Building an AI content engine that doesn't sound like a robot
AI will happily produce a month of content in an afternoon. Most of it will be unusable in a way that's hard to articulate — technically fine, saying nothing. The trick is knowing precisely which parts of the job to hand over.
What it's genuinely good at
AI is excellent at the work that surrounds writing rather than the writing itself. Turning a rambling voice note into a structured outline. Producing twelve headline variants so you can pick one. Reformatting an article into a newsletter, six captions and a script.
It's also very good at the unglamorous jobs nobody wants: alt text, meta descriptions, transcript cleanup, first-pass summaries of long documents. This is where the hours actually go, and where automation pays for itself without anyone noticing a drop in quality.
Where it consistently fails
It cannot tell you what's worth saying. It has no view on which of your clients' problems matters most this quarter, no opinion formed by having been wrong before, and no access to the specific thing you learned on a call last Tuesday.
Ask it for strategy and you get the average of everything ever written on the topic. That average is, by definition, unremarkable — and your audience has read it already, many times, in slightly different words.
It's also confidently wrong about specifics. Statistics, dates, platform features, what a tool does. Anything checkable needs checking.
The line we draw
The rule that's held up for us: AI never decides what to say or makes the final call on how it reads. It handles everything in between.
A human decides the angle, supplies the specifics, and edits the output. AI expands, restructures, reformats and repurposes. Framed that way it stops being a writer you're disappointed in and starts being a very fast assistant.
The workflow
In practice, a single piece of content moves through something like this. The human steps are short; the AI steps are the ones that used to eat afternoons.
- Human: pick the angle and the one thing the reader should take away
- Human: record a rough voice note — messy is fine, specifics matter
- AI: transcribe and structure it into an outline
- Human: fix the outline, cut what's obvious, add the real example
- AI: draft from the corrected outline
- Human: edit hard, especially the opening and the ending
- AI: repurpose the finished piece into captions, newsletter and script
The tells to edit out
Unedited AI writing has a recognisable texture, and audiences are getting quicker at spotting it. A few patterns to hunt down before publishing:
- Openings that restate the title in different words
- Sentences built on 'not only… but also' and 'it's important to note'
- Three-item lists where the third item is filler for rhythm
- Endings that summarise rather than land a point
- Hedging everywhere — 'can help to', 'may potentially assist'
- Zero specifics: no numbers, no names, no example that could only be yours
Quality at volume is a process problem
Teams that publish good work quickly aren't using better prompts. They've worked out which decisions are theirs and refused to delegate those, while automating everything around them.
The measure worth watching isn't how much you published. It's whether anyone would notice if you stopped — and whether a reader could tell your content from a competitor's with the logo removed. AI makes hitting the first number trivial and the second one harder. Spend the time you save on the part that only you can do.