Stop using AI like a chatbot: think like a manager

Most people start with AI the same way: type a question, read the answer, type the next question. It works. But it treats a capable agent like a search box with manners, and it leaves most of the value on the table. The shift that changes everything is simple. Stop thinking of it as one assistant. Start thinking like a manager.

A chatbot answers. An agent does the work

A chatbot answers questions. An agent does the job. It reads your files, searches the web, writes documents, runs tools and checks its own results, step after step, until the work is done.

That changes what you can ask for. “Write me an email” is a chatbot request. “Go through these 40 quotes, compare them, and draft the email to the best two suppliers” is an agent request.

You stay in charge. The agent works with your rights, shows you what it did, and you decide what is good enough.

Not one assistant: a manager with a team

Working with one person, you are limited by their brainpower, their memory and the hours in their day. An AI agent that can run helpers is more like a manager who can hand work to as many staff as the job needs. Each helper, called a subagent, takes one part of the job and reports back.

So the right question is not “what can one assistant do this afternoon?” It is “what would I give a whole department?”

Here is the same request sized two ways:

• One-assistant version: “Find me a cheaper courier.”
• Manager version: “Find every courier that services our three regions, compare their prices for our five most common parcel sizes, check reviews for late deliveries, and recommend the best two with reasons.”

The second would take a person days. A manager with staff puts a helper on each region at once and has a comparison back while you do something else.

Try saying:

Treat this like a team job. Put a helper on each of our three regions, have each one find the couriers there and their prices, then give me one comparison table and your top two picks.

Size the job to the team, not to one person. Ask for ten options instead of one, every file instead of a sample, three independent reviews instead of a quick look.

Delegate outcomes, not steps

You would not tell a good employee which keys to press. You tell them what you need, why, and by when. Agents work the same way. Dictate every step and you get exactly your steps, including the ones you got wrong. Describe the outcome and the agent can find a better route.

• Steps: “Open the March sales file. Add up column D. Then open February and add up column D. Tell me the difference.”
• Outcome: “Find out why March sales are lower than February’s, and tell me which customers explain most of the gap.”

The first gets you a number. The second gets you an answer, and the agent may spot that a big customer simply paid late.

Say what done looks like

The most common reason an agent hands back the wrong thing is that nobody said what the right thing was. Tell it the finished form.

Try saying:

Done looks like this: a one-page Word document, written for the board, with the three biggest causes, the dollar amount for each, and one recommendation.

A good brief has four parts: the goal and what it is for, the context, what done looks like, and the limits (what not to touch, and what to ask you about first).

Check the plan, not every step

For a long job, ask for the plan before the work. Fixing a plan takes one message. Fixing a finished report built on the wrong idea takes many.

Try saying:

Before you start, show me your plan in five lines or fewer. Wait for my go-ahead.

Helpers, in plain words

You do not need special commands to get helpers. Ask the way you would ask a manager to put people on something.

Try saying:

Give each of these four supplier contracts to a separate helper. Each one should list the payment terms, the notice period and anything unusual. Then bring their notes together for me.

A good manager also does not give every task to the most senior person. Mix sizes: a small, fast model for gathering and sorting, a strong one for the hard thinking.

Try saying:

Use Haiku helpers to pull the key dates out of each of these 30 leases. Then use Claude Opus to check the ones that look unusual and write up the risks.

In Interweave OS, helpers can be any model from any family your organisation connects, and a panel beside the chat shows each one, its model and how long it has run. You can click in to read a helper’s work as it happens.

Parallel work: split it, then bring it together

Anything that splits into separate parts can run at the same time: one helper per supplier, per document, per region, per month. Ten helpers working at once take about as long as one, so a job that would fill your week can come back in an afternoon.

What splits well:

• Parts that do not depend on each other. Reviewing twelve CVs splits well. Writing chapter two after chapter one does not.
• Parts with the same shape of answer. Ask every helper for the same headings or columns, so the results fit together.
• Parts big enough to be worth a helper. Ten one-line questions are faster asked in one message.

Splitting is the easy half. The value is in the join: one table, one recommendation, and a clear note of where the helpers disagreed or something does not add up. Say how you want it combined when you hand out the work.

Try saying:

Split this market research into five parts, one per competitor, and run them at the same time. Each helper uses the same headings: prices, strengths, weaknesses, recent news. When all five are done, put them in one table and point out anything that does not add up.

Picture Priya, who needs to compare quotes from eight printers before Friday. She asks for one helper per quote, each pulling out price, turnaround and paper stock into the same columns. The main agent builds one table, flags that two quotes leave out delivery, and drafts follow-up emails to both. Priya reads the table and sends the emails.

Use another family for reviews

Models from the same family tend to share blind spots. A good manager gets a second pair of eyes from someone trained differently. If Claude did the work, have a GPT, Gemini or Grok helper review it, and the other way round. Ask for the review to be done independently, without seeing the first answer, or it tends to agree.

From doing to reviewing

Once agents do the work, your job changes. You spend less time making things and more time deciding whether they are right. That is the manager’s job, and it is where your experience matters most.

Review, do not redo. When something is wrong, say what and why, and let the agent fix it. “Too formal, these are long-time clients” gets a better second try than quietly fixing it yourself.

Do not hover. You do not need to watch it work. Hand out the job, get on with your own work, and come back to the result.

Keep the manager’s head clear. A chat can only keep so much in mind at once. On a big job, ask the main agent to give the reading to helpers and keep only their summaries, so it holds the whole picture for longer.

Try saying:

This is a big job. Don’t read the files yourself. Give the reading to helpers, have each one send back a short summary, and keep your own attention on the overall plan.

Own the result. Whatever the agent makes, you are responsible for it. Check in proportion to the stakes: a draft email needs a read, a figure in a contract needs its source.

A manager’s morning

A useful habit: start the day by handing out the long jobs, then do your own work while they run. Review the results between meetings. And before starting any task yourself, ask one question: could I hand this over? Research, first drafts, comparisons, summaries, checking and tidying are usually yes.

The things that stay with you are the things that always mattered most: relationships, judgement calls, and knowing what really matters to your clients.

How Interweave OS does it

In Interweave OS, every chat’s agent works as a manager with an unlimited team. Ask in plain words and it gives each part of a job to its own subagent, in whichever model and family suits that part, all working at once. The main chat keeps only what it asked for and what came back, so it stays quick and focused, and you stay free to keep talking while the team works.