By JayOctober 2026
You Are the Bottleneck
When AI drafts your email, your quote or your summary, you are the step that finishes it. The draft arrives in seconds. The finishing happens at your speed, one item at a time, and that is where the work now waits.
A software engineer named the same problem from the other side. After a conference full of "software factory" demos built around task boards, Josh Bleecher Snyder argued in an exe.dev post on running fewer agents that starting a new agent every time you get blocked only moves the cost onto the human.
concurrency is really rough on humans. It's stressful. It trashes flow state and thrashes our mental page caches.
Every draft waits on a person
In Using AI Isn't the Same as Running On It we looked at two 2026 surveys of small businesses. Of small business workers already using AI, only 6% automate workflows with minimal human involvement. The other 94% still do the work themselves, with AI's help.
Picture an owner who has AI draft ten customer quotes a week. The drafting takes minutes. The reading, correcting and sending takes the owner's evening, every week, and a second AI tool drafting follow-ups or invoices means more to read.
Adding another assistant adds another thing to check, and a person's attention does not scale the way a subscription does.
Every person in the loop is slow, and checking by eye is unreliable
Every review step costs time. In OpenAI's GDPval paper, using GPT-5's output as it came was 90 times faster than an unaided expert. Once an expert checked each result and did the task themselves when it fell short, it was only 1.12 times faster. For the older GPT-4o, the checking made the job slower than the expert working alone. These are October 2025 models, so read the ratios as direction, not as current speeds.
A check by eye is also unreliable. In a 2023 radiology experiment, 27 radiologists (the doctors who interpret scans) read scan results with a suggestion presented as coming from an AI, and some suggestions were wrong. Very experienced radiologists got 82.3% right when the suggestion was right and 45.5% when it was wrong. Inexperienced radiologists got 79.7% and 19.8%.
Experience reduced the pull of the machine's answer without removing it.
Which is why the check should not rest on a person's eye alone.
Decide in advance what deserves your attention
Most interruptions are the tool asking a question it did not need to ask. Anthropic's guide to Opus 5.5 suggests adding a rule to your AI's saved instructions that says when it may interrupt you: "Stop and ask only when you can't continue without me, or before anything destructive". The owner version names what a finished quote looks like and when to stop and ask, for example A finished quote has the customer's name, the itemized total from my price list and a due date. Stop and ask me only if a price is missing.
Everything outside that sentence should arrive finished. What to Stop Prompting Now covers the matching habit of putting standing rules in the instruction file instead of retyping them in chat.
Put a check between the AI and you
If you confirm every action by hand, you are the quality control for everything the AI touches. In A Change Is Not Done When the Command Succeeds we described the habit we run instead: check the same change from both ends, independently, so a failure surfaces before a person has to look for it. For an owner that can be as small as a second AI pass that checks each quote's totals against your price list before the quote reaches you.
Checking is a job, and it should not default to the busiest person in the building.
Our own publishing runs on the same principle. Every new post on this site passes a tone reviewer and a fact validator, both agents, before a person reads it. The validator caught a sentence in What to Stop Prompting Now that credited When Your AI Agent's Instructions Contradict Themselves with a re-testing method that post never describes. The fix landed before a person read the draft, so the person's time went to judgment calls and not to fact-checking.
Automate the whole task, not the drafting
Drafting faster leaves you in the loop on every item. Taking the task out of the loop takes two parts:
- An agent that runs the whole recurring job.
- A review layer that shows you what the agent did, so you are not checking every run by hand.
That is the Plan, Deploy, Review process addAI.dev runs for every client: we map the task with you, build it, and give you the visibility to trust it. You stay for the judgment calls, and those evenings go back to you.