Perplexity email humanizer

An email drafted in Perplexity fails on three separate axes, and it is worth separating them because each has a different fix. The lexical axis is stock phrasing. The structural axis is the silhouette: formal opener, three matched body sentences, an empty closing. The provenance axis is Perplexity-specific residue, meaning bracketed citation numerals, attribution constructions and a closing caveat about currency.

This rewrite is pointed at all three. It runs on our AI server, accepts 1 to 12,000 characters, allows 5 free runs per rolling 24 hours without an account, and exposes one optional instruction field of 300 characters for constraints such as a word ceiling.

These phrases are not bad English. They are high-frequency and low-information, which is a different defect: the reader has processed them so often that they carry no signal, so the first line of the email does no work.

"I hope this email finds you well"

Maximum frequency, zero information. It is the single most recognisable opener in professional correspondence.

"I am writing to inform you"

A performative that restates the medium. The information itself is a strictly better opening.

"As per our previous conversation"

Formal register standing in for a specific referent. Naming the call or the thread carries information; this does not.

"Please do not hesitate to reach out"

A closing with no requested action, in a document whose purpose is usually to request an action.

Agentless passive

"It has been determined that" removes the actor from a message whose subject is who does what next. It is also the most checkable of these by eye.

One server round trip, followed by a verification pass that is not optional.

Step 1

Paste the body only. The signature block will not survive as formatting and does not benefit from rewriting.

Step 2

Optionally constrain the output through the 300-character instruction field, then run it. One round trip to our AI server.

Step 3

Verify dates, figures, proper nouns and any sentence expressing a commitment, against the original, before the email leaves your drafts.

The structural tell precedes reading. A formal opener, a body of three sentences in a narrow length band, and a closing that requests nothing produce a recognisable outline, and readers classify outlines faster than they read sentences.

Correspondence written under time pressure is asymmetric: a long sentence carrying the substance, a short one carrying the ask, an uneven closing naming a person and a date. This asymmetry is the harder half of the problem, because a writer deleting stock phrases by hand almost always leaves the shape intact.

Perplexity's pipeline is retrieval then synthesis with attribution. Every artefact of that pipeline is misplaced in correspondence: the bracketed numeral refers to a source list that did not travel with the copied text, the attribution construction hedges a claim the sender should simply make, and the currency caveat is a search engine disclaiming its own index.

In outreach the costliest artefact is the research paragraph itself. Summarising a recipient's organisation back to them inverts the information gradient: they know it better than the retrieval did, so the paragraph spends the reader's attention and returns nothing. The rewrite compresses it and moves the request forward.

Competing tools in this category claim that action items, deadlines and requests are preserved. That is a claim about the behaviour of a language model across all inputs, and it is not enforceable by the party making it. We decline to make it.

The accurate statement is narrower and more useful: the rewrite is instructed to preserve meaning and factual content, and you verify four classes of item afterwards, namely dates, quantities, proper nouns, and sentences expressing a commitment. Those four are where a rewording turns into a changed obligation, and they are quick to scan because you already know what the email was supposed to say.

Output is plain text, so formatting, signature blocks, quoted threads and inline attachments are not preserved. The subject line is out of scope unless pasted with the body. There are no tone presets: the only control is one free-text instruction of at most 300 characters, and describing it as a preset menu would misrepresent the interface.

No detector claim is made anywhere on this route. We do not test against GPTZero, Copyleaks, Turnitin or Originality.ai, hold no affiliation with any of them or with Perplexity AI, and make no prediction about what any of them would report.

Outreach produced in the same session as the research, which is where the provenance axis dominates. Sensitive internal messages, where stock phrasing reads as evasion. Second-language correspondence, where the register is formal by caution rather than by intent.

It is not worth a run on a short reply. Where the email is shorter than the verification pass it requires, the rewrite costs more attention than it saves.