AI for a Design Studio: Where It Fits in Concept, Review and Delivery

A small design studio should use AI most in concept exploration, admin and review preparation, sparingly in production, and not at all for final marks, signed-off artwork or anything where ownership must be certain. Put a human approval gate at the end of every stage and log where each asset came from.

The stage a studio picks first matters more than the tool. Concept work is private, so an odd result costs nothing; delivery work is contractual, so an odd result costs a relationship. Mapping AI to the stage where it sits, and writing down who approves what, is what keeps a studio from drifting into ad hoc use that no one can explain to a client.

Follow me on Instagram@sagnikteaches

The map: three stages, three questions

For each stage ask what AI may do, who approves, and what you record. The table is the summary; the sections after it give the detail.

Connect on LinkedInSagnik Bhattacharya

Illustration throughout: a five-person studio that does brand identity, packaging and web design for small clients, with a typical project of six weeks.

Subscribe on YouTube@sagnikteaches

Stage one: concept

Where AI is genuinely useful

The first use is the unglamorous one: turning a messy brief into a clear one. Paste in the client's emails and call notes (checking your terms first), and ask for a summary of goals, audiences, constraints, mandatory elements and open questions. The open-questions list is the valuable part. Say the output of a brief for a bakery rebrand includes "no mention of budget for printed packaging", which the lead then raises in the next call instead of discovering it in week four.

Below are notes from three client calls and an email thread. Produce:
1) A five-line summary of goals and audience.
2) A list of constraints the client stated (budget, timing, brand rules, things they dislike).
3) A list of questions the notes do not answer that a designer would need answered.
Do not suggest design solutions. Quote the notes where you rely on them.

The second use is volume of variation for internal discussion: twenty directions for a naming list, ten layout ideas for a poster described in words, a dozen colour mood descriptions. The point is to beat the blank page and to show the team ten things not to do. Treat the output as raw material that a designer reworks. Nothing generated should go straight to a client as a proposal, because its provenance is uncertain and the client will pay for your judgement, not for a mood.

Image generation for moodboards

Image tools such as Midjourney, Adobe Firefly or Canva's AI features can produce reference imagery fast. Pick based on terms, not novelty. Adobe's Firefly pages say outputs from its own models can be used commercially, add that partner-model content is the user's responsibility to assess, and offer IP indemnification only to eligible enterprise customers. That matters for a small studio: if you cannot get indemnity, keep generated images as internal reference only, and rebuild anything that ships using licensed or original assets. For the practical tools by cost, see AI graphic design software for a small studio.

Two rules for concept images. First, never prompt with a living artist's name or a competitor's brand. It invites derivative output and an awkward conversation. Describe qualities instead: "flat colour, thick outlines, muted autumn palette". Second, record the prompt and the tool next to the image from the moment you save it. You will not remember in week five which of 40 files was generated.

The concept gate: label every item

A lead designer reviews every item that will be shown to the client and marks each as original, licensed, or AI-reference. AI-reference items are labelled as such on the slide ("direction only, to be redrawn"). Nothing moves to the next stage without that mark.

Stage two: review

Review is where internal reviewers and clients check work against the brief and each other. AI is useful here as a prep tool, not as a judge.

Feedback consolidation. A client sends four people's comments by email, a screenshot with arrows and a voice note. Paste the written comments and ask the model to produce a table: comment, who said it, which deliverable, whether it conflicts with another comment, and whether it appears to change the brief. Conflicts are the key column. When the marketing lead says "make the logo bigger" and the founder says "more white space", the designer takes the conflict back to the client in one message instead of silently choosing.

Scope checking. Ask the model to compare each comment against your signed scope and list those that look outside it. The answer is a prompt for the project manager, not a verdict. Whether to absorb a change or bill it is a commercial decision.

Consistency checks. Give it the copy from a ten-page brand guideline and ask for inconsistencies: "Colour" and "Palette", "logo lock-up" and "logotype", a hex code written two ways. It is quick and catches what tired eyes miss. Run the checker's findings past a human before changing anything, because it also flags deliberate choices.

Accessibility drafts. Ask for a first draft of alt text for the images in a web handover, then edit for accuracy and purpose. The model will describe what it sees but not why the image is there.

What to avoid in review

Do not let a model approve work ("does this meet the brief? yes or no"). It will say yes with confidence. Do not auto-reply to client feedback. Replies carry tone and commitment, and a clumsy sentence about scope creates a dispute. Draft replies with AI if you like, but a person edits and sends. For the client-side process in more depth, see AI in product design client feedback rounds.

The review gate: a signed change list

The project lead signs off a written change list before designers act on it. The list names which items are included in the quote and which are extra.

Stage three: delivery

Delivery is the stage with the most contractual risk and the best low-risk uses.

Handover documentation. Give the model the final file list and your notes and ask for a plain-English handover: what each file is, which format to use where, fonts and licences, and what to do when something breaks. Edit it, and attach it. Clients rarely get this right on their own.

