How Reckitt is beating the AI odds with its approach to pilots

The approach — which Bastien Parizot, Reckitt’s svp of IT and digital calls “functional reinvention” of the marketing workflow — has delivered, speeding up creative asset adaptation by 30%. In this conversation, Parizot walked Digiday through Reckitt’s pilot process, how he’s managed expectations from corporate leadership and whether it’s affected Reckitt’s agency ties.

This interview has been lightly edited for conciseness and clarity.

Reckitt began organizing its use of generative AI beyond ad hoc implementation last year; what marketing applications have you found for the tech?

The intent was not to run a pilot. The intent was to build solutions that we could embed into marketing teams. Think of the process of approving [campaign] artwork; there’s a lot of regulatory, brand [and] legal rules that you need to respect. Those are known in advance, so we built a solution to make sure that is checked automatically. Then the marketer can focus on the design [of the asset].

We’ve created gen AI solutions for [brand performance monitoring], with which they save 20-40% of their time. They combine Nielsen data and other data into a report … in about 15 minutes. On insights, the idea is different. We sit on a goldmine of data. We connected about 35 different data sources, external and internal, and they have a tool where they can access those insights and create product concepts with it. 

Phase two [in 2025] is to deploy to all of our markets. My team was deploying in Brazil yesterday.

How did you design and deploy the pilot schemes?

Before we built any solutions, we did a time and motion study. We looked at the tasks marketers are doing on a daily basis: which are the ones that are very repetitive? Where it might not be they had the highest value? What are the ones that they enjoy the least? Discussions at the coffee machine can’t be addressed with AI. But artwork you’re approving [repeatedly], this you can address. All of that built up to 10-12 work solutions we believed we could impact. 

Then, we entered the pilot phase. We don’t call them proofs of concepts, but proofs of value; we wanted to check if they worked, if they brought efficiencies, and if they brought better or at least the same quality [of output]. We always have a sample group of human-only… alongside the [group] augmented with AI. We are not looking at anything related to full automation. We have a very strong belief that [AI] should be used to augment humans.

[Now,] we’re monitoring the marketing community … asking them how much time it takes to get basis brand reports, how much time it takes to improve work. We’re doing that on a constant — constant meaning quarterly — basis.

How are you navigating the ‘buy or build‘ question?

I like the word assembling, because we’re not truly building, we’re assembling from different components. We don’t train our own models. We use mostly OpenAI and [sometimes] Gemini…

We are accessing those models through a layer of technology that is on top, namely Databricks. That was a conscious choice because we wanted optionality … and potentially to be future proof. Also, some models are better at certain tasks than others.

[The] biggest benefit [from a direct enterprise deal with OpenAI] would have been receiving the latest technology. In reality, for what we do, we don’t need that. We need strong reasoning. We need strong models, and we need to overlay that with our context. 

Is it difficult to manage expectations from Reckitt leadership?

Where we really make a difference is when we capture tacit knowledge. Capturing our thought processes, making them explicit and then codifying into an AI solution that use an existing LLM, whether it’s OpenAI or something else. People [then] understand that the hard work is in making sure the way that we go about an analysis, the way that we go about artwork approval, the way we brief creative, at scale.

Some of Reckitt’s portfolio lies in areas with regulatory restrictions (such as Durex, for example). With that in mind, what guardrails have you put in place around the use of AI?

We have put a lot of things in place to mitigate hallucination. We’ll use a different LLM as a judge. These are solutions to augment humans … that means that every output is always for a human to make a decision to use or not to use. Last, we’ve put in place a responsible AI framework.

Has your AI usage led you to depend less on agencies?

That’s not what we found so far. I think it’s the same as what happened in media 15 years ago when CPGs or companies started to in-house or “right-house” some competencies. We never really truly went fully internal and typically balanced our relationships with agencies. I think it’s building a healthier relationship with agencies. It creates a decoupling between technology and operations and whether the agencies, creative or media, are really bringing value. There’s still very much reliance and partnership with agencies [such as WPP, which won Reckitt’s European business earlier this month] on driving what we do.

In the future, as use cases for AI increase, will more of Reckitt’s media move in-house? Perhaps bidding algorithms?

There’s a tension here which is hard to resolve for an advertiser, which is the depth of data and the sophistication of the algorithms that are being provided by digital platforms … [it’s] going to be very hard to match from an advertiser’s perspective. 

An overlay that helps you to have a better view of where you allocate across channels, with data that makes sure you have better audience targeting … these are things we already do and I think they will remain and be strengthened. But to completely replace at some point the algorithms? I am very doubtful of that, but I could be wrong; it’s very difficult to be right in predicting anything with AI. But that’s my two cents.