Building an AI knowledge base for creative teams

The Native Advertising Institute (NAI) has spent years collecting and analysing marketing campaigns from around the world. The challenge was turning that archive into a practical resource for creatives and account managers. Together, we built NAI+ Studio, an AI tool that helps advertising teams research clients, find inspiration and develop client proposals.

In a blind test, campaign concepts generated by the tool scored higher than those from an award-winning advertising studio, as judged by their own team. In a second test, 94% of its concepts were rated ready to send to a client without further work. 

But this isn't a story about AI outperforming people. The model itself wasn't what made the difference. What we built around it was, and none of that would have existed without human creativity and craft. Let's take a closer look at how we put both to work with NAI+ Studio.

From campaign archive to AI knowledge base 

The Native Advertising Institute has built up a rich archive of marketing inspiration, with thousands of award entries, blog posts, podcasts, and insider content available to members. Its users include brand studios from around the world.

NAI wanted to make that archive easier to use in day-to-day creative work. The idea was to bring its industry knowledge together with each studio's own published work, so creative teams could spend less time searching and more time developing ideas.

The knowledge already existed, but it was almost impossible to use when someone needed it. In a client meeting. While looking for inspiration. Against a deadline. It was scattered across studio files, NAI's archive, and the heads of the people who had been in the room.

Our goal was to make that knowledge searchable in one place. 

How the NAI+ Studio AI knowledge base works 

We built a knowledge base with a chat interface, combining NAI's own content and each member studio's published work.

Studios talk to it much like they would message a colleague. The tool interprets the request and routes it to one of eight specialised functions, from drafting pitches to finding inspiration to generating campaign debriefs and even creating content.

Each function looks for different information and uses it in a different way. No single prompt tries to do everything.

That's context engineering in practice: deciding which data matters for a particular task, how to find it, and how to pass it to the model in a format it can use.

Three use cases: how NAI+ Studio helps creative teams

1. Meeting prep in minutes

You’re heading into a client meeting and need a quick refresher on the campaigns you’ve done for them before, where they were published and the briefs behind them. 

Instead of digging through folders, emails and spreadsheets, you can ask the tool and get the relevant brief and context in one place. It cuts meeting prep from an hour to a few minutes.

2. Finding inspiration with the strategy behind it 

As a creative, you might be looking for inspiration to get an idea moving. Other times, the idea is already there, and the real question is whether anyone has tried something similar and how it performed.

Your questions can be broad or specific: Show me this year's award winners. What have people built around a podcast? Which campaigns actually drove sales? The tool responds with relevant work from studios around the world, along with the creative rationale behind each campaign. 

What you get is a shortcut to the research behind the idea, plus the answer to the only question that really matters: why did this work?

3. Developing ideas from real campaign data

The campaign concepts created by NAI+ Studio hold up because they're specific and grounded in the actual brief, the studio's own published work, and real examples of what's worked before for similar products and audiences, not in whatever the model already assumes about the industry. 

It also isn't one prompt doing the work. Two agents are put on the task, each with its own job:

• The detective takes the brief apart, researches the client and the market, and pulls the most relevant information from the archive. 

• The creative works with that research to develop the idea. The result stays grounded in real context rather than generic internet knowledge. 

Once the campaign concept is approved, the same tool can draft the actual advertorial or native content, using the same brief and archive and writing in the studio's house style.

How is NAI+ Studio different from generic AI chatbots? 

Most AI-generated content sounds generic. That's part of what fuels the worry that AI flattens creativity instead of supporting it.

Usually, the problem is a lack of relevant information. A model with little context falls back on broad patterns and average, familiar ideas. Give it the actual brief, the client's history and relevant examples, and it has something specific to work from. 

The pieces were always there. The important part is putting the right ones together. That's what we built around the model: a system that decides which pieces matter, finds them, structures them and puts them in front of the model at the right moment. The human touch is built into the context the model works from. 

Search that doesn’t guess

Say a team is working on a podcast campaign and wants to see what has worked recently. They might ask: Best award-winning podcast campaigns from the Nordics in the last three years, targeting young adults. 

Several criteria are packed into one sentence. The tool pulls them apart and checks each one against values that actually exist in the archive. If a criterion cannot be matched, it leaves it unresolved rather than guessing. 

The phrase “targeting young adults” was never tagged in the data. So the tool searches for that part in two ways at once: one based on the words used, the other on their meaning (also called hybrid search). It then merges those results with the structured search. The result can be precise where the data is structured and flexible where it isn't.

Data pipeline built around campaign context

Before anything gets indexed and goes into the knowledge base, content is broken down into the information that matters most: 

• Goal and audience – who it was for and why.

• Strategy – how the team planned to reach them.

• Creative choices – the format, channel and idea.

• Results – how it performed.

That gives the AI more to work with than keywords alone. Instead of simply finding a campaign that mentions podcasts, for example, it can understand why a podcast was used, who it was meant to reach and what the campaign achieved.

Human expertise, baked into the context 

We started with one question: How do we make sure the AI isn’t simply remixing what’s already been done? We knew we didn’t want to recycle old ideas and dress them up as new. The NAI archive gave us something useful here: not just the campaigns themselves, but the thinking behind them. Why a particular audience insight led to a strategic decision, and how that choice shaped the creative idea. 

So when NAI+ Studio is asked to develop a new idea, the system gives the model relevant examples of how professionals approached similar creative problems. It can then use that reasoning as a starting point for the new concept, rather than just drawing from a library of past ideas.

How NAI+ Studio performed in blind tests  

We ran several tests to check whether the system holds up in practice:

• NAI+ Studio vs an award-winning studio
  We tested NAI+ Studio's campaign concepts against concepts created by a native advertising studio using the same brief. Nine people from the studio's content team, including its lead, judged the concepts without knowing who had created them. NAI+ Studio scored 8.06 against the studio's 7.85, ahead on every measured criterion: how well it used the client's products, how relevant the pitch was to the client, and how strong and creative the idea itself was.

• NAI+ Studio vs an LLM without added context
  We gave the same brief to a frontier LLM without access to NAI's archive, the studio's work or its house style. 94% of NAI+ Studio's concepts were rated ready to send to a client without further work. For the bare model working from the brief alone, that figure was 9.5%. The difference came from the context available to each system.

• NAI+ Studio vs standalone LLMs at scale
  A separate test compared NAI+ Studio with regular LLMs using the same brief, but with no access to the knowledge base behind NAI+ Studio. used the ordinary way (a capable model and a prompt, no dedicated retrieval or context engineering) evaluated across the same criteria by eEight people from the client's content team and a separate three-person evaluation team assessed the results. NAI+ Studio came out ahead by 22% and 24% across the two evaluations.

The average wasn't the most telling part, though. Standalone LLMs were much less consistent, with some scores dropping as low as 2 out of 10 even where their overall averages looked solid. NAI+ Studio's lowest score across the same tests never dropped below 5.9. The average tells you which one is better on a good day. The range tells you which one you can actually rely on.

How to turn company knowledge into better AI outputs 

Most companies have an archive full of valuable knowledge. The problem is finding and using it when it matters.

AI can help, but simply uploading that archive to a chatbot isn't enough. The underlying system still needs to structure the data, select the context that matters for each task and keep the model grounded in what the company actually knows.

That's what we built for NAI. A way to turn years of accumulated knowledge into something people can use in their daily work. 

Planning an AI product around your own data and expertise? At Applifting, we can map the right sources, prepare them for the model and scope the first use case. Send us a brief, and we’ll show you where we’d start.