Our AI use policy and intent.

2. Regulatory context

We treat AI use as subject to existing law, not exempt from it. The following frameworks are most relevant to our marketing, social, and content work:

• UK advertising (CAP and BCAP Codes, enforced by the ASA).Existing advertising rules apply to all content regardless of how it is produced. AI does not create a new category of advertising or exempt content from the standard. Disclosing that AI was used does not cure a message that is misleading on its merits.
• EU AI Act (Article 50).Transparency obligations for AI-generated content apply from 2 August 2026. As a deployer (a user of third-party AI systems rather than a builder of them), our main duties are to label deepfake-style content and any AI-generated text on matters of public interest, and to use tools whose providers support machine-readable marking of outputs. The provider-side marking obligation has a grandfathering extension to 2 December 2026 for systems already on the market.
• Data protection (UK GDPR and the Data Protection Act 2018).Inputs to AI tools that contain personal or client data are handled under the same obligations as any other processing.
• Competition and consumer law (CMA).Strengthened consumer-protection enforcement (DMCC Act 2024) means misleading or non-compliant marketing carries materially higher financial risk than before.

Our role in the chain

We are a deployer of AI systems, not a provider of them. We state this in client responses so that obligations are allocated correctly.

3. Our position on AI

We see AI as a powerful accelerator, not a replacement for human expertise, judgement, and creativity.

3.1 Core beliefs

Our position rests on four principles:

• AI as a collaborator.AI works only when it augments human skill. We use it to handle repetitive tasks, generate starting points, and surface insight, freeing our team to focus on the strategic and creative work that delivers real results.
• Transparency first.We are open with each other and with clients about where and how AI is used, and we disclose AI involvement to consumers wherever that is required or wherever non-disclosure could mislead.
• Quality over speed.AI can accelerate delivery, but never at the expense of quality. Every AI-assisted output goes through the same quality standards as any other piece of work.
• Continuous learning.The AI landscape moves fast. We commit to ongoing experimentation, knowledge-sharing, and honest assessment of what is working and what is not.

3.2 What AI does not replace

AI does not replace strategic thinking, client relationships, creative direction, or professional accountability. A named team member is always accountable for any AI-assisted output that reaches a client or the public.

4. How we use AI in our workflows

Practical guidance on where AI fits into day-to-day work across departments.

4.1 Approved use cases

The following are areas where AI tools are actively encouraged, provided outputs are reviewed and refined by a team member:

• Content drafting and ideation.Generating first drafts, blog outlines, social copy variations, captions, and meta descriptions for review and refinement by the content team.
• Design exploration.Creating mood boards, layout concepts, and visual inspiration to accelerate the early stages of the design process.
• Development scaffolding.Generating boilerplate code, component structures, and design-to-code translations (for example, Figma to WordPress or Shopify) to speed up build phases.
• Data analysis and reporting.Summarising analytics data, identifying patterns, generating report narratives, and supporting GA4 audits and funnel analysis.
• SEO and PPC support.Keyword research, ad copy generation, bid strategy recommendations, and competitor analysis.
• Internal operations.Meeting summaries, process documentation, project scoping, and workflow automation.

4.2 AI-generated imagery, video, and influencer content

Because we work with lifestyle, beauty, and cooking brands, generated visuals carry specific risk. The following rules apply:

• We do not use AI-generated or AI-altered imagery to depict product results, efficacy, or before-and-after effects. AI imagery is treated exactly as we treat retouching: it may not create an impression the real product cannot deliver.
• Where AI is used to produce a fully synthetic person, presenter, or influencer, or to manipulate a real person's face or voice, this is labelled clearly to the audience.
• Influencer content remains subject to advertising-disclosure rules (for example #ad) in addition to any AI labelling. The two are separate obligations and we apply both.
• Final published imagery, video, or animation generated with AI is signed off by the account lead against the client's brand and the accuracy rules in section 6.

4.3 Boundaries and restrictions

The following are not permitted:

• Publishing AI-generated content without human review and editing.
• Presenting AI-generated final creative (images, video, copy) to clients as original human work without disclosure.
• Inputting confidential client data, credentials, or sensitive business information into any third-party AI tool that has not been approved for that purpose.
• Relying on AI-generated code in production without peer review and testing.
• Using AI tools that have not been vetted and approved (see section 9 and Appendix A).

