Should I let an AI reply to my email in my voice?

Should you let AI reply to email in your voice?

For inbound triage and first-draft replies, voice-mimicking email AI like Goldfish saves real time. For relationships that matter, never let it send unattended. The brittle part is not the prose — it is judgment about what the relationship needs right now, which the model cannot see.

Goldfish launched on Product Hunt in mid-June 2026 with one of the more honest taglines in the AI inbox space: "an AI that understands your work and replies in your style." Several friends have asked whether they should turn it on, or its competitors (Shortwave's Voice, Superhuman's AI replies, the new ChatGPT inbox integration). I have been running a voice-trained reply layer over my own email for six weeks. I want to write down what I have actually learned, because the marketing and the real experience diverge in ways that matter.

The good news first. Once a voice-mimicking email AI has read a couple of months of your sent folder, the prose it generates is genuinely uncanny. It picks up that I open with the person's first name, that I write short paragraphs, that I close with a verb-led next step rather than "let me know." For the right tier of email — recruiter outreach, light vendor pings, scheduling threads, "thanks for the intro" replies — the drafts are 80–90% there. I save an honest fifteen to twenty minutes a day, every day, just by accepting first drafts with one or two edits. That is real. If anyone says these tools don't work, they have not used the current generation.

The bad news, which the tools don't advertise. The thing your real voice carries that the model cannot replicate is judgment about the relationship, not stylistic tics. When a founder I have coached for a year writes me a long, slightly-too-cheerful email, the actual signal is usually that something hard is happening underneath. A voice-trained model reads the surface text and produces a chipper reply that matches the tone of the message. A human who knows them writes back two sentences asking what's actually going on. The model doesn't know that distinction exists. It cannot, because the signal is not in the email — it is in the history of the relationship and what's at stake in it.

This is the failure mode Paul Bloom describes in Psych when he talks about the fundamental attribution error and how much of communication is situational rather than dispositional. We confuse our own consistent voice across emails with our consistent character. They're not the same thing. My voice in an email to a board member is not the same as my voice in an email to a coaching client in crisis, even when the surface vocabulary looks identical. A model trained on the surface cannot tell those situations apart, which means that the more the model speaks for you, the more your communications converge toward an average self that is no one's friend.

The Co-Active coaching framework names a related principle that I take seriously here: that change in someone's life often happens in the small, deliberately personal moments — not in the polished correspondence. If those moments are also being written by your AI, you have outsourced the actual relational equity of your work. Cal Newport in Slow Productivity makes the more pragmatic version of the argument: high-leverage work is rare and dense, and protecting the few interactions that compound is more valuable than draining the inbox faster. Email AI is excellent at draining the inbox. It is dangerous at the compounding interactions, because it makes them look like inbox work.

The three rules I have settled on after six weeks. First, the AI drafts but never sends unattended. Auto-send modes ("respond to anything from a stranger within an hour") are tempting and they will eventually send something wrong to the worst person. Second, any thread with a person who matters — a board member, a customer, a coaching client, a friend — gets manually written by me, not edited from a draft. I learned this the hard way after sending an AI-drafted reply to a long-time mentor that read fluently and said almost nothing personal; he noticed, and told me. Third, every two weeks I read my own outbox, not to audit the AI but to audit myself: am I still recognizable in here? If the answer is "barely," the dial is set too high.

So the honest verdict on Goldfish and the category: yes, turn it on if you process more than fifty emails a day and you can tell yourself the truth about which threads matter. The category will save you real hours. But the part of email that is actually executive work — knowing which message to send personally, which to delay until you can be present, which to pick up the phone for — is exactly the part the model cannot help with. Buy the tool to free up time for that judgment. If you buy it to avoid that judgment, you have automated the wrong layer of your life, and you will not feel it for months.