Summary
This webinar, hosted by Navaneetha (product manager at Lyzr, building marketing agents) and Nirupam (product engineer who built and maintains Skott, Lyzr’s internal marketing agentic OS), laid out a blueprint for building an internal marketing intelligence system. They opened with a live, unscripted demo: Nirupam asked Scott (accessible directly inside Slack) to pull banking case studies and check deal status on existing opportunities, showing it querying Lyzr’s internal knowledge base (via Gemini embeddings) and CRM integrations in real time, with answers returning in roughly 8-11 seconds. They noted the point wasn’t speed but that the answer had existed in company systems for months and nobody could easily find it, illustrating the core problem: fragmented tools with rich memory but poor recall.
They framed the opportunity around three shifts that now make internal agent systems buildable in weeks rather than months: models can now hold context over long time horizons rather than just a single prompt, systems have become connectable (CRMs, CMSs, and other tools can plug into models via things like MCP servers), and agents can run unattended on schedules without needing a human to trigger each action. From there, they introduced a four-layer blueprint, built bottom-up, where each layer is dependent on the one beneath it:
- Memory (bottom layer): centralizing CRM data, team chat, and published content so the system can answer “have we done this before” without a human needing to know where to look. They stressed that teams often skip chat, but chat is where reasoning and rationale actually live, not just outcomes.
- Access: putting the agent where the team already works (e.g., Scott living inside Slack) rather than requiring a new tool or tab. Their test: if anyone needs training to use it, the access layer is broken. They argued most so-called “AI adoption problems” are actually access problems in disguise, and fixing access (embedding Scott in Slack) dramatically increased usage across teams, including outside sales.
- Agents: the layer that actually performs repeatable tasks, classified into three types: scheduled/nightly agents, continuously running agents, and trigger-based agents (fired by a threshold or event). They proposed a three-part test for whether a task deserves an agent: does it repeat on a schedule (not one-off), is the input already available in the memory layer, and can a human still make the final call.
- Approval: ensuring agent outputs come with a visible reasoning trail so a human isn’t just judging the output but evaluating the process behind it, treating approval as a calculated decision rather than blind trust.
Key Takeaways
- Fragmentation, not lack of tools, is marketing’s core problem. Marketing has more tools and stored memory than ever, but that memory is scattered across disconnected systems, so the real challenge is making existing knowledge recallable and actionable, not buying more tools.
- Build bottom-up: memory, then access, then agents, then approval. Most teams start at the top (jumping straight to agents or flashy automation) and skip memory and access, which is why systems fail to get adopted or produce unreliable results.
- Most “adoption problems” are actually access problems. Placing the agent inside tools teams already use (like Slack) rather than requiring a new login or interface removes friction points that otherwise get misdiagnosed as AI or adoption failures.