Delegate to AI. Don't outsource to it.

Delegating work to AI makes you faster and better. Outsourcing work to AI makes you faster and, sooner or later, wrong. The two words sound alike, and most AI mistakes, hallucinations and “slop” come from mixing them up.

Two words that sound the same

Delegating is handing someone a task while you keep the outcome. You set the intent, explain what “done” looks like, check the result, and put your name on what ships. The work moves to someone else and the accountability stays with you.

Outsourcing is handing over the task and the judgment. You describe roughly what you want, accept what comes back, and pass it along.

In business, outsourcing can work because there’s a contract around it: a vendor who answers for the work and a service level you can hold them to. AI has none of that, and a model can’t be held accountable for what it writes. So when we “outsource” to AI, the accountability doesn’t move to the tool. Nobody holds it until someone downstream finds the mistake, sometimes a judge.

An old management principle applies: you can delegate authority, but you can’t delegate accountability. AI doesn’t change that, though it makes it easier to forget.

What outsourcing to AI looks like

The headline AI failures of the last few years share one pattern: polished output that nobody owned.

In Mata v. Avianca (2023), two attorneys filed a brief citing court decisions that ChatGPT had made up, and kept defending them when questioned. The judge fined them and their firm $5,000, and also noted there’s nothing “inherently improper” about using AI, but that the rules give attorneys “a gatekeeping role” to ensure their filings are accurate.

In Moffatt v. Air Canada (2024), the airline’s chatbot told a grieving customer that a bereavement discount could be claimed after flying. The airline’s own policy said otherwise. Air Canada argued the chatbot was responsible for its own answers, and the tribunal disagreed: the company was responsible for everything on its website, chatbot included.

In 2025, Deloitte agreed to partly refund the Australian government for an AU$440,000 report that contained fabricated references and a made-up quote from a court judgment. The corrected version disclosed that a generative AI tool chain had been used.

The everyday version is smaller and far more common. Researchers from BetterUp Labs and Stanford, writing in Harvard Business Review, named it “workslop”: content that looks polished but lacks real substance and pushes the thinking onto whoever receives it. They report that 41% of workers have run into it, and that each instance costs nearly two hours of rework.

In every one of these cases, AI output reached a court, a customer or a client before anyone had taken ownership of it.

Why outsourcing to AI is unstable

Language models are built to produce plausible text, and plausible isn’t always true. Most of the time the two line up. When they don’t, the wrong answer arrives in the same confident tone as the right one, which is why hallucinations are hard to spot: they don’t look like errors.

Outsourcing also skips the checks we already trust. Agile teams have a definition of done, acceptance criteria, peer review and the sprint review, because every producer of work, human or not, makes mistakes. Outsourced AI output tends to bypass all of them, because it looks finished.

Errors compound downstream. A wrong number in a status report becomes a wrong decision in a steering committee, which becomes a wrong plan for next quarter. The further an unchecked output travels, the more it costs to unwind.

And your own judgment gets rusty. If AI always writes the first and last draft, you stop practicing the skill that lets you spot when it’s wrong.

What good delegation looks like

Delegating to AI looks a lot like delegating to a capable new teammate.

1. Start with intent. What outcome do you need, for whom, and what does done look like? It’s the same discipline as writing a good user story with acceptance criteria.
2. Give context and constraints: the audience, the tone, the data to use, and the data not to use.
3. Slice it small. Ask for an outline before a full draft, one section before ten. Small batches are easier to check, the same way small stories are easier to review than one giant one.
4. Review it like a teammate’s work. Check every number, name, date and source. If you can’t trace a claim to something you can open, it doesn’t ship.
5. Own the result. If your name is on it, you should be able to defend every line without saying “the AI wrote that.”

A few examples from everyday project work:

• For a leadership status report, outsourcing means pasting the Jira export, asking for an executive update and forwarding it. Delegating means deciding what leadership needs to hear, letting AI draft from the export, then checking every figure against the board and rewriting the risk section in your own words.
• In a retrospective, AI can cluster forty sticky notes into themes in seconds. The team still decides which themes are true and what to act on.
• In backlog refinement, AI can draft acceptance criteria for a story. The product owner and the team still challenge them, because they know the edge cases the model doesn’t.
• After a meeting, AI can write the summary. A person confirms the decisions and owners before it goes out, because a wrong owner on an action item is worse than no summary at all.

A status report I almost sent

I once asked AI to help me build a project status report for five teams, based on their work and comments in Jira. It was a long report, and at first glance it looked exactly how I wanted: an executive summary, actions, follow-ups and a general summary.

Right before sending it, I read it a second time. For one team, the AI had done what I asked. It summarized their Jira stories and pulled examples from them. The trouble was that those stories were still full of placeholder text, so my report said Team A had all their work in place “thanks to their latest Lorem Ipsum implementation.”

Lorem Ipsum is one of the most common placeholder texts there is, and the AI had treated it as the name of a real implementation. The summary was accurate to the data. Only a careful read showed that the data itself wasn’t real.

A quick test before you hit send

• Could I explain and defend every line of this?
• Do I know where each fact and number came from?
• Would I be comfortable if the reader knew exactly how this was made?

If any answer is no, you’ve outsourced, and the work isn’t ready to send.

Questions to sit with

• Where in your week have you drifted from delegating to outsourcing?
• What do you check when a teammate hands you work that you skip when AI hands it to you?
• Which skills do you want to keep sharp, even if AI could do them faster?
• If your AI-assisted work were wrong, who would notice first: you, or the person relying on it?

Keep the reins

In my own work, I mostly use AI to build systems and reusable skills rather than to hand off the work itself. I keep several human-in-the-loop checkpoints along the way, and I keep using, testing and refining those skills until I trust that the rework left for me is minimal. Trust, but still verify.

AI is one of the most useful teammates any of us has had. Use it to move faster and clear out the busywork, and keep the reins yourself. The work is still yours, and so is getting it right.