We hosted our third webinar this week, this time with Aidan Lynch and Billy Owen from Hightouch, alongside our own Anthony Fontana. The topic was agentic marketing, what it takes to get there, and what becomes possible once you do. The short version: the technology is further ahead than most data foundations are.
Why this topic, and why Hightouch
One of the principles we hold strongly at Inlight is composability. When we build digital ecosystems for clients we deliberately choose best-of-breed technology that is modular and API-first, chosen on merit rather than because it came bundled with something else.
Agentic marketing is where that principle is about to be tested hardest. Every marketing platform is shipping agents. Very few organisations have the data underneath to make them useful. Hightouch sits directly on the customer data warehouse rather than duplicating it, which makes them a good example of composability applied to the layer that matters most right now.
Watch the webinar
Where the room actually is
We ran a live poll early in the session. Just under half the room were experimenting with AI in campaigns but had nothing live. A third had it running in a few workflows. Only 9% were using it to drive decisions at scale, and only 11% had nothing in play at all.
So almost everyone has started, and very few have got past the pilot phase. That matches what we see in client conversations. The gap is in scaling, and the blocker is almost never the model.
Activation fails because the data layer is weak, not because the AI is weak
Anthony opened with the pattern he sees repeatedly. Siloed data, fragmented stacks, copies of the same customer sitting in three systems with no shared identifier. Agents cannot read that, and they certainly cannot act on it. It is not a new problem. It is just that agentic workflows expose it far faster than a campaign calendar ever did.
“You cannot measure what you cannot join.
Matthew Niederberger, Martech Therapy
The research has been saying this for years. McKinsey found eight in ten companies cite data limitations as the roadblock to scaling agentic AI. Gartner found 67% of organisations had onboarded a CDP, but only 17% reported high utilisation. Salesforce found only 25% of marketers are satisfied with their ability to use data for personalisation.
Anthony set out four things worth getting right before you invest in agents: define your success signals across both transactional and non-transactional journeys, agree your customer identifiers and maintain them across every dataset, build context layers that are readable by an agent rather than a human, and decide how often data needs to refresh.
What composability changes at this layer
The traditional CDP collected, managed, segmented and activated inside one closed system. A composable setup keeps the warehouse as the source of truth and assembles the rest from components. That matters more in an agentic model than it did in a manual one, because an agent is only as good as the context it can reach. A closed system decides that for you.
Scott Brinker's recent report makes the same call, describing composable architecture as the new default assumption and pointing out that it leverages the data infrastructure you have already paid for rather than duplicating it.
What the demo showed
Billy walked through a day in the life of a marketer at a brand like Nike, with one goal: drive repeat purchases. Three moments stood out.
The diagnosis. He asked the agent why repeat purchases were slipping. It confirmed the drop, isolated the cause to first-time football kit buyers returning at a materially lower rate, factored in World Cup seasonality as the reason, sized the opportunity, and recommended what to do next. That is the analyst layer most marketing teams queue up behind.
The audience build. The same thread built the audience from warehouse building blocks, customer attributes, order history, behavioural events, and propensity and lookalike models, with the audience size updating as it narrowed. No data ticket, no CSV, and syncs running on a schedule to paid and owned channels. Holdout groups were configured alongside it to test whether the marketing drove any incremental lift at all.
The approval loop. Creative was generated against the brand's own templates with personalisation pulled from the warehouse, then reviewed by legal and creative agents against internal guidelines before a human signed off. For anyone who has lost weeks to brand and legal review cycles, this was the part that landed hardest.
It closed with a performance report written back to the team, including what to change next. Paid creative generation was covered too, though the approval loop is the more useful story for most teams in the room.
The part that is about people, not platforms
The thread running through the whole session was a change in what marketing teams do. Less time building campaigns and setting scheduling rules. More time setting goals, guardrails and brand context, then judging what the agents bring back. Aidan's read is that teams are becoming more cross-functional, organised around pods rather than channels.
That is a meaningful shift, and it depends entirely on the foundations underneath. Guardrails only work if the context they draw on is accurate and maintained.
Know where your data foundations stand
We built the Data Readiness Scorecard for exactly this. It is a short self-complete diagnostic and we send back a one-page view of where you stand against what agentic marketing requires, plus the recommendations we would action first.
Questions from the session:
What does a martech stack need to look like before any of this is possible?
The non-negotiable is a cloud data warehouse acting as the source of truth, with customer identifiers resolved consistently across the datasets feeding it. After that it matters less which tools you run than whether they are API-accessible and can be written to on a schedule. Most organisations we talk to already own enough tooling. What they are missing is the joined data layer underneath it, which is why we usually start there rather than with a platform decision.
Traditional campaigns have deliberate approval and governance steps. What should we get right before handing decisions to agents?
Brand guidelines, legal requirements and tone rules get encoded into the context layer, so compliance is checked as the work is produced rather than after it. Hightouch flags non-compliant output proactively during the workflow, with the aim of arriving at something close to approved before a human sees it. Strategists and legal handle the final checks rather than the first ten drafts. Our view is that this only holds if someone owns the context, because guardrails inherit the quality of what they are built on.
Are you seeing teams restructure around this, or adapt in place?
Both, but the direction is more cross-functional. The pattern Aidan described is a few foundational pillars in the business, with people coming together in pods to move through cycles faster than a channel-based structure allows. His view is that agents will reshape the marketing department significantly over the next few years.
In a composable ecosystem, how do you stop different teams creating their own customer definitions?
Two layers. Definitions live in the schema layer against the underlying warehouse data, and they are reinforced in context, including how audiences are constructed. The definition follows the data rather than the team asking the question.
Does the context layer become unmanageable at scale? Thousands of lines of facts across multiple teams sounds hard to maintain.
Separate spaces keep context segmented by brand, with guardrails set at that level. There is also an agent that reviews the context you have loaded and looks for redundancy, overlap and contradiction, then recommends fixes. Context also updates itself from the warehouse and from response data flowing back from channels. Worth adding our own view here: any static context set will go stale, so the review process matters as much as the initial build, whatever platform you are using.
Does onboarding start with a large project to assess and clean the warehouse?
No. Start with one high priority use case and load only the data that use case needs, then expand. Start with the use case and work backwards to the data, rather than trying to get the warehouse perfect first.