File naming and checks. Ask for a naming convention applied to a list of 60 files ("project_asset_variant_size_v03"). A spreadsheet formula can do this too; AI is simply faster when conventions are messy.

Resizing and variants. Tools that extend a background or generate format variants save time on social assets. They also hallucinate edges and add artefacts. Inspect each at 100% zoom before export, and log any asset that used generative fill.

Final marks and signature elements. Avoid AI for logos and key identity elements. The reasons are practical: many tools give no assurance of originality, trademark clearance is harder when the origin is a model, and you cannot hand a client a mark and say you know where it came from. Use AI to explore, then draw the mark by hand or in vector.

The delivery gate: a second check of the log

Before files leave the studio, a second person checks the provenance log against the delivered files. No log entry, no delivery.

The asset provenance log

This is a spreadsheet, one row per asset that will reach a client. It takes two minutes per asset to keep and saves hours when a client or a rights-holder asks a question. Columns:

• Asset ID and filename
• Source type: original, stock (with licence reference), client-supplied, AI-generated, AI-edited
• Tool and date for any AI use, plus the plan or licence it was used under
• Prompt or instruction or a link to where it is saved
• Human work added: a one-line description of what a designer changed
• Approved by and date

A filled row (illustrative): "PKG-014 bakery box back panel; AI-edited; Firefly generative expand, 14 Oct; prompt 'extend wheat pattern to the right, same colours'; designer redrew the border and corrected the pattern repeat; approved by lead, 15 Oct." Nothing fancy, and it answers nearly any later question.

A worked example: one bakery rebrand, six weeks

A bakery with two shops asks the studio for a new logo, packaging and a five-page website. Here is how the stages play out (illustrative).

Week 1, concept. The project manager pastes call notes into the brief prompt and gets a summary plus nine open questions, three of which the client answers by return. The designers use a chat assistant to generate 40 possible names for a seasonal range and 20 descriptive phrases for the colour mood. They create two moodboards using generated reference images, each marked "direction only". Time: about 4 hours against a usual 8.

Weeks 2 to 3, concept to review. Two designers draw three logo routes by hand. No AI is used for the marks. A model helps write the rationale text for each route from the designer's bullet points, which the designer then edits. The client gives feedback; the PM pastes it into the consolidation prompt and spots that the owner and her business partner disagree about whether the old wheat symbol should stay. One email resolves it.

Weeks 4 to 5, production. Packaging artwork is built from the approved mark. The team uses AI to extend a background pattern across a larger box panel, logs it, and a designer retouches the seam. Alt text and a first draft of web copy are generated and then rewritten by a copywriter. The website's accessibility pass includes a model-assisted check of heading order and contrast descriptions, confirmed with a proper tool.

Week 6, delivery. The handover pack is drafted from the file list. A second designer checks the provenance log against the exported files and finds one stock image with no licence reference, which is replaced. The client receives originals plus a one-page note on what AI was used for.

The total saving in this illustration is perhaps two to three working days across the project, none of it in the places where the client is paying for judgement. That distribution is the point: use AI where the work is preparation and admin, and keep people on the decisions.

Costs: what a studio actually pays

Keep the toolset small. A typical stack for the illustration: one business-plan chat assistant for the people who need it (ChatGPT Business and Claude Team both list Standard seats at $25 a month or $20 billed annually, with a two-seat minimum; Claude Pro for one person is $20 a month), one image tool, and whatever your design suite already includes. These are the list prices on the vendors' pages in November 2026:

• Adobe Firefly: Standard at $9.99 a month with 2,000 generative credits, Pro at $19.99 with 4,000. All paid plans include unlimited standard image generations; the credits go on premium features such as video, audio and partner models.
• Midjourney: Basic $10, Standard $30, Pro $60 and Mega $120 a month, with 20% off for annual billing. Its plan comparison says Stealth Mode, which keeps images private, is only on Pro and Mega, and that a company with more than $1,000,000 in gross annual revenue must buy Pro or Mega. For unreleased client work, budget for Pro.
• Canva: Pro at $144 a year for one person (only one person may use a Pro account) and Business at $250 a year per person. AI features draw on an allowance that rises with the plan: Pro has ten times Free's allowance for Premium and Ultra AI tools, Business twenty times.

A quick sum for the five-person illustration: three chat seats at $25 ($75), one Midjourney Pro seat for private generation ($60), Firefly Standard ($9.99) and Canva Business for two designers (about $42 a month at the annual rate) comes to roughly $187 a month. Give all five people every tool and it passes $300. Audit usage after 90 days and cancel what nobody opened.

