The trap is handing everything over and leaning back, which is exactly where AI degrades the output while looking productive. You delegate the one thing so you can be more present on the other.
None of this requires a new tool on day one. It requires a list, written with the team in the room, because the Anti-To-Do List is a team exercise. The single user who adopts a tool just gets personally faster while everyone else works as before, and that is a redistribution rather than a gain.
So before the next tool demo, I would ask one thing of your team: name the three tasks on your desk this week that a competent agent could take off it entirely. If you cannot name three, you have not looked closely enough. If you can, you have your Anti-To-Do List, and you have not spent a cent yet.