## Key Takeaways

- **The time AI saves is being spent supervising AI.** In the [Work AI Index 2026](https://www.glean.com/work-ai-institute/reports/work-ai-index) from Glean's Work AI Institute, a survey of 6,000 digital workers, people report saving about 11 hours a week through AI automation and then spending 6.4 hours a week making AI usable.
- **Four buckets, and only some of the work is worth doing.** The same report splits that time into feeding context (2.3 hours), supervising output (2.2 hours), debugging (1.7 hours) and cleanup or tool switching (0.2 hours).
- **Re-pasting context is the biggest and most avoidable bucket.** If the AI already sits where the conversation and the files are, most of that typing disappears.
- **Reviewing high-stakes output is not waste.** Keep it, make it fast, and stop reviewing the same thing forever once it has been right twenty times.
- **Running the same request through three tools is a symptom, not a strategy.** It means no single tool has enough context to be trusted, so you compare guesses instead.
- **Track the supervision hours the same way you track the saved hours.** Untracked supervision work is why leadership sees a productivity story that nobody on the team recognizes.

Our ops lead ran a small experiment on herself in the spring. For one week she kept a note every time she typed something into an AI tool that the tool should already have known: which client, which quarter, which spreadsheet, which naming convention, what we decided in the thread two days earlier. By Friday the note had 41 entries. None took longer than four minutes. Together they cost her most of a working day, and the output still needed a read-through because the tool had never seen the underlying files.

That is the shape of the problem nobody budgets for. The work AI removes is visible and gets celebrated. The work it adds is invisible and lands on whoever is closest to the output.

![Where AI time actually goes](/images/blog/how-to-cut-time-spent-babysitting-ai/where-time-goes.webp)

## Why does AI create work while saving time?

Because most AI tools are strangers to your business. Every session starts from zero, so a person has to hand over the context first, then verify what comes back, because a stranger's confident answer cannot be trusted on sight.

The Work AI Index 2026 gives that hidden layer a name, botsitting, and a size: 6.4 hours a week, more than the time workers spend using AI to produce the work. The report also found 87% of digital workers use AI at work and 75% say it makes them more productive, while only 13% say their organization performs significantly better because of it. Those two findings only fit together if the gains are being eaten somewhere between the desk and the business result.

Our own read, from watching teams do this daily: the leak is almost always structural rather than personal. Nobody is prompting wrong. The tool simply has no memory, no access, and no shared thread with the team, so a human keeps stepping in as the connective tissue.

| Where the hours go | What it looks like on a Tuesday | Worth keeping? |
| --- | --- | --- |
| Feeding context | Pasting the same client background, last quarter's numbers and the tone rules into a fresh chat | No. The system should already have this |
| Supervising output | Reading a summary against the source spreadsheet before it goes to a client | Yes, on anything that leaves the building |
| Debugging | Re-prompting, adding detail, switching models until something usable appears | Rarely. Usually a missing-access problem in disguise |
| Comparing tools | Running the same request through three different assistants because none felt right | No. Pick the one that can see your work |
| Cleanup after handoff | Fixing work a colleague shipped without reading it | No, and it lands on the wrong person |

## How do you kill the context tax?

Stop retyping what already exists in your tools, and put the AI where the context lives instead of carrying context to the AI.

Concretely, that means three things. The work chat where the decision was made is where the request should be made, so the thread itself is the brief. The tools that hold the facts should be connected with real read access, so numbers get pulled rather than pasted. And the corrections you make should be stored, so the second request starts from the corrected version rather than from the original mistake.

The difference is easiest to see in one request. Here is the same job in a tool that starts from zero versus in a channel where the AI employee already has access:

```prompt
@Viktor pull the July numbers for the Northvale account from Stripe and HubSpot, compare them to June, and post a short read on what moved in this thread. Use the same format as the one you did for Redwell.
```

Nothing in that message explains what Northvale is, where the data lives, or what the format should be. That is the point. Nine of the 41 entries in our ops lead's note were format instructions she had already given.

A useful test before you adopt anything: give it a task on Monday, come back on Thursday and reference that task without re-explaining it. If the thread continues instead of restarting, the context tax mostly goes away. If it asks you to restate everything, you have bought yourself another supervision job. We wrote the longer version of that test in [how to give your AI employee memory](https://viktor.com/blog/how-to-give-your-ai-employee-memory).

## Which supervision should you keep?

Keep review on anything that touches money, customers or the public record, and make the review cheap rather than optional.

Cheap review has two properties. The output arrives next to its sources, so checking it is a glance rather than an investigation. And it arrives as a draft in the same thread as the request, so approving it is one reply instead of a copy-paste into another tool.

- **Money and legal:** always reviewed, no exceptions, no matter how many times it has been right.
- **Client-facing writing:** reviewed until the voice is consistently right, then spot-checked.
- **Internal reporting:** reviewed for the first month, then reviewed monthly against the source.
- **Research and first drafts:** read, not audited. The cost of a bad draft is a rewrite.

