I can delegate more work to AI and still be just as busy.
Producing an output is not the same as completing delegated work. If you still have to choose between proposals, correct the result and approve every step, much of the work remains in your hands. To know whether AI is helping, you need to count that effort too.
The coordination still needs doing
I see this in my own work. I enjoy building with AI, and my agents genuinely move my projects forward. While one is working, I tend to start another task. With several projects running in parallel, the results and questions eventually pile up. Before stepping in, I have to recover the context and decide what comes next.
That availability sometimes keeps me tied to work, even on vacation. An idea comes along, and I know I could move it forward immediately. Paying for that capacity also makes me want to use it.
I am still adjusting how I work. What I want is to finish more useful work without multiplying the things waiting for a decision from me.
Here are three questions to clarify before delegating.
1. What does finished work look like?
A general request leaves plenty of room for interpretation. Be more specific about what you want to have in hand at the end.
The assignment needs to specify the deliverable, approved sources, constraints and points to check. The work stops at a version ready for your review.
The assistant now has an objective you can assess. It does not need to produce more variations, revisit the content strategy or keep the conversation going. It needs to deliver the requested result.
Also define what would make the result unusable: an unsourced fact, a promise you have not approved or an essential piece of information left out. An example of an acceptable result can make your expectations clearer than a long list of adjectives.
2. Which decisions can it make on its own?
If every detail comes back to you, you keep much of the coordination work. Useful delegation allows some room to act within clear limits.
The assistant can choose the order of the sections, shorten a sentence or remove repetition. You do not need to approve each wording change before it prepares the draft.
It must not, however, invent a number, add a commercial promise or send a document on its own initiative. External actions, sensitive information and commitments to a client need explicit rules and the necessary approvals.
Specify which sources it may use and which actions remain under your control. Match that freedom to the tool's actual capabilities and the access you provide. Clear instructions cannot fix missing data or give software a feature it does not have.
3. When should it come back to you?
Some questions genuinely block the work. Others concern details the assistant can resolve within the assignment's limits.
Ask it to flag a contradiction between two sources, missing essential information or a decision outside its authority. If it can prepare the rest without making a risky assumption, it should continue and group the questions for your review. Otherwise, it should stop and explain exactly what is missing.
Specify what the handoff should contain: the result, the checks completed and the decisions that remain yours. You need to understand what is ready and what requires your judgment without reading a running account of every step.
This gives you clear checkpoints. Approval before an external action remains in place, even as you reduce the back-and-forth during preparation.
Measure all the work involved
To try this approach, choose a limited, recurring task. Write a few lines answering the three questions, then test the assignment on a real case.
Compare it with your usual way of working. How much time does the complete task take, from the start to a usable result? Include preparation, answering questions, checking and corrections. Also note how many times you had to return to it and whether the quality is acceptable.
Fast output can hide a long review. An assignment that takes longer to define may become easier to repeat. Comparing a few real cases will show where the effort falls and where the process still needs improvement.
If the same errors keep returning, examine their cause before adding another instruction. The problem may be an incomplete source, a poorly defined step or a limitation of the tool. Sometimes you need to change the process or keep a step manual.
The number of busy agents or requests used is not enough to measure a gain. Look at the useful work that gets finished and the attention it still needs from you.
Make AI fit the way you work
I still enjoy building with these tools. Using them every day also helps me see what needs to be organized around them so they can fit into the work.
That is the thinking behind Nord Paradigm's AI Implementation Partnership: working directly with you on one concrete process, testing it and defining the human checks it needs.
If you already use AI but spend too much time directing it, we can work together to make it more useful in your real work. We start with one priority and a result we can verify.