Where AI fits in a creative workflow, stage by stage

What does a creative workflow look like before AI enters it?

Before deciding where AI goes, write down the workflow you already have. Ours is the one named on our services page: six stages that move a project from insight to growth. Many studios and in-house teams run some version of it, even if they call the stages something else.

1. Discovery and strategy: research into the market, the audience and competitors, ending in a brief and a positioning everyone agrees on.
2. Creative direction: the idea, the visual system and the campaign narrative.
3. Production: filming, photography, motion graphics, 3D and the edit.
4. Technology and build: websites, apps, e-commerce and the tools behind them.
5. Marketing and growth: campaigns, social media, search and analytics.
6. Optimise and scale: testing, measuring and improving what is already live.

AI does not add a seventh stage. It changes the cost of particular tasks inside each of the six. So the useful question is never 'should we use AI?' but 'which tasks in this stage are drafts, which are mechanical, and which are decisions?'

Draft, speed up, decide

• Draft: AI produces a first version that a person rewrites, combines or throws away. Nothing it drafts is published as it stands.
• Speed up: AI does a mechanical job that a person can check quickly, such as transcribing an interview or resizing a set of exports.
• Decide: a named person makes the call and answers for it. AI may inform the decision; it never owns it.

One test sorts most tasks: the more a mistake would cost, and the harder it is to spot, the further that task should sit from unreviewed AI output. A clumsy caption draft costs a minute to fix. A wrong figure on a product page, or a generated face that looks like a real customer, can cost a client's trust.

The rest of this guide maps AI in creative workflows stage by stage, and marks at each one what AI can draft, what it can speed up and what stays with people. It is how I think about planning AI into our projects at The House of Edits, where my work sits between creative direction and emerging technology.

Discovery and strategy: can AI help with research and briefs?

Yes, as long as you treat what it produces as reading notes, not findings. Discovery is where a team gathers far more material than anyone can read closely, which is exactly where a model helps. It is also where a model's confident mistakes do the most damage, because everything later is built on them.

What AI can draft

• Summaries of long material: the client's brochure, past campaign decks, annual reports and research PDFs.
• A first list of competitor claims gathered from public pages, for a person to verify one by one.
• Kick-off question lists tailored to the sector, so the first meeting covers what matters.
• A skeleton brief with the headings filled in from your notes, ready for a strategist to rewrite.

What AI can speed up

• Transcribing stakeholder and customer interviews, and grouping the quotes by theme.
• Turning scattered meeting notes into one organised document.
• A first-pass translation of Kannada or Hindi material for internal reading. Anything that will be published still needs a fluent reviewer.

What stays human

The positioning, the insight worth building on and every statement of fact. Language models are built to produce plausible text, and plausible is not the same as true: a model can attribute a figure to a report that does not contain it. Every number that reaches a deck should be traced to a source a person has opened and read.

The most useful output of discovery, with or without AI, is a facts file: one document that lists what the brand may state, with the source of each figure. On the Sumadhura logistics website, we started from the client's written brief and transcribed the printed park brochure into one source of facts. Where figures disagreed, the park pages follow the brochure, and figures that had not been published, such as clear height, were left out. A file like that is what makes AI drafting safe later, because every draft can be checked against it line by line.

Creative direction: why the idea and the taste stay human

Creative direction is choosing one idea and saying no to everything that does not serve it. That is the part of the job I would not hand to a model. My method is to decide how a piece should look and feel first, then make every production choice serve that decision. A model can offer a hundred directions in a minute, but it cannot tell you which one is true to this brand, this audience and this moment.

Where AI helps a director

• Rough mood references to show a route quickly, before anyone spends a day on it.
• Variations on a layout, palette or composition, to test against the idea that has already been chosen.
• Raw lists of names, lines and taglines, to be mined for a word or a rhythm rather than used as they are.
• Quick mock-ups of a route on real applications, such as a poster, a shopfront or a social post.

Why taste does not transfer

Generative models lean towards the most likely answer. Direction is the opposite job: finding the specific choice that belongs to one brand and no other. When a team picks from whatever the tool offers instead of steering it towards a decided idea, the work starts to look like the work of everyone else using the same tool.