4.4 Approved tools

We maintain a register of vetted AI tools that the team is approved to use (Appendix A). It is reviewed quarterly. Before a new tool is adopted, it is assessed for data-privacy compliance, commercial licensing, output quality, and fit within our workflows.

5. Transparency, disclosure, and labelling

We distinguish between two kinds of transparency, and we apply both.

• Transparency to clients.We tell clients where AI has contributed to their work, and we will provide a usage summary on request.
• Transparency to consumers.Where AI involvement is material to how the audience understands the content, or where it is required by the EU AI Act or advertising rules, we disclose or label it so the consumer is not misled. AI-assisted content is tagged in our project management workspace against the asset itself, so the contribution history is clear on review.

Two levels of AI involvement

We also distinguish between two levels of AI involvement:

• AI-assisted.AI is used to assist a human-led process (for example drafting copy a person then rewrites, or removing a stray background object). This generally does not require consumer disclosure, in the same way routine post-production has never required it.
• AI-generated.AI produces a substantial finished asset (for example a synthetic image, a generated video, or a synthetic presenter). This is reviewed against the labelling tests above and labelled where required. We do not produce purely AI-generated final assets for publication without explicit client agreement.

Our guiding principle

Disclosure is not a shield. We never make a claim or create a visual impression that is misleading and then rely on an "AI was used" label to excuse it. If a message would mislead, we change the message.

6. Accuracy and product claims

Any factual or efficacy claim in client content is checked against client-supplied evidence before publication. We do not let an AI tool invent, infer, or embellish product claims.

Efficacy claims for health, beauty, food, and supplement products are held to the relevant sector substantiation requirements. AI-generated supporting visuals must not imply results the evidence does not support.

Statistics, quotes, citations, and any figures produced by an AI tool are verified against primary sources before use.

The reviewing team member is accountable for the accuracy of the final output, whether or not AI contributed to it.

7.1 Data handling principles

We only use AI tools on team, business, or enterprise plans with commercial data-protection terms in place. We do not upload client data, brand guidelines, strategy documents, or any other sensitive material into consumer or free-tier AI tools. Because the plans we use contractually exclude customer inputs from model training, nothing we provide to these tools is used to train the underlying models or made available to other users.

We never input client-sensitive data into AI tools unless the tool's data-handling and privacy terms have been reviewed and approved. This includes customer databases, login credentials, financial data, personally identifiable information (PII), and any data covered by NDAs or contractual obligations.

We understand how each tool handles data. Where a service uses inputs for model training, we opt out where possible, or we do not use it for client work.

We use anonymised or synthetic data when testing AI workflows that need realistic inputs, and we strip identifying details before pasting content into any AI interface.

We respect UK GDPR and the Data Protection Act 2018 at all times. If in doubt, we consult the document owner before proceeding.

7.2 Tool assessment criteria

When evaluating any AI tool for client work, we assess:

• Where data is stored and processed (UK and EU preferred; US acceptable with adequate safeguards).
• Whether inputs are used for model training and whether opt-out is available.
• Data retention policies, meaning how long inputs and outputs are stored.
• Compliance certifications (SOC 2, ISO 27001, and similar).
• Sub-processor and third-party data-sharing arrangements.
• Commercial licensing and indemnity terms for enterprise or business use (see section 9).

8.1 The human-in-the-loop principle

Every piece of AI-assisted output must be reviewed, refined, and approved by a qualified team member before it reaches a client or goes live. This is non-negotiable. AI outputs are starting points and require the same editorial judgement, design critique, code review, and strategic scrutiny as any other work.

8.2 Known risks and mitigations

• Factual inaccuracy (hallucination).Always verify facts, figures, and claims generated by AI against reliable sources.
• Bias.Review outputs for unintended bias, particularly in content targeting diverse audiences.
• Intellectual property.AI-generated content may reproduce protected material. We run plagiarism checks on substantial text and reverse-image-search visual assets where appropriate.
• Over-reliance.If we find ourselves accepting AI outputs without critical assessment, we step back and recalibrate.
• Security vulnerabilities.All AI-generated code passes our standard review and testing process before deployment.