Putting it in the contract and the studio policy

Add a short clause to your terms covering: that AI tools may be used in concept and production; that a designer reviews and approves every deliverable; that you will tell the client where generated content appears in final artwork; and that final logos and identity marks are created without generative AI unless agreed in writing. See how to mention AI use in client contracts and proposals for wording options. A one-page internal policy is enough for a small team: approved tools, banned uses (living artists' names, confidential files in free accounts), the three stage gates, and the provenance log.

A first-month rollout for a five-person studio

Do not switch everything on at once. A month is enough to put the gates in place and learn which uses stick.

1. Week 1: agree the one-page policy and the provenance log template. Choose one admin use (the brief summary prompt) and one concept use (name lists). Everyone tries both on a real, current project.
2. Week 2: add feedback consolidation to review on one live project. The project manager keeps a note of every time the table helped or misled.
3. Week 3: try one delivery use, the handover document, and run the first provenance check with a second person. Fix the template where it was awkward.
4. Week 4: hold a 30-minute review: what saved time, what produced rework, what should be banned. Update the policy and drop any tool nobody used.

A realistic result is two or three uses that stick. That is a good outcome; a studio with three well-run AI uses is ahead of one with ten unmanaged ones.

Another example: a one-week website project

A small web design studio takes a five-page site for an accountancy practice in one week. Where does AI fit when time is this tight? Concept: a chat assistant turns the client questionnaire into a sitemap draft and a list of page goals, which the designer corrects in 20 minutes. Production: the designer uses placeholder copy generated for layout only, labelled "lorem replacement, not for publishing". Review: the client's comments are consolidated into a table with the conflicts highlighted. Delivery: alt text drafts, a handover note and a list of editable areas for the client's admin person. The studio does not use generated imagery in the final site because the accountancy client's brief asks for real team photos. The log records two AI-assisted items: the sitemap draft and the alt text, with the designer's edits noted. The whole log took ten minutes.

Handling a client who dislikes AI

Some clients will say no. Respect it and record it in the project file: "AI not to be used on this account", and tell the team on day one. Concept admin that never leaves your own notes is still arguably in scope, but ask rather than assume. Another client may want AI-generated imagery for speed and a low budget. Then the contract should say that you provide generated assets without guarantees of originality, that the client accepts that risk, and that final marks are excluded. Keeping each client's preference in the project brief means designers do not have to remember it.

Where studios go wrong

• Using free personal accounts for client work. Confidential briefs end up under a personal account's settings and history. Move to a team plan.
• Showing generated work as if it were designed. Clients feel misled when they find out. Label it.
• No log. The one time you need it, nobody remembers which tool made the file.
• Chasing quantity. Fifty AI options are not better than five considered ones. The client pays for selection.
• Skipping the gate when busy. Gates are what protect you in a hurry. If a gate keeps being skipped, make it shorter, not optional.

If you would like to map this against your own studio, that is what my AI implementation consultation covers. For related reading on the brief and revision cycle itself, see design studios using AI for briefs, moodboards and revisions.

Questions studio owners ask

Can we deliver AI-generated images to a client as final artwork?

Only if your contract and the client's expectations allow it, and the tool's terms permit commercial use. Check ownership and indemnity terms, because many tools give no guarantee that outputs are free of third-party rights. For logos and trademarks, avoid AI-generated final marks.

Do we have to tell clients we use AI?

Tell them. Put one clause in your proposal or terms covering where AI may be used and that a designer approves everything delivered. It prevents surprises and answers the question before the client asks it in the middle of a project.

Which stage is safest to start with?

Internal concept exploration and admin. Nothing leaves the studio until a designer has chosen and reworked it, so risk is low, and you learn the tools' limits before they touch client deliverables.

How much should we budget for AI tools?

A five-person studio sharing a few seats might spend roughly $150 to $200 a month on a chat assistant, one image generator and a design platform; giving everyone every tool can pass $300. Check current vendor pages, as credits, plans and prices change often, and cancel anything unused after the first quarter.

Further reads

• Best AI for Interior Design Mockups: Client-Ready Images for a Studio — Client-ready mockup images and their limits.
• How Interior Designers Use AI for Mood Boards and Presentations — Moodboards and presentations in an interior studio.
• Best AI for Brand Identity: Logos, Palettes and What Still Needs a Designer — What still needs a designer in brand identity work.
• How to Use AI for Graphic Design: Brand Kits and Templates for Clients — Brand kits and templates handed over to clients.
• AI for Interior Design Studio Admin: Sourcing, Quotes, Updates — Sourcing, quotes and updates: the admin side of a studio.
• Best AI Tools for Creative Agencies: Briefs, Concepts and Client Sign-Off — Ten jobs from brief intake to sign-off, with Gamma, Canva, ChatGPT and Claude, each with an example, plus rights and disclosure checks.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Adobe Firefly plans page and FAQ; Midjourney plan comparison (docs.midjourney.com); Canva pricing page (archived copy of 1 November 2026); ChatGPT Business and Claude pricing pages; checked 4 November 2026.