Artificial Intelligence in Design: Who Approves the Output?

Artificial intelligence in design uses computational models to generate, modify, or assess creative work - but for professional services firms, the critical challenge is governance. 

Your team can use generative tools to produce a polished proposal, but the harder decision is who holds the authority to approve what that proposal promises.

For a firm preparing a rebrand, the useful boundary lies between producing an asset and authorising it to represent the business.

The exposure is already being reported. 

That covers AI generally, not design alone, and does not establish your firm’s likelihood of an incident.

The practical response belongs within brand management: decide what can be generated, what evidence must support the result and who can release the final version.

For firms rolling out a new identity, those controls should sit inside the wider brand operating system, alongside approved assets and everyday production workflows.

How do generative models fit into brand design? 

Automated design tools use computational models to generate, modify, or assess creative work, including images and layouts. 

While generative systems create net-new content, other algorithmic capabilities assist in editing or analysis. 

McKinsey’s 2024 implementation guidance recognises image generation as a distinct capability and recommends controls tailored to the use case:

• Production permission: Defines approved tools and inputs.
• Business authorisation: Assigns responsibility for commercial claims.
• Release approval: Covers the final asset and its public use.

Artificial intelligence in design uses computational models to generate, modify or assess creative work within a human-defined brief and approval process.

What are we paying professionals for when AI can generate designs?

You should be paying for a defined result and the judgement needed to deliver that result. Generating options is one part of the work.

Ask the agency to separate concept generation from the work needed to make the selected concept usable: research, art direction, production files and accountable review.

Take an illustrative task: a consultancy needs imagery for a new infrastructure-advisory practice. Generating an impressive bridge image does not answer whether that bridge implies a completed client project, whether the visual direction distinguishes the consultancy or whether the files work across a website and printed report.

Those decisions remain in the brief, whether the first image comes from a model, a photographer or an illustrator.

McKinsey’s March 2025 report associates workflow redesign with reported financial impact from generative AI. The finding supports examining the whole process rather than assuming faster generation produces a proportionate reduction in the total job. The research does not establish a design-agency fee benchmark.

Can an AI-generated image replace the agreed brief?

No. Under the governance model proposed here, an AI-generated image can inform a brief change, but the person with authority over the brief must approve that change.

Suppose a partner prefers a generated image showing a large international team—however, the agreed positioning centres on direct access to a small specialist group. The image preference now raises a commercial question: which service experience should the firm represent?

Return to the decision criteria. Ask what the partner likes: the composition, the colour treatment or the suggestion of global capacity. The first two may fit the existing brief. The third may require a different proposition.

A new visual preference is useful feedback. Record any resulting change to positioning, scope or cost before commissioning further production.

What should an agency explain before the client asks?

Require a short AI-use statement at proposal stage. The statement should identify intended uses, permitted inputs, approval responsibilities and any output restrictions requiring further review.

Keep the statement specific to the engagement. “We use AI responsibly” gives a managing director little to evaluate.

For an identity project, the statement might distinguish internal concept exploration from final asset creation. For a proposal-template project, the statement might permit layout assistance while excluding the generation of client results or service guarantees.

The relevant professional responsibility is to explain the decisions behind the deliverable, including those that software cannot authorise on the firm’s behalf.

Is our confidential material safe in the tools our team uses?

Treat confidentiality as a condition to establish before upload. The available evidence does not justify a blanket assurance about every AI design tool or account.

McKinsey’s 2024 implementation guidance identifies third-party exposure, privacy and intellectual property among the risks organisations should assess. McKinsey also recommends controls across the workflow rather than relying entirely on the underlying model.

For a rebranding firm, start by identifying the material people might upload: an acquisition presentation, an unpublished client case study, partner portraits or interview transcripts.

Each input raises a different permission question. A public logo and a confidential transaction diagram should not inherit the same approval merely because the same application can process both.

Do protections apply to the account and model actually being used?

Approve the tool, account, model and use case together. Permission to generate an internal abstract background does not automatically include permission to upload a client’s acquisition plan.