The order matters. On the VicWide website, the brief already set the idea: a scroll story built around one home, which became one illustrated house that explains every inspection. Only then did we generate imagery, all of it centred on one consistent reference house, so every film, still and 3D view shows the same property. On Sumadhura, the idea was that the visitor arrives at the park the way a truck does, and the generated film follows one truck from the highway to the dock to carry that line.

Signs that direction has slipped to the tool

• Nobody can state the idea in one sentence, but there are forty generated options in the folder.
• Feedback rounds are spent choosing between outputs rather than refining one route.
• The visuals look polished but could belong to any brand in the category.
• The brief is being rewritten to fit what the tool produced.

Production: where generation, previsualisation and editing assists earn their place

Production is where AI tools are most visible, and where the line between help and harm is easiest to draw. I sort production uses into three groups, each with its own rules.

Generation: imagery a camera cannot capture

Generated stills and video earn their place when they show something that does not exist yet, or could not reasonably be filmed, and when they are presented honestly as illustration. Two of our website builds used generation this way.

• For Sumadhura, we generated six native 4K chapters based on the official Hoskote renders. One Indian fleet truck with a rust-orange 40 ft container runs from the highway through the gate to the dock and ends on an aerial view. The clips were then cut into 180 frames that play as the visitor scrolls. Basing the generation on the developer's own renders ties the film to the client's material rather than to an invented park.
• For VicWide, the stills and short films (blueprint to finished build, dusk, pool and interior loops) were generated with Higgsfield, and each is labelled as an illustration, never as an actual property or real inspection work.

Generation does not earn its place when it stands in for the real thing a customer will buy or meet: your product, your premises, your team, your customers or your past work. In those cases the camera is the point.

Previsualisation: deciding before the shoot

Rough generated frames, storyboards and animatics let a client and a crew agree framing, pacing and locations before the shoot day, while changes are still cheap. Treat previs as a planning document, not a deliverable. The shoot still records the real people, products and places, and the previs simply means everyone arrives knowing what the film needs.

Editing assists: the mechanical half of post

• Transcripts of interviews, so the editor can find selects by searching instead of scrubbing through hours of footage.
• First-draft subtitles, checked by a person for names, brand terms and Kannada or English spellings.
• Audio clean-up for room noise and hum.
• Masks and rotoscoping that once took hours of frame-by-frame work.
• Upscaling and noise reduction on older or low-light footage.
• Draft vertical reframes of landscape edits, which an editor then corrects shot by shot.

What stays with people is the cut itself: story structure, rhythm, the choice of performance and the final look. Our video production process moves from a rough assembly to picture lock before colour, sound and graphics are finished, and AI assists do not change that order. They make each stage faster; they do not remove a review. For what happens at each step, see how post-production works, from ingest to final delivery.

Technology and build: chatbots, automation and content systems

Our AI Laboratory work, listed as AI workflows, content systems, chatbots and process automation, sits mostly in this stage. Here AI becomes part of what a client's customers use, so the standard is higher: it has to be correct when nobody from the team is watching.

Chatbots that answer only from approved content

A website assistant is useful when it answers the questions a visitor would otherwise phone about, and risky when it improvises. Two of our website builds show what grounded in approved content means in practice.

• Ask VicWide says plainly that it is an AI and answers only from an approved answers file. Without an API key, it falls back to a local engine.
• The chat on the Uniq Builders website answers from the site's own packages, services and FAQs, offers quick-reply chips and hands off to a call, WhatsApp or email for anything else.

Before any chatbot goes live, settle these points in writing:

• Source: the exact documents it may answer from, and who keeps them current.
• Refusals: the topics it must decline, such as prices that are not on the site, legal questions or promises of outcomes.
• Hand-off: a clear route to a person by phone, WhatsApp or email.
• Disclosure: it says it is an AI, in its first message.
• Review: someone reads a sample of conversations regularly and updates the approved answers.

Content systems: volume with a fact check built in

A content system is a facts file, a set of templates, a review step and a publishing schedule working together. On VicWide, the first 97 guides were written against a single facts file that lists the only VicWide details a guide may state, and each guide goes live on its own date without a redeploy. On Uniq Builders, 480 guides across eight topic hubs were all written against a single facts sheet. Whether a paragraph is drafted by a person or a model, the facts file is what keeps volume honest: the reviewer checks every claim against one document instead of their memory.

