AI Assigning Tasks to Employees: Why Numbered Steps Matter

When an AI assigns a task to an employee, the task must carry numbered steps: what to open, what to do, what finished looks like, and what access is needed. A task with only a title is a claim that work exists. It does not tell the person how to do it, so it waits until someone has spare attention, and spare attention is rare.

We learned this in August 2026, from our own backlog. This post covers what a stale task actually looked like, the rule we wrote after it, and a check you can run on your own task list this week.

Why do tasks an AI assigns to people sit untouched?

They sit because the person receiving them has none of the context the agent had. The agent read the source data, the history and the goal. The employee gets a title. Rebuilding that context costs more effort than the task feels worth, so it slides down the list.

The attention budget is already thin. Microsoft’s 2025 Work Trend Index, published 23 April 2025, reported that 80% of the global workforce says it lacks the time or energy to do its job, and that employees are interrupted by a meeting, email or ping about every two minutes. A task that demands twenty minutes of detective work before the real work starts loses that contest every time.

Nobody is being lazy here. The instruction gap, not the work, is what makes the item stale.

What did a bare task title cost us?

In August 2026 our task backbone, the single store where every task lives, held a pile of items that agents had raised for people and that had gone nowhere. The chief-of-staff agent had been filing them with short titles like “close the reconciliation work to audit grade.”

Read as a human, that title is unactionable. Which statement? Which period? Match against what? Where does the evidence go? The agent knew all of that. The task did not say any of it.

Compare the version we now require. “Pull the bank statement for the first to the last day of the month. Match each posted settlement against it line by line. Flag any that do not reconcile. Attach the statement to the task.” That is a short job for someone who has the statement in reach.

The work was the same in both versions. Only one of them got done.

What is the rule we adopted?

The rule, set by our founder in August 2026: whenever an AI sets a task for a person, the task carries clear step-by-step instructions. It applies to every agent that writes to the task backbone, including the watchdog that raises stuck-work tasks, and to any coding session that raises work on someone’s behalf.

A ruling is not an instruction. When a decision creates work, the decision text is not what the assignee needs. Someone, or something, has to write the steps separately for whoever acts on it.

Every task body now needs four parts:

1. Numbered steps, in the order a person would do them.
2. The specific system, page or file to touch.
3. What done looks like, stated so it can be checked.
4. Any access or credential the person will need, named but never pasted.

What separates a task that sits from a task that gets done?

A task that gets done tells the receiver everything they would otherwise have to go and find. The differences are mechanical, which is good news: you can write a template for them and make the agent fill it in rather than hoping it improvises well.

Element	Task that sits	Task that gets done
Title	A goal (“close out the audit work”)	An action (“Match July settlements to the bank statement”)
Steps	None, or one sentence	Numbered, in doing order
Location	”In the system”	The exact page, file or report
Done means	Unstated	A visible artifact: attachment, status, count
Access	Assumed	Named, with who to ask
If blocked	Silence	Who to message, and what to say

The last row matters most for agents. A person who gets stuck on a vague task will not file a question. They will quietly leave it. Giving them a named person to ask turns a stall into a message.

Why does this matter more as agents take on more work?

Because the volume of agent-created tasks grows faster than people’s capacity to decode them. Asana’s State of AI at Work report (2025) found that workers plan to hand off nearly half their work to AI agents within three years. Each handoff creates a thread that eventually reaches a human, and every one arrives with whatever context its author thought to include.

Visibility is a related weak point. Atlassian’s State of Teams research (2025, 12,000 knowledge workers and 200 executives) found that only 13% of Fortune 500 executives say teams have complete visibility into each other’s priorities and progress, a figure Atlassian published in June 2025. If humans cannot see each other’s priorities, a task from an agent with no explanation has even less chance of being slotted in.

Our post on AI task follow-up automation covers the chasing side: how an agent notices that a task has gone quiet and verifies the outcome. This post is the upstream half. A chase works far better when the task was clear in the first place, because the person has something to act on rather than something to ask about.

Should the AI also ask the CEO before chasing?

No, and this was the second rule from the same month. An item blocked on a missing fact is not a decision. Fetching a fact is an errand, and routing every errand to the top of the company turns that person into a gatekeeper.

The founder’s instruction, in August 2026, was to chase people directly and bring decisions, not questions. The agent checks the current state first, so nobody is nudged about something they already finished. Then it contacts the person who holds the fact, with the steps attached. The founder sees the recommendation and the facts already gathered, not the errand.

The two rules work together. Direct chasing without steps produces a nudge the recipient cannot act on. Steps without direct chasing produce a clear task that nobody is reminded of.

What is still open?

The rule is written and the agents that file tasks follow it. What we have not done is audit the full history of older tasks, and the tasks raised by the stuck-work watchdog were the last to be brought in line. As of October 2026, I would not claim every older item has been rewritten. Treat the rule as applying going forward, with the backlog cleanup still in progress.

Which check can you run this week?

You do not need new software. You need your task list and one hour.

1. Export or list the open tasks that an automation, a bot or an AI tool created in the last month.
2. Pick ten at random.
3. For each one, ask a teammate who did not write it: could you start this in five minutes, with no questions? Score yes or no.
4. For every no, note what was missing: steps, location, definition of done, or access.
5. Add the missing field to your task template so the next one arrives complete. If your tool allows a required field, require it.
6. Check how many of the same ten are stale. If the tasks that scored no are also the ones that sat, you have your answer.

Run it again in a month. The goal is to see the “could start in five minutes” score climb, and to see old tasks stop collecting dust.

Where does this fit?

Every tomte, our word for one production agent with one defined job, ends up writing to a human at some point. The quality of that last message decides whether the work completes. A fractional AI officer sets the template, the ownership rules and the escalation path so that every agent in your operation hands off cleanly, and keeps them current as you add more. You can see the fleet this practice comes from on the agent fleet proof page.

If you want to see how an agent-created task looks in an operation like yours, start async. We reply within one business day, and there are no meetings.