An AI-generated logo arrives as a 1024-pixel PNG that looks finished. It is not a logo yet. It is a picture of a logo. Turning it into something that works on a favicon, a shirt, an invoice and a dark screen is a specific craft with a specific sequence. This is what a brand designer does, in order, and what "done" means at the end.
Why the PNG is not the logo
A logo is a system, not an image. It has to survive being 16 pixels wide, being printed in one colour, being placed on a photograph, being embroidered, and being sent to a client who will put it in a slide at the wrong size. A generated concept has none of that yet. It also has, almost always, three problems that a trained eye sees immediately and a founder does not: inconsistent stroke weights, letterforms that are nearly but not actually a real typeface, and detail that will vanish or turn to mud when small.
There is also a legal point worth knowing. The US Copyright Office's standing position, set out in its Copyrightability report and unchanged as of 2026, is that AI-generated material is protected by copyright only where a human has determined sufficient expressive elements. Prompts alone do not count; human selection, arrangement and modification can. A brand designer's finishing work is what gives the mark a human author. That matters when you want to own your own logo.
The sequence
1. Choose a direction, and say why
AI produces candidates cheaply, so requesters usually arrive with four to ten. The first task is to pick one, and the reason matters more than the pick. A designer evaluates each against the brand's actual context (who sees it, where, next to what) rather than against how nice it looks at full size. The reason gets written down; it becomes the first line of the brand rules.
2. Redraw, do not trace
Auto-tracing a PNG produces a vector with hundreds of anchor points, wobbly curves and no geometry. A designer redraws the mark from scratch, using the concept as reference: real circles, real angles, consistent radii, a grid. This is where the mark becomes parametric, which means it can be adjusted later without starting over. It is also where most of the AI artefacts disappear: the not-quite-symmetry, the stroke that thins for no reason, the counter that is closing up.
3. Fix optical weight and spacing
Mathematically equal is not visually equal. Horizontal strokes need to be thinner than vertical ones to look the same. Round shapes need to overshoot the baseline. Letters need different gaps depending on their neighbours. AI output does not know any of this. The designer adjusts by eye, at several sizes, until the mark reads as one calm object.
4. Test at small sizes, and cut what fails
Render at 16, 24, 32 and 48 pixels. Print at 10 mm. Anything that becomes noise gets simplified or removed. Often this produces a second version of the mark, a small-size variant with thicker strokes and less detail. A logo that only works large is half a logo.
5. Build the colour system, with contrast
The generated palette is a mood; it is not a system. The designer picks one primary and defines exactly how it behaves: on white, on dark, as text, as a fill, as an accent. Every combination that will carry text is checked for WCAG 2.2 AA contrast (4.5:1 for normal text, 3:1 for large text and graphics). A dark-mode version is designed, not just inverted, because most brand colours need adjusting to hold on black.
6. Typography, if there is a wordmark
AI-generated letterforms are almost never a real typeface, which means they cannot be typed, extended or matched. The designer either identifies a licensed typeface that carries the intent and sets the wordmark properly, or draws custom letters as vectors with consistent construction. Either way, the result can be reproduced exactly and paired with body text that agrees with it.
7. Variations
A usable identity ships as a set: primary lockup, stacked lockup, mark alone, wordmark alone; each in full colour, single colour black, single colour white (reversed), and dark-mode; plus favicon and app icons, which usually means the mark on a shaped background at fixed sizes. Each variation is drawn and checked, not exported by changing a fill.
8. Clear space, minimum size, misuse
Three rules prevent most future damage: how much space must surround the mark, the smallest size it may be used at, and a short list of things not to do (stretch, recolour, put on busy photos, add effects). One page. Illustrated.
9. Files and hand-off
SVG for everything that can be vector. PNG at three sizes with transparent backgrounds. A favicon set (ICO, SVG, PNG sizes) and app icons. Fonts or font licence details. A short PDF of rules. Named consistently. A designer who has done this before delivers a folder a developer can use without asking a question.
A definition of done
What to give the designer
The generated candidates as files, the prompt if you have it, a sentence on where the logo will mostly live, anything that must not change (the name, an existing colour), and what went wrong with previous attempts. Then let them propose a tighter definition of done if they see something you missed. Good designers will.
How this was made: written by the After AI Work team, drawing on the process used to build our own identity from a generated concept. Copyright position from the US Copyright Office, Copyright and Artificial Intelligence Part 2 (Copyrightability) and its registration guidance, both still current in 2026. Contrast thresholds from WCAG 2.2, the version adopted by the 2026 EN 301 549 standard. AI tools helped outline the sequence; every step was rewritten and checked by a person.