Anthropic’s Claude now embeds an invisible watermark directly into the text it generates using patterns of text it chooses, a requirement driven by the EU AI Act’s transparency obligations. Every model launched from that date carries it by default, and older models are being retrofitted. Image and file outputs carry a different mark: signed C2PA provenance metadata confirming the file was created or touched by AI.
Platforms are moving at the same pace. In July 2026, LinkedIn added a “seems like AI slop” button to every post, after research found that more than 40% of long-form posts on the platform were fully AI-generated. Over a million people used it in the first two weeks, and flagged posts get buried even more in the feed.
Regulators are catching up too: New York’s Synthetic Performer Disclosure Law, in force since June 2026, requires advertisers to conspicuously flag any AI-generated performer in a commercial, and the FTC has confirmed that undisclosed AI-generated endorsements can breach Section 5 of the FTC Act.
Put those three together and the picture is consistent. The platforms your content lives on, the regulators overseeing your market, and the readers scrolling their feed are all now taking steps to highlight AI. That’s the backdrop for everything below: not a reason to avoid AI, but a reason to be specific about how we use it, rather than leaving it as a line in a policy document nobody reads.
Olev Intelligence: our AI ecosystem
That’s the environment we’re building in, so here’s what we’ve actually built. AI at Velo runs through Olev Intelligence, our own AI-powered ecosystem, connecting our proprietary IP, market and performance data, and sector insight into one system our teams draw on daily. It strengthens every part of the agency: research, buyer profiling, ideation, content, campaign delivery and production.
It isn’t a single chatbot bolted onto the business. It’s a connected set of purpose-built tools, each one carrying its own guardrails and each one built on close to two decades of Velo’s own intellectual property: our frameworks, our tone of voice rules, our templates, our methodologies to structure a brief, a wash-up deck or to solve a problem.
It’s built in a way that is secure by design- our client roster includes Nasdaq, FTSE, AIM and NYSE listed companies, each with their own AI governance, data handling and disclosure requirements, so Olev Intelligence must hold up to that scrutiny by default.
How we use it to think, plan, do
Olev Intelligence is designed around three stages of work, and a different type of AI does the heavy lifting at each one.
Think
Everything starts with ideation, research synthesis and first-pass analysis, turning a blank page or a pile of source material into a starting point a strategist can react to. Three things make that reliable rather than generic:
- Ring-fenced client-specific GPTs, used only for specific clients and to act as the context hub, learning constantly through use and maintenance. This acts as the golden client-specific thread
- Process meeting notes, interview transcripts of provided assets to distil into structured insight around specific questions to accelerate discovery phase, particularly around success story interviews.
- Separate, controlled instruction agents specific to a task that follow specific steps and outputs. Knowledge files carrying Velo’s own sector research and proprietary IP, so the model is using proven knowledge, not inference from the internet. Each one of our 100+ models is configured for one kind of thinking task rather than one general-purpose tool trying to do everything. It means consistency, quality and agility for every Velo client.
Plan
Once there’s something worth building on, our planning tools take that thinking and turn it into a structured plan grounded in Velo’s own frameworks, not generic best practice. That’s down to three things:
- Programming each tool with Velo’s colour palettes, template layouts and design systems, so a plan comes out looking and reading like Velo work from the first draft
- Assessing impact using synthetic personas, past campaign performance and market insight as grounding material, so new plans build on what’s already proven rather than starting from a blank slate.
- Locking each tool to a fixed output format, such as our briefs, creative guidelines, messaging or ABM audience insight, so nothing gets invented or skipped for accelerated concepts.
Do
From there, agentic workflows carry the plan through to a finished, formatted output, running a sequence of steps on their own rather than waiting for a person to move each one along by hand. That only works within limits:
- Every agent operates inside a structured, controlled workflow with a defined start and a defined end, not an open-ended task.
- GenAI models are used to produce print, digital, video and audio output, automating, where possible, production steps. Steps such as cut-outs, retouching and image extensions lean on AI where appropriate.
Think, plan and do sound like three separate jobs, but they run as one line: research feeds the plan, the plan shapes what an agent goes on to build. What holds that line together isn’t the technology, it’s three things underneath it, which is what the rest of this piece covers. The glue is our people.
A person checks the work, not just the result
Our people sit inside the process, not just at the end of it. Every workflow above has defined checkpoints where someone reviews the work before it moves to the next stage, and checks built in: research gets checked before it’s used to write anything, a first draft gets checked before it’s formatted into a client-ready document, and the finished piece gets signed off by the person who owns that discipline before it reaches a client. We’ve learnt that shorter-form outputs such as social media posts are more suitable to AI production than long-form case studies, even though AI as a sub-editor or proofreader can be excellent. Each of those checkpoints is owned by an expert in that specific area, which is what makes it a real check rather than a formality.
A second model checks the first
We use models to provide additional diligence at key stages. Where accuracy genuinely matters, such as cited statistics, dates, regulations or client figures, we don’t rely on a single model’s output. A second, different AI model checks the first model’s claims and references before anything is finalised. Different models make different mistakes, so running a claim past two systems built differently catches errors a single pass would miss, before it ever reaches a human reviewer.
Copy is where this gets tested hardest
Those three checks matter everywhere, but copy is where they earn their keep, because the risk with AI-written copy isn’t usually facts, it’s tone and consistency. A person builds the structure first: the argument, the flow, the headline logic. AI never decides what a piece should say before a human has. From there, AI drafts section by section within that structure, so the piece is assembled to a plan rather than generated in one pass and hoped into shape. Before delivery, a final pass checks the whole piece for typos, adherence to Velo’s brand voice rules, and specific vocabulary, such as how we refer to particular tools, products or platforms by name, so terminology stays consistent even when different sections were drafted separately.
Client policy still comes first
All of the above is Velo’s own baseline, and it sits underneath whatever a specific client asks for, not above it. Where a client has their own rules about what AI can and can’t touch, whether that’s a blanket restriction, an approval process, or a required disclosure format, that policy governs for that account. Where there’s a conflict, we follow the stricter standard, not our own default.
Why we’re telling you this
None of this is a policy document we keep for ourselves. If AI played a meaningful part in something we’ve delivered, you’re welcome to ask how, and this is the honest answer: Olev Intelligence, scoped tools built by the people who know the work, research before drafting, a person and a second model checking the output, and a named individual who signs off the final version. That’s how it works, and it’s the standard we’re happy to be checked against.