How to use AI in a design workflow, step by step

AI is fastest at the parts of design that were never the hard part. A human-directed, AI-built workflow puts the brief, the type and the final call with you, and uses the models for breadth, with the rights and provenance checks that 2025-26 law now expects.

A Jacquard loom in a museum with a long chain of punched cards that program the woven pattern
Photo: Stephencdickson, CC BY-SA 4.0, via Wikimedia Commons

A good AI design workflow is human-directed and AI-built: you write the brief and make the decisions, and the models do the breadth work of research, moodboards and variants. Keep typography, layout and the final artwork in your own hands, use tools whose licences allow commercial use, keep the Content Credentials, and log what was generated. That order matters because research shows image generators make designers fixate on the AI's first outputs, and because the US Copyright Office says prompts alone don't make you the author.

Key takeaways
  • 7 steps, alternating human decisions and machine generation; the human has the first and last word.
  • Sketch before you prompt: in a CHI 2024 study of 60 people, AI users fixated more and varied less.
  • Five AI ideas raised story novelty by 8.1% but made stories more alike (293 writers).
  • Prompts alone don't make you the author, per the US Copyright Office (Jan 2025).
  • India's IT Rules require labels on realistic synthetic content from 20 Feb 2026.
  • Keep Content Credentials and a one-line generation log for every asset.

Most "AI for designers" guides are tool lists. Tools change every quarter; the workflow problems don't. This guide answers four questions in the order you meet them on a real job:

  • Where does AI help without flattening the ideas?
  • Which parts must stay manual?
  • Who owns the result?
  • What must you disclose in India after the February 2026 amendment to the IT Rules?

It is the method I use, as someone self-taught in graphic and brand design since 2013 who now builds marketing systems that are "human-directed, AI-built".

Jan 2025US Copyright Office Part 2 report: prompts alone don't create authorship
60designers in the CHI 2024 fixation study
20 Feb 2026India's synthetic-content labelling rules take effect
2.3current C2PA Content Credentials spec version

What does an AI design workflow look like, step by step?

It runs in seven steps, alternating between human decisions and machine generation: brief, research, divergent sketching, AI variants, human selection and build, rights and provenance checks, and a generation log. The human owns the first and last word in every loop.

HumanAI1 Brief+ criteria2 Researchsummaries3 Sketchby hand4 Variantsand mood5 Select,type, layoutRetouch,upscale6 Rights7 LogEvery AI step starts from a human decision and ends in one.
The zig-zag is the point: sketch before you generate, select before you build, and check before you ship.
  1. Write the brief. Audience, job of the design, constraints (sizes, scripts, print or screen), three success criteria and what is off-limits.
  2. Research with AI, verify by hand. Summaries of competitors, category conventions and references, each checked against the original source.
  3. Sketch before generating. Ten to twenty rough thumbnails on paper, so your ideas exist before the model's do.
  4. Generate variants. Use image models for moodboards, textures, colour studies and alternatives to your sketches.
  5. Select and build by hand. Choose against the brief's criteria; set type, grid and layout yourself; retouch and upscale where needed.
  6. Check rights and provenance. Licence, training-data terms, Content Credentials, labelling duties.
  7. Log what was generated. Tool, model, date, prompt and what you changed, kept with the project files.
The story

On 15 September 2022, Kris Kashtanova registered Zarya of the Dawn, an 18-page comic, with the US Copyright Office. The application did not say that the images came from Midjourney, and when the Office saw Kashtanova's social-media posts saying so, it moved to cancel the registration. Kashtanova's lawyer replied that the work had taken hundreds or thousands of prompts and hundreds of intermediate images, and that Kashtanova had retouched Zarya's lips in Photoshop. In a letter dated 21 February 2023, the Office kept the text and the selection and arrangement of words and pictures, but excluded the Midjourney images. The lip edit, it said, was "too minor and imperceptible" to count. What Kashtanova wrote, chose and arranged was protected; what the model drew was not.

US Copyright Office: Re: Zarya of the Dawn, letter of 21 February 2023

Why write the brief before you open any AI tool?

7 stepsalternating human and machine
1st and lastword always belongs to the human

Because AI images anchor your thinking. In a CHI 2024 experiment with 60 participants, people who used an image generator while designing fixated more on the example they were shown and produced less varied ideas than people with no inspiration tools.

Fixation moves, it doesn't disappear

Design fixation is an old finding: show designers an example and they unconsciously copy its features. The new result is that a generator does not cure it; it moves it.

In that study, fixation shifted from the starting example to the AI's own outputs, and much of it began while participants were writing prompts.

A prompt is already a solution sketch, so prompting first means the model's interpretation of your words becomes your starting point.

