Blog Book a call

Duet Fails at Judgment. That's When AI Automation Always Fails.

Duet Fails at Judgment. That's When AI Automation Always Fails.
Short answer

Claude's Duet agent works well when it collects data and surfaces what you already need to see. It fails when it tries to understand tone and make judgment calls. I see the same pattern in every automation I try to fix: the ones that work execute a clear instruction fast and reliably. The ones that stall are the ones where I asked the AI to be smart.

Anthropic's Claude Duet agent had one job: be a personal assistant who understands your work life, picks up on nuance, and knows what matters. A week into testing it for The Verge, the reviewer wrote that Duet 'felt like interacting with a slightly neurotic personal assistant'. It spotted a draft waiting for edits before 9 AM and panicked, missing that the editor was still asleep on the west coast. It drafted a Slack message that was a little too good, shortening 'signature' to 'sig' in a way the reviewer never would. Duet tried to be smart about tone and nuance, and it failed. The best part of the review: 'A person can pick up on nuances in tone to understand what's a big deal and what isn't. A person will pick up on communication norms and deploy them sensibly.' This is exactly the gap I have hit in every automation that stalled, and the reason the ones that worked actually worked.

What Duet Got Right

The review also described what Duet did well. It could read email, calendar, Slack, and Google Docs, and spot when something said in a meeting connected to a direct message the reviewer had missed. It offered to draft a Slack message and update a shared doc with weekly notes. It highlighted a planning gap the reviewer did not catch: time off would leave only a couple of working days before a big draft was due. The reviewer called this 'surprisingly helpful' and said it worked best as 'a backstop for my own organizational systems, not as a replacement for them.' That backstop job is exactly what I built that actually moved the needle. A daily sales report used to come out in a format nobody had time for and nobody read it. I changed the delivery: now a short summary lands on the owner's WhatsApp every morning, and it actually gets read. The data did not change. The order did not change. What changed was where it arrived and how long it was. Most reporting problems are not data problems. They are delivery problems. The right number is already there. It is just in a place nobody looks, in a format nobody has time for.

Where It All Falls Apart

Duet tried to do more. It wanted to help write things. It tried to anticipate what mattered and what did not. It attempted to understand whether something was a real deadline or a soft commitment. It failed every time it had to make a judgment that depended on understanding tone, context, or what matters most to the person using it. This is the pattern I see when an automation stops working. I automated three things at my own distribution business: reporting, stock tracking that watches for stock-outs and dead stock, and the path from an order coming in to it being dispatched. Those three things save the team roughly 800 to 900 hours a month. That is a rough estimate, not a measurement, but even that number shows what I learned: most of those hours were paperwork about the work, not the work itself. The stock tracking does not decide whether to order; it just flags when stock drops below a level we set. The order-to-dispatch system does not decide who gets the job; it just moves the order from one queue to the next and marks off every step. Reporting does not interpret results; it assembles what happened and delivers it. None of these automations need to be smart. They need to be fast, consistent, and get in front of the person who does have to be smart.

What This Means When You Build Automation

Duet's problem was that Anthropic built it to be a personal assistant, not a tool. A personal assistant has to understand tone. A tool just has to execute an instruction correctly every time. When you decide what to automate in your practice or your business, look for the work that has three things: it repeats, it follows a clear rule, and someone has to do it the same way each time anyway. Do not automate the decision. Automate the work that comes after the decision, or the work that surfaces what someone needs to decide about. A reminder that goes out the day before an appointment is automated. Deciding who needs to be reminded is not. Generating a list of overdue invoices is automated. Deciding which ones to pursue is not. Drafting a follow-up message is automated. Deciding when to send it is not. Duet wanted to be smart. Every automation that collapsed on me wanted to be smart too. The ones that work never try to be.

The Cost of Scope Creep in Automation

There is another reason some automations fail that the Duet review hints at: scope. Duet got access to email, calendar, Docs, and Slack all at once. That sounds powerful. It meant the system could see everything and connect things. It also meant the system could be wrong about everything, all at the same time, in places the owner was already drowning in information. An automation that runs a few times a day and does one thing can stay cheap to run and easy to fix. An automation that fires every time someone sends a message, or every order comes in, or every day at a scheduled time, can grow expensive to operate if you build it wrong. I use n8n because it can run on a client's own server, which protects privacy. But I also use it because tools that charge per task add up the moment you automate something that fires at volume. Success becomes a growing bill. A self-hosted tool keeps the cost flat while the business grows, which is exactly when you want it to be. That matters more as you scale, but it matters from the start: keep the scope narrow and the execution clear, and the whole thing stays manageable.

A Question to Ask Before You Start

When Duet tried to draft a message about a story waiting for edits, the reviewer noticed something: Duet copied their millennial habit of writing in lowercase, pointed it out in 'a slightly self-congratulatory aside', and used an abbreviation they use for their CMS. The system was trying to be person-like. It was trying to be them. This is the moment every automation I have seen fail tries to make the jump from doing work to being the person doing the work. Do not let yours. Before you automate anything, ask whether the system needs to make judgment calls or understand tone. If the answer is yes, do not automate it yet. Automate the part before it, or the part after it. Automate the delivery, not the thinking. That is what works.

Sources

Working together

Got something like this to automate?

We take on a small number of builds at a time, as ongoing engagements with a defined scope rather than one-off tasks. If you have a process that is costing your team real hours every week, tell us what it is and we will tell you straight whether it is worth automating.

Keep reading