Ask procurement, IT and the relevant legal adviser to establish:

• Which account and contractual terms cover the work?
• Which models or connected services can process the inputs?
• What retention, access and training-use provisions apply?
• Which categories of client or personal information are permitted?
• What commercial-use restrictions and contractual protections apply to the intended output?

These are due-diligence questions, not assurances that a particular supplier offers a particular protection.

Require agencies and freelancers to follow the same approval boundary. If an agency proposes a different account or model during the project, review the change before work continues.

The application’s name is not the approval record. Record the permitted configuration and purpose.

What provenance should we retain?

For this proposed framework, provenance refers to the record of where an asset originated and which permissions support its intended use.

Retain enough information to answer a later challenge without reconstructing the project from someone’s memory:

• The originating tool, model and generation date.
• The source assets supplied to the workflow.
• The applicable permission or licence records.
• The final selected version and material human edits.
• The release decision and any restrictions.

Keep the record proportionate. An internal concept board does not need the same release pack as imagery supporting a public launch.

Do not retain confidential prompts indefinitely merely to document the process. Agree what must be recorded, where the record belongs and when the record should be deleted.

This recommendation follows McKinsey’s broader guidance to combine technical controls, contractual provisions, metadata and operating responsibilities. The exact record described here is an editorial proposal, not a documented industry standard.

Who can authorise what AI-generated design says about our firm?

Separate permission to produce from permission to represent.

Permission to produce allows someone to create work within approved boundaries. Permission to represent allows a specific asset to communicate a claim about the business.

A designer may have the first permission without holding the second. A practice lead may approve a service claim without being qualified to approve the production file.

That distinction extends the sensible argument for brand guidelines. Approved colours, typography and image rules give teams a shared starting point. 

Automated checks can help identify departures from those rules. Neither function by itself establishes that a proposed service promise has supporting evidence.

The underlying survey was conducted in July 2024. Respondents in business, legal and other professional services were more likely than respondents in other industries to report reviewing all outputs.

Review coverage matters. So does the reviewer’s remit.

“Marketing approved it” is incomplete if marketing checked the layout while assuming a partner had approved the claim.

Who should approve claims in AI-generated proposals?

Assign a business owner to each material claim and a release owner to the final proposal.

The business owner checks whether the firm can support the statement in the relevant context. The release owner confirms that the approved statement appears correctly in the version being sent.

Use the following as a starting allocation, then adapt the roles to your partnership structure. The table is a proposed decision framework, not a claim that these job titles carry identical authority across all firms.

McKinsey’s 2024 guidance supports assigning business-led risk assessment and cross-functional review to AI use cases. The proposed table applies that principle to brand production; however, McKinsey did not test it.

Avoid sending every row to the CEO. Reserve leadership decisions for positioning, changes to distinctive identity, and exceptions above delegated limits. Give other owners authority they can actually exercise.

What can go wrong when the design looks correct?

Consider a hypothetical 120-person accountancy firm preparing a rebrand after an acquisition. This is an illustration, not an Inkbot Design client case.

The marketing manager uses an approved AI workflow to develop a proposal cover. The final deck follows the correct colour palette and typography. A generated section heading reads: “Your dedicated partner, available whenever you need us.”

The visual review passes.

The service claim has not yet been approved. The acquired practice operates under a different contract model, and no one has established whether the promise applies across the combined firm.

The proposed control works like this:

1. The designer checks visual execution.
2. The practice lead checks the contract commitment against the service being offered.
3. The proposal owner replaces or qualifies the heading if the evidence does not support the wording.
4. The release owner approves the corrected version for the named prospect.

The same test applies to imagery. A generated international-office montage may imply geographic reach. A project visual may imply delivered work. Ask what a reasonable buyer could infer, then identify who can verify that implication.

The example demonstrates the framework’s logic. No measured reduction in complaints, rework or commercial loss is claimed.

Should previous approvals become automatic rules?

Only after the relevant decision owner confirms that an approval expresses a reusable rule.