How a brief protects you

A written brief protects against this in two ways:

  1. Criteria first. It gives you selection criteria that exist before any image does, so you judge outputs against a goal rather than against each other.
  2. Your sketches. It makes the sketches in step 3 yours.

The top answer in Stack Exchange's thread on copying other designers draws the same line: borrowing one or two elements is normal, but a design that is essentially someone else's with a few changes is not ethical.

Stacks of punched pattern cards for a hand-loom used to weave silk saris in India
Jacquard pattern cards for silk sari weaving. The loom does the repetitive work, but a person designs the pattern before a single card is punched: the oldest version of human-directed, machine-built.Photo: McKay Savage from London, UK, CC BY 2.0, via Wikimedia Commons

Where does AI help most: research, moodboards or variants?

60people in the CHI 2024 fixation study
293writers in the story-novelty study
8.1%novelty gain from five AI ideas
9.0%usefulness gain from five AI ideas

AI helps most where breadth is cheap and judgement is still yours: summarising research, building moodboards, generating colour and texture options, and producing variants of an idea you already have. It helps least on the decisions that define a design: concept, hierarchy, type and the final crop.

The sameness trade-off

There is a trade-off you should know before you lean on it. A 2024 study in Science Advances gave 293 writers zero, one or five AI story ideas.

AI help raised the rated novelty and usefulness of individual stories, but the AI-assisted stories were more similar to each other. Individually better, collectively more alike: for a brand trying to look different from its category, that is the risk in one sentence.

AI ideas raised individual scores (change vs no AI)
See the data as a table
One AI idea (%)Five AI ideas (%)
Novelty5.48.1
Usefulness3.79.0

Source: Doshi and Hauser, Science Advances, 2024 (293 writers, 3,519 evaluations) · The same study found AI-assisted stories became more similar to one another, so the collective diversity of ideas fell.

TaskAI roleHuman roleMain failure to watch
Category and competitor researchSummarise, cluster, list conventionsCheck every claim at the sourceInvented facts and sources
MoodboardsGenerate textures, light, palettesPick against the briefEverything looks like the model's house style
Concept variantsExpand your sketches into optionsChoose and combineFixation on the first good output
Typography and layoutSuggest pairings, check copy lengthSet type, grid and hierarchyUneditable lettering baked into pixels
Photo retouch and extensionGenerative fill, clean-up, upscaleDecide what is honest to changeAltering facts in a documentary image
Production at scaleResize and adapt formatsApprove the master and the rulesErrors multiplied across 50 sizes

A practical rule that follows from both studies: generate wide, then select hard. Ask for many variants, but force yourself to name what each fails on before you pick one. For what this does to the job of a designer more broadly, see what changes when you design with AI.

Why keep typography and layout manual?

Because type and layout are where a design is edited, handed off and reused, and generated pixels can't be.

Lettering inside a generated image can't be kerned, re-set in another language, made accessible or swapped by a developer; a type layer in a layout file can.

There are three practical reasons to keep them in a layout tool:

  • Precision: spacing decisions are measured in units of a thousandth of an em. You can't prompt your way to a consistent 20-unit tracking across a campaign; you set it. The details are in kerning, tracking and leading explained.
  • Systems: a grid makes 40 assets look like one family. A generator invents a new layout each time. Build the grid by hand with the method in the typographic grid system guide, then let tools fill it.
  • Hand-off: the top answer in the Stack Exchange thread on handing designs to developers says to move into the real medium, HTML and CSS, as early as possible. A flattened AI image is even further from that medium than a layered mock-up.
A wooden type case full of metal movable type in the Gutenberg Museum in Mainz
A letterpress type case. Each letter was a separate, reusable object, which is exactly what text inside a generated image is not.Photo: dronepicr, CC BY 2.0, via Wikimedia Commons

Not the purely generated parts, at least in the US. The US Copyright Office's Part 2 report (January 2025) says that providing prompts alone does not make the output protectable, while human-authored material that is perceptible in the output, and human selection, arrangement or modification of AI material, can be protected.

Courts agree on the baseline. In Thaler v. Perlmutter, decided on 18 March 2025, the US Court of Appeals for the D.C. Circuit held that the Copyright Act requires a human author, and refused registration for an image that Stephen Thaler listed as made by his AI system alone.

The court left the harder question (how much human input is enough) to be decided case by case.

For a working designer, this has practical consequences:

  • A logo that is a raw generated image may be unprotectable, so a competitor could use something very close. This is one more reason the logo design process should end in vector work drawn by a person.
  • Your sketches, your type setting and your arrangement are where your rights live, so keep the files that prove them.
  • Indian copyright law on AI outputs has not been settled by a specific ruling that I could verify, so treat the US position as a guide, not as Indian law, and check a lawyer for high-value identities.