The report's own framing is worth borrowing here: verifying high-stakes output is productive supervision, and reloading context or comparing three tools is not. Sort your week into those two piles and cut the second one.

![Two kinds of supervision](/images/blog/how-to-cut-time-spent-babysitting-ai/two-piles.webp)

## What about the work nobody sees?

Put a number on it, because untracked effort never gets fixed.

The Work AI Index found that frequent botsitters are 73% more likely to be actively looking for another job. Whether or not that holds in your team, the mechanism is familiar: someone quietly absorbs the cleanup, gets no credit for it because it is not on any plan, and eventually stops absorbing it. The version of that failure we see most often is a colleague who stops checking output at all, which moves the mess downstream to someone with even less context.

Two habits fix most of it. First, name an owner per recurring AI job, so cleanup has an address instead of landing on whoever opens the file. Second, count corrections. If ten outputs a week need a human fix and that number never falls, the corrections are not being stored anywhere, which is a tool problem rather than an effort problem. We break the counting down in [how to tell if your AI employee is paying off](https://viktor.com/blog/how-to-tell-if-your-ai-employee-is-paying-off).

```prompt
@Viktor every second Friday, list the recurring jobs you run for us, how many times each one needed a correction in the last two weeks, and which ones nobody has read the output of. Post it in this channel as a table.
```

That request is deliberately uncomfortable. An AI employee that reports on the jobs nobody reads is more useful than one that only reports wins.

## How do you get the first hours back this month?

Pick the single job you re-explain most often, move it to where your team already talks, and connect the two tools that hold its facts.

The sequence that works: choose one weekly job, write the request once with its sources named, keep review on for a fortnight, then correct in the thread every time something is off. After two weeks you are not typing the brief anymore and the output arrives before you ask. That is the whole mechanism. It is unglamorous and it compounds.

What not to do: buying a second AI tool because the first one keeps missing context. That trades a context problem for a switching problem, and the Work AI Index found 77% of workers already juggle multiple AI tools weekly, with a third using four or more. If you want the deeper version of that argument, [the hidden cost of tool sprawl](https://viktor.com/blog/the-hidden-cost-of-tool-sprawl) covers it.

## Frequently Asked Questions

### What is botsitting?

Botsitting is the term the Work AI Index 2026 uses for the unrecognized work of making AI usable: feeding it context, supervising its output, debugging its mistakes, and cleaning up after it. The report puts it at 6.4 hours a week for the average digital worker.

### Why do I spend so long explaining things to AI?

Because most AI tools have no memory of your business and no access to your systems, so each session starts from nothing. Reducing that means giving the AI persistent memory plus real read access to the tools that hold the facts, rather than writing longer prompts.

### Is reviewing AI output a waste of time?

No, on anything that reaches a customer, a regulator or a set of accounts. It becomes waste when it is the only defence, when the same thing gets fully re-audited forever, or when the review requires opening five tabs to find the sources.

### Does using more AI tools help or hurt?

Usually it hurts. Running one request through several tools produces answers to compare rather than an answer to use, and the context has to be loaded into each of them. One tool that can see your actual work beats three that cannot.

### How do I measure the time my team loses to supervising AI?

Keep a one-week note: every time you retype context, re-prompt, or fix AI output, log the task and the minutes. Most teams find the total lands close to a full day per person per week, which is enough to justify changing the setup.

### Where does an AI employee fit differently from an assistant?

An AI employee lives in the work chat, keeps memory across weeks, and has access to the tools where the facts sit, so the context work happens once instead of at every request. An assistant in a separate tab needs the context handed over each time.

### Can I stop reviewing eventually?

On narrow, repetitive jobs with a clean record, yes, and it should be a deliberate decision rather than drift. We set out where that line sits in [when to let your AI employee act without review](https://viktor.com/blog/when-to-let-your-ai-employee-act-without-review).

## Where to start

Take the job you have re-explained three times this month. Move the request into the channel where the work already gets discussed, connect the two systems that hold its data, keep review on for two weeks, and correct in the thread instead of in your head. Then count the minutes you did not spend on it.

---

**Viktor is an AI employee that lives in Slack, connects to 3,200+ integrations, and does real work for your team.** [Add Viktor to your workspace -- free to start →](https://viktor.com/?utm_source=blog&utm_medium=cta&utm_campaign=how-to-cut-time-spent-babysitting-ai)

Related reading:

- [How to give your AI employee memory](https://viktor.com/blog/how-to-give-your-ai-employee-memory)
- [How to tell if your AI employee is paying off](https://viktor.com/blog/how-to-tell-if-your-ai-employee-is-paying-off)
- [The hidden cost of tool sprawl](https://viktor.com/blog/the-hidden-cost-of-tool-sprawl)
- [When to let your AI employee act without review](https://viktor.com/blog/when-to-let-your-ai-employee-act-without-review)