Automation that does not need AI

Some of the most useful automation is plain code. The VicWide build has a launch guard that refuses to build while placeholder business details remain, and its enquiry forms return a unique reference that can be sent on to WhatsApp. Before reaching for a model, ask whether a rule would do: rules are predictable, cheap to run and easy to audit.

Good candidates for automation, with or without AI:

• Naming and filing exports to an agreed convention.
• Resizing and encoding images and video for each placement.
• Routing enquiries to the right person with the details attached.
• Checking pages for missing alt text, missing titles or leftover placeholder copy before launch.

Marketing, optimisation and scale: what AI can measure and repeat

Once work is live, AI is good at two things: producing variations to test and summarising what the numbers say. It is poor at knowing which number matters, and why.

Marketing and growth

• Caption and ad copy variations around one approved message, for testing rather than for publishing unread.
• Draft alt text, meta descriptions and page titles, edited against the page they describe.
• Cut-downs and resizes of approved creative for each placement.
• A first draft of a monthly content calendar built from the brand's themes and key dates.

Optimise and scale

• Plain-language summaries of campaign and analytics reports.
• Patterns worth testing next, framed as hypotheses for a person to choose from.
• Repeating a format that has proved itself, at the volume a channel needs, from templates the team has approved.

What stays human

Choosing the metric that reflects the business goal, deciding whether a result is a real signal or noise, killing or scaling a campaign, keeping the brand's voice, and replying to people. A brand that automates its replies on social media risks no longer having a conversation at all.

One caution: AI can produce more variants than anyone can review. Test only what you would be happy to see published, because in a live test it is published. For how formats are shifting as these tools mature, see the future of video production for brands.

Where AI should stop: people, proof, rights and final sign off

These four lines hold whichever tool you use and however good it becomes.

People

Do not generate or alter a real person's face, body or voice without their informed, written consent, and never make a real person appear to say or do something they did not. Platforms set rules here too: YouTube requires creators to disclose when they use AI to meaningfully alter or generate photorealistic content. Check the current rules of every platform you publish on before a campaign goes out.

Proof

Anything a customer reads as evidence must be real: testimonials, reviews, case-study images, product shots, results, team photos and client logos. The VicWide brief set this as a rule from the start: generated imagery is always shown as illustration, never as client work. On Uniq Builders, imagery that is not the firm's own work is captioned as illustrative. If a generated image could be mistaken for proof, label it or leave it out.

Rights

• Read each tool's terms for commercial use before its output goes into client work.
• Check whether a tool may train on what you upload, and keep confidential client material out of any tool that allows it.
• Ownership of generated material is a legal question, and the answer can depend on the country. For a decision that matters, take advice from a qualified intellectual property lawyer.
• Keep a record of which tool, prompt and source material produced each asset. On VicWide, the generated imagery is tracked with a manifest and an ingest script in the project repository.

Final sign off

Every published piece needs a named person who has read or watched all of it and approves it. Not 'the team' and not 'the tool': a person. Their check covers facts against the facts file, names and spellings, labels on generated material, rights and the brand's voice.

How do you add AI to your own workflow without breaking it?

Introduce AI the way you would bring in a new team member: one task at a time, with a review, before you trust it with more.

1. Map what you do now. Write the stages, the tasks inside each and who signs off, on one page.
2. Sort every task. Mark it draft, speed up or decide. Leave every decide task with a person.
3. Write the facts file first. Before any AI writes copy, list what the brand may state and where each fact comes from.
4. Pick one bottleneck. Choose a task that eats time and is easy to check, such as interview transcripts or first-draft subtitles.
5. Run it on a real, low-risk project. Compare quality, time taken and the number of corrections against your usual way.
6. Name the reviewer. Write down who checks the output and what they check it for.
7. Set the labelling rules. Decide how generated images are captioned and how an assistant introduces itself, before anything is published.
8. Review after a month. Keep what held up, drop what created more checking than it saved, then pick the next task.

Mistakes to avoid

• Buying tools before mapping the workflow they are meant to fit.
• Letting generated drafts skip review because they look finished.
• Using generation where real footage or photography is the point, such as products, premises and people.
• Publishing figures that nobody traced to a source.
• Measuring success by how much was produced instead of how good the published work is.

If you want help mapping AI into your own workflow, from grounded assistants to content systems, tell us what you are working on. More writing on the subject is collected under AI.