The two-shifts problem: why AI slows teams down today

The two-shifts problem, defined

The two-shifts problem is the new tax AI added to knowledge work. For every piece of output your team ships, there is now a meeting shift and a cleanup shift. Most teams used to do one. Now they do both, and call it progress.

The cleanup shift is small per task. Five minutes to paste a transcript. Three minutes to write the right prompt. Eight minutes to fix the parts the model got wrong. Two minutes to drop it in Notion or Linear. None of it is hard. None of it was on anyone's calendar a year ago.

Multiply that by every recurring meeting and every owner, and the second shift quietly becomes a job. The painful version is that it's invisible. It hides inside the gaps between calls, on the commute home, after dinner. The work moved, but it didn't go away.

Where the second shift came from

It came from how we shipped AI: as a sidebar, not a participant.

Almost every workplace AI tool sits next to the work, not inside it. The meeting happens. Then someone goes to a different tab, opens an AI, hands it the artifact (transcript, doc, dataset), prompts it, reviews the output, and moves the result to wherever it needs to live. That last sentence has eight verbs. Each one is a step the human has to do.

This is not a bad design. It's the obvious one when the underlying model is a chatbot. Chatbots are reactive. They wait for prompts. They have no idea your team is in a meeting. So the only way to use them is to context-switch out of the meeting, do the AI work, and switch back. Multiplied by a busy week, that's the second shift.

The hidden tax, with numbers

Pick a typical product team. Five recurring meetings a week per person. Each meeting needs notes, decisions logged, follow-ups created. Pre-AI, this got done badly or not at all. Post-AI, it gets done, but each meeting carries roughly 15 minutes of cleanup work somewhere downstream.

That's 75 minutes per person per week. Across an eight-person team, that's 10 hours. Across a quarter, it's a full sprint of engineering capacity spent on the second shift, paid quietly out of evenings and small gaps in the day.

The kicker: the team would have called this overhead obvious before AI showed up. We tolerate it now because the AI looks helpful in the moment. Each individual prompt feels like a win. The compounding cost is invisible because no one is timing it.

Microsoft's own Work Trend Index has been pointing at the same gap from a different angle: a majority of knowledge workers say they don't have enough uninterrupted focus time. The second shift is one of the things eating it.

Three patterns I see in the wild

When I talk to teams that already use AI heavily, the second shift shows up in three repeating shapes.

The recap stayer

Someone, usually the most senior person in the room, stays online after the meeting to write the recap. They do this because the AI summary missed the actual decision, or named the wrong owner, or buried a critical follow-up under filler. The recap takes longer than the meeting did because they're now translating audio in their head while editing a draft they didn't write.

The prompt librarian

One person on every team keeps a private doc of prompts that work. "Summarize this meeting in 5 bullets, owner-action format, ignore filler before timestamp 02:00." It's an entire skill, learned in spare time, and the team's AI quality depends on it. If the prompt librarian goes on vacation, the recaps get worse.

The decision miner

Three days after a meeting, someone messages the channel: "wait, did we agree to ship this on Friday or next Friday?" Then someone else opens the transcript, scrubs to the right minute, copies the relevant text, pastes it into a chat with their AI, asks it what was decided, and posts the answer. The decision was made in the meeting. It got mined out of it three days later.

Each pattern is rational. Each pattern is the second shift in disguise.

Why "use AI more" doesn't fix it

The instinct, when you spot the second shift, is to add more AI. Better summaries. A meeting bot. An agent that watches Slack. Sometimes it helps. More often it adds a third tool and a fourth shift.

The reason: the second shift is a structural problem, not a tooling one. It exists because the AI is outside the moment work happens. Adding more outside-AI stacks more cleanup. The fix is to move the AI inside the moment, so the work finishes before the meeting ends.

If the meeting isn't closed when people leave the call, you have a second shift, no matter how many AI tools you bought.

What one-shift work looks like

One-shift work has a simple test. When the meeting ends, the meeting is done. The recap is in the right tool. The decisions are logged with owners. The follow-ups are created in Linear or Asana with the right context. The next person who needs the answer can find it without opening a transcript.

To get there, the AI has to do four things while the meeting is still happening:

1. Listen with shared context. It knows what your team decided last week, what the doc says, who owns what. Not just the audio in front of it.
2. Move on its own. If a decision lands, it logs it. If a follow-up is named, it files the ticket. No one has to remember to ask.
3. Speak when useful. Not constantly. When the room is missing context the AI has, it offers it. The rest of the time, it stays quiet.
4. Land work in the right tools. Slack, Notion, Linear, Gmail, Calendar. The output goes where the team already lives, formatted correctly the first time.

Do those four, and the meeting closes itself. The second shift never starts. (We wrote up the design constraints for this in how AI should behave in a real meeting.)

The bar voice AI has to clear

Voice AI is the category that can do this, but only if it clears a real bar. Bots that just transcribe and email a recap are the second shift dressed up. The honest test:

• Did the meeting end with decisions filed in the right tool, with owners?
• Did anyone have to type a prompt during the call?
• Did anyone have to clean up the AI's output before it was useful?
• Could a teammate who missed the meeting catch up in 60 seconds, without opening a transcript?

If the answers are yes / no / no / yes, the second shift is gone. If any of them flip, you still have it.

Where relly sits in this

relly is built around exactly this test. It joins your meeting on Zoom, Google Meet, or Microsoft Teams, holds shared team context across calls, drafts decisions and follow-ups while the room is still talking, and files the result into the tool that owns it. The goal is one shift, not two.

If your team is doing the second shift right now and you can feel it, early access opens before the public launch with 50% off your first 12 months. No card needed until launch.