9.1 Commercial safety and licensing

• We use AI tools on business or enterprise plans where client work is involved, so that commercial-use rights, confidentiality, and (where offered) IP indemnity apply.
• We do not use consumer or free tiers for client deliverables where those tiers permit training on inputs or lack commercial-use terms.

9.2 Provenance and auditability

• Approved tools and their plan level, data location, and training status are recorded in the register at Appendix A.
• Where AI tools support content provenance signals (for example C2PA Content Credentials), we retain them rather than stripping them, in anticipation of the EU AI Act marking regime.
• For AI-generated final assets, we keep a record of the tool used and the responsible reviewer, so any asset can be traced after the fact.
• Generated assets are stored in the client's project workspace under version control alongside human-produced work, so the history of a deliverable is reconstructable.

9.3 Internal guardrails

Tool approval, data-handling rules, the human-in-the-loop requirement, and named accountability together form our guardrails. They are set out in this document and owned by the AI Governance Lead.

10.1 Our commitments

• We are honest with clients about our use of AI and explain openly how AI contributed to any piece of work.
• We do not misrepresent AI-generated work as entirely human-created where the distinction matters (for example awards submissions, case studies, and testimonials).
• We protect the interests of our clients by handling their data securely and representing their brand accurately.
• We invest in our people rather than replace them. AI adoption should upskill the team and create space for higher-value work.
• We stay informed about the ethical, legal, and environmental implications of the AI tools we use and factor these into our decisions.

10.2 Accessibility and inclusion

AI-generated content, designs, and code meet the same WCAG and brand accessibility standards we apply to all work.

11. Governance and review

• Document owner.Head of UX Design & Development (CTO), acting as AI Governance Lead, responsible for maintaining and updating this document and the tools register.
• Review cycle.Reviewed at minimum every six months, or sooner if a significant shift in AI capability, legislation, or agency practice warrants it. The next scheduled milestone is the EU AI Act transparency deadline in August 2026.
• Feedback.Any team member can propose changes or raise concerns. Suggestions go to the document owner.

Appendix A: Approved tools register

The register below is maintained by the document owner. For each tool, it tells us: whether it is commercially licensed for our use, whether it trains on our inputs, where data is processed, and who approved it.

• Anthropic Claude (Projects, Skills, Claude Code).Proofreading, copy refinement, tone adjustment, caption improvement, note analysis and summarisation, research, content trends, competitor review, trend analysis, tech and troubleshooting support, concept development, image generation (background and graphic elements, disclosed to client when used), market and context checks, and translation review. Team plan — does not train on inputs — commercial terms apply.
• Canva (Magic Studio).Concept and design exploration; graphic and layout production; building visual creator briefs, campaign documents, and influencer decks. Team plan — does not train on inputs — not used for efficacy visuals.
• CapCut.Short-form video editing: auto-captions and subtitles, background removal, stabilisation, quality enhancement, transitions, and filter effects. Pro plan — does not train on inputs.
• Lightroom.Photo editing: AI denoise, masking, and selective adjustments for grading and retouching. Pro plan — does not train on inputs.
• Adobe Podcast.Audio enhancement: removing background noise and improving clarity on voice-over and recorded audio for video content. Pro plan — does not train on inputs.
• Photoshop (Adobe Firefly).In-built in Photoshop. Image editing: removing imperfections, cleaning up subjects, and removing backgrounds. Pro plan — does not train on inputs.

Appendix B: Notes for client reviews

• Scope check.When a client raises EU AI Act compliance, we confirm whether their content is placed on, or directed at, the EU market, since that determines whether the Act's transparency obligations bite. We respond as a deployer.
• Evidence pack.We can share a representative, labelled sample of AI-assisted or AI-generated work for a defined client and period, with context on how each item was produced and reviewed, rather than an undifferentiated export.
• Policy evidence.This document is the artefact we share to evidence our internal policies and guardrails.