An exception for one tender may depend on the audience, client permission or a contractual condition. Recording “approved” without those conditions can turn a narrow decision into an unintended default.

Give exceptions a scope and review point. For example: approved for this proposal, for this recipient, using this wording; not approved for the website.

McKinsey’s November 2025 survey found that AI high performers were more likely to report defined processes for deciding when model outputs require human validation. That association supports taking validation design seriously; the research does not prove that a particular approval workflow causes higher performance.

The replacement directive is specific: encode repeatable decisions as rules, and keep conditional decisions attached to their conditions.

Is AI-generated artwork ready for production, and who can change it afterwards?

Treat generated artwork as a candidate deliverable until the agreed production checks are complete.

A persuasive preview does not establish that the file is suitable for every intended use. Ask the production specialist to test the selected asset against the actual destination: website, proposal, event display or printed report.

The required checks depend on that destination. A small proposal illustration and a large exhibition graphic have different file requirements. A decorative background and a diagram explaining a client process have different accuracy requirements.

McKinsey’s implementation guidance recommends managing risk at multiple points in the workflow, including controls that prevent inappropriate direct release into production. The following release checks apply that principle to design work.

What should the final production check cover?

Agree acceptance criteria before generation begins.

For the selected asset, check:

• Resolution, dimensions, crop and file format.
• Legibility of any visible wording or labels.
• Consistency with approved identity assets.
• Accuracy of depicted details where accuracy matters.
• Permissions and restrictions for the intended use.
• The final exported version, rather than only the working preview.

Assign each check to someone able to judge the requirement. A print supplier can advise on production suitability; a practice lead must verify the depicted service or process.

Do not confuse a completed generation with an accepted deliverable. The contract should describe what the client receives, including editable files where required and feasible.

Who can approve changes after sign-off?

Define approval as applying to a specific version and use.

Under the proposed change-control rule, later edits reopen the relevant approval whenever they affect meaning, permissions, identity, or production suitability.

Correcting an agreed typographical error may stay within delegated production authority. Replacing a project image, adding a performance claim, or changing an identity device should be returned to the decision-maker.

A partner’s seniority does not make every edit purely a matter of personal preference. If the partner changes a service promise, record the business approval. If the partner regenerates the artwork, confirm whether the new asset requires fresh rights and production checks.

Keep superseded versions out of the release location. Retain the audit record separately in accordance with the firm’s retention rules.

Where should we start without creating another approval bottleneck?

Start with a bounded workflow, not an organisation-wide permission to “use AI”.

For a firm of 50–200 people, a suitable pilot might be generating internal visual directions for a proposal template using public or otherwise approved material. The pilot should have a defined owner, permitted inputs and a clear boundary between exploration and publication.

Do not choose that pilot if the urgent problem is a positioning dispute. Agreeing how the firm should present itself comes before automating the disputed message.

McKinsey’s 2024 guidance recommends adapting existing governance where possible rather than creating an entirely separate set of approval bodies. Apply that principle to the people already responsible for brand, client information and proposal release.

Before the pilot begins:

Measure the complete job. Track time to approved output, review effort and material corrections, rather than counting generated options as the result.

For a small asset library and a single production team, a shared register and a controlled folder may be enough to test the rules. 

Consider additional software when the pilot exposes a specific access, versioning or workflow problem. The coordination discipline is explored further in BrandOps roles and workflows.

Delegate production within limits

The business can permit image generation while retaining control of the promises, evidence and exceptions attached to the result.

That is a workable starting position for a rebrand: keep low-risk production moving within agreed limits, and route meaningful changes to named decision owners. The proposal cover does not need a board meeting. A new promise about partner access does need someone accountable for delivering it.

If you need an independent review of your positioning, messaging and visual consistency, request Inkbot Design’s free Brand Equity Audit™. The written diagnostic identifies priorities you can address internally or with external support. It is not a substitute for legal or security review.

Before the next AI-assisted proposal leaves the firm, name who approves what the proposal says.