• LinkedIn
• Twitter
• Instagram
• YouTube
• RSS Feed

Apple •Spotify• Pocket Casts •Youtube •Overcast •RSS

What’s up folks, welcome to our 4 part series of crawling through the dungeon of martech architecture. You’ve arrived at Part 4: The Dispatch Tower.

Summary: Every vendor in your stack just turned on an agent. Your ESP has one. Your CDP has one. Your MAP has one. None of them know what the others are doing, none of them were told to check, and nobody volunteered to referee it. Welcome to the final floor. The boss here is the agent avalanche, and the team that controls the dispatch layer controls what the rest of your stack is actually allowed to do.

In this Episode…

• What Happens When 30 Vendors Turn On AI at the Same Time
• Why Too Many Tools Break AI Agent Performance
• BOSS BATTLE: The Agent Avalanche
• Why You Need a Central Referee, With Agent Guardrails
• The Dispatch Layer that Controls Routing Logic
• The Layer Above the Dispatch is the Interface
• What Production AI Agents Look Like on a Governed Data Foundation
• The Deterministic Layer: Why the Interface Can’t Run on Guesses
• System Stewardship: The Role That Emerges When You Clear This Floor

What’s up folks, welcome to our 4 part series of Crawling through the dungeon of martech architecture. You’ve arrived at our final episode: Part 4: The Dispatch Tower.

Welcome back to the Dungeon of Martech Architecture.

You’ve arrived at the final floor. Parts 1 through 3 built the foundation, the meaning layer, and the causal memory. Part 4 is the floor most organizations have been living on the whole time, often without knowing it.

Episode 1: CRM Gravity. You’ll conquer the source of truth and discover that the data warehouse replaces the CRM with portable audiences.

Episode 2: The Eye of Context. You’ll learn why AI fails without shared meaning, why context engineering is the layer between data and agent authority, and why the industry built the wrong kind of meaning infrastructure in 2012.

Episode 3: The Correlation Masquerade. You’ll escape the correlation trap and build the causal memory layer that separates agents that optimize correctly from agents that confidently scale the wrong behavior.

Episode 4: The Dispatch Tower (you are here). You’ll tackle the governance chaos of 30 vendors all claiming authority, and confront the interface decision that most organizations already made without realizing it.

Let’s finish our descent.

FLOOR 4: THE DISPATCH TOWER

Congratulations martech crawler, you’ve arrived to the final floor of the dungeon. Let’s face it, most of your fellow crawlers will never make it this far. They’re still lost in political battles debating why the CRM shouldn’t hold product activation data or they ended up falling prey to the correlation trap and I forever stuck in the boomerang room. 

But you’ve made it. You’re proudly wearing the 3 badges of honor on your jacket: data janitor, context king and causal brainiac. 

This floor is the hardest because it combines collaborative political battles but some serious technical obstacles as well. The layout of this floor is best described as a hot mess of agent spaghetti. 

Every vendor in your stack has an agent now. Your ESP has one. Your CDP has one. Your MAP has one. Your CRM has one. Your ad platform has one. Each one has separate logic, separate assumptions, separate permissions, and a roadmap slide that says “autonomous.”

None of them knows what the others are doing.

And nobody volunteered to referee this stuff.

At the top of this floor, there is a door most organizations never noticed. Whoever controls that door controls what instructions get through, which systems can act, and which agent gets authority over the customer.

What Happens When 30 Vendors Turn On AI at the Same Time

Your inbox is probably already full of emails from every platform in your stack, all saying some version of “Turn on AI today.”

Each one promises autonomous intelligence. Each one wants permission to write, recommend, segment, suppress, score, enrich, personalize, or trigger something. Each one is operating from its own model of the customer. None of them can explain what happens when two agents try to update the same record, message the same person, or optimize against conflicting goals.

You are now the referee. Congratulations. 

Rich Waldron, CEO of Tray.ai, describes how this feels on the ground:

“If you’re the marketing ops leader and your 30 vendors are telling you to go turn on the AI for their application, you sort of become like an AI referee. You have to figure out, well, do I turn it on for this one and not for this one? And is this one gonna overwrite what occurs here? You feel it already. Your inbox floods with vendors begging you to ‘just turn on our AI capability’ — 30 different platforms all promising transformation. Suddenly you’re an unwilling AI referee asking impossible questions: which AI systems should I activate first? What happens when one AI’s decisions contradict another’s? Are all these systems feeding sensitive data into the same models?”

RICH WALDRON, Episode 162

Rich is describing the nightmare version of the final floor: every vendor ships an agent, and marketing ops gets stuck deciding which ones are safe, which ones are redundant, which ones are risky