Which AI tools are safe for commercial design work?

Prefer tools that publish where their training data came from and what you may do with outputs. Adobe, for example, says its Firefly models are trained on licensed content such as Adobe Stock plus public-domain material, not on customer content, and offers IP indemnity to enterprise customers for Firefly output.

Partner models

Read the fine print on partner models. Adobe's apps now also offer third-party models (Adobe's October 2025 Firefly release lists models from Black Forest Labs, Google, OpenAI, Runway and others).

Adobe says it labels when a partner model is used and that users must decide whether a partner model suits their project. In other words, the commercially safe promise covers Adobe's own models; the partner models carry their own terms.

CheckQuestion to answerWhere to find it
Commercial useDo the terms allow client and paid use of outputs?Tool terms of use
Training dataIs it licensed, public domain or scraped?Vendor's AI approach page
IndemnityDoes the vendor defend you if output infringes?Enterprise contract
Model usedWas it the vendor's model or a partner's?App label, Content Credentials
InputsDid you upload images you have rights to?Your own asset records
Names in promptsDid you prompt with a living artist's name or a brand's marks?Your generation log
Inside the glass-curtain workshop wing of the Bauhaus building in Dessau, with sunlight falling across the floor
The workshop wing of the Bauhaus in Dessau, a school that set out to design for machine production. The machines changed; the idea of a designer setting the rules for them did not.Photo: Anne Stengel, CC BY-SA 4.0, via Wikimedia Commons

What are Content Credentials, and what does India's 2026 labelling rule require?

20 Feb 2026labels required on realistic synthetic content in India
Jan 2025US Copyright Office: prompts alone are not authorship
1 linegeneration log to keep for every asset

Content Credentials are a signed record of how a file was made and edited, built on the open C2PA standard (version 2.3), whose steering committee includes Adobe, Google, Microsoft, Meta, OpenAI, Sony, the BBC and others. India's amended IT Rules, in force from 20 February 2026, require platforms to label realistic synthetic audio and visual content.

Photo capturedcamera signsCrop and coloureditor adds entryGenerative fillAI use recordedExportmanifest attachedManifest: who, which tool, what changed, which modelSigned, so tampering shows; stripped metadata means no proof
Each step adds a signed entry to the file's manifest. Exporting through a tool that strips metadata breaks the chain, which is why the "keep credentials" step comes before publishing.

What India's rule covers

The Indian rule, notified by MeitY in February 2026, covers synthetically generated information: audio, visual or audio-visual content created or altered by a computer so that it looks real and could pass for a real person or event.

CoveredExcluded
Realistic content created or altered by a computer that could pass for a real person or eventRoutine, good-faith editing
Material manipulation of any part of the contentTools that improve quality, clarity or accessibility without manipulating a material part

How labels work

The final rules dropped the draft's 10% label-size proposal, replacing it with a requirement that labels be prominent and easy to notice.

Significant social media platforms must also ask users to declare whether content is synthetic and verify it.

What this means for designers in India:

  • A stylised illustration or an obviously graphic poster is far from "appears real". A photorealistic generated person in an ad is not, so label it.
  • Colour correction and clean-up are routine editing. Adding a person, a product or an event that was not there is manipulation.
  • Keep the Content Credentials when you export: they are the easiest way to prove what was and wasn't generated if a platform asks.

This sits inside the wider India AI Governance Guidelines that MeitY released on 5 November 2025, which lean on existing laws and voluntary measures rather than a new AI Act.

How do you record what was generated?

Keep a one-line log entry for every generated asset that survives into a deliverable: date, tool and model, prompt, input files, and what you changed by hand. It takes seconds and answers every later question about rights, labelling and authorship.

FieldExample entry
Assethero-banner-v3.psd, background layer
Tool and modelPhotoshop Generative Fill, Firefly model
Prompt"monsoon clouds over a ghat road, dusk"
InputsOwn photo, DSC_0412
Human workComposite, colour grade, all type and layout
LabellingNot photoreal person or event; credentials kept

Use this to check what you just read. Score a deliverable before it goes out: set each check to 1 if done, 0 if not.

AI design hand-off check

If you are building a portfolio, the log doubles as process evidence: it shows a reviewer exactly where your judgement came in, which is the point of a portfolio built without clients.

For the same discipline applied to ad production, see the Marketblog's weekly AI ad creative pipeline, and my wider work on these systems is on the Applied AI page.

What does the research say?

The Effects of Generative AI on Design Fixation and Divergent ThinkingWadinambiarachchi S, Kelly RM, Pareek S, Zhou Q, Velloso E · CHI 2024 (ACM)

Sixty participants, 20 per condition, sketched as many chatbot avatar ideas as they could in 20 minutes after seeing an example avatar, with no help, with Google Image search, or with Midjourney V4. Bayesian models showed the AI group fixated more on the example (+0.32 on a probit scale versus baseline) and produced less varied ideas; originality differences were negligible. Fixation also shifted onto the AI's own outputs and began during prompt writing. The limit: a single 20-minute task, a small sample and an older model, so the size of the effect may differ for experienced designers on real briefs.

Read the paper (free)
Generative AI enhances individual creativity but reduces the collective diversity of novel contentDoshi AR, Hauser OP · Science Advances · 2024

In an online experiment, 293 writers wrote short stories with no AI, one GPT-4 idea or five GPT-4 ideas, and 600 evaluators made 3,519 ratings. Five AI ideas raised rated novelty by 8.1% and usefulness by 9.0%, but AI-assisted stories were more similar to each other and to the AI's suggestions. The limit: eight-sentence stories, not visual design, so read it as evidence about idea generation in general.

Read the paper (free)
Copyright and Artificial Intelligence, Part 2: CopyrightabilityUS Copyright Office · Official report · January 2025

The Office reviewed public comments and existing law and concluded that current principles are enough: material generated by AI from prompts alone is not protected, while human expressive contributions visible in the output and human selection, arrangement or modification can be. It recommended no new protection for AI outputs. The limit: it describes US law and Office practice; other countries, including India, may decide differently.

Read the report (free)

Questions people ask

Will AI replace graphic designers?

It is replacing tasks, mainly production and early exploration, faster than it replaces judgement. The research above shows AI raises individual output but narrows the range of ideas, which makes the person who sets the brief and picks the direction more valuable, not less.

Is it ethical to use AI images in client work?

It is when the client knows, the tool's licence allows it, you have not prompted to imitate a living artist or another brand, and realistic synthetic content is labelled where the law requires. Put AI use in your proposal so it is never a surprise.

Can I trademark a logo made with AI?

Trademark protects a mark's use in trade, not authorship, so registration questions differ from copyright. But a purely generated logo may lack copyright protection in the US, which weakens your position against copies. Redraw the final mark as vector artwork by hand.

Do I have to label every image I edited with AI in India?

No. The 2026 rules target content that looks real and could pass for a real person or event, and exclude routine good-faith editing that does not manipulate a material part of the content. Clean-up and colour work are routine; adding people or events that did not exist is not.

Sources

  1. Wadinambiarachchi et al.: The Effects of Generative AI on Design Fixation and Divergent Thinking, CHI 2024 (60 participants, fixation and variety results)
  2. Doshi and Hauser: Generative AI enhances individual creativity but reduces the collective diversity of novel content, Science Advances (2024; novelty and usefulness figures)
  3. US Copyright Office: Copyright and Artificial Intelligence report, Parts 1 to 3 (2024–2025)
  4. US Copyright Office NewsNet: Part 2 report on copyrightability released (Jan 2025; prompts alone not protectable)
  5. IPWatchdog: D.C. Circuit upholds human authorship requirement in Thaler v. Perlmutter (Mar 2025)
  6. Adobe: Our approach to generative AI with Firefly (training data, indemnity, partner models)
  7. Adobe newsroom: Firefly at Adobe MAX 2025 (Oct 2025)
  8. C2PA: Coalition for Content Provenance and Authenticity (spec 2.3, steering committee)
  9. Khaitan & Co: MeitY notifies the IT Amendment Rules 2026 (SGI definition, exclusions, labelling, dates)
  10. S.S. Rana & Co: MeitY unveils India AI Governance Guidelines (Nov 2025)
  11. Figma: Config 2025 press release (May 2025)
  12. US Copyright Office: Re: Zarya of the Dawn (Registration # VAu001480196) (21 Feb 2023) (story: 15 Sep 2022 registration, 18 pages, "hundreds or thousands" of prompts, lip edit too minor, images excluded)
  13. Graphic Design Stack Exchange: Is it ethical to copy the work of other designers?
  14. Graphic Design Stack Exchange: Efficient workflows for handing off designs to developers
ai design workflowgenerative AIdesign fixationAdobe FireflyContent CredentialsC2PAcopyrightIT Rules 2026
Sapta

Sapta (Saptarshi Sarkar) is a mechanical engineer turned marketer in Bengaluru and a solo touring rider. He writes the seven Saptaverse blogs. Work with him or read more of Artblog.

Keep reading in Artblog

From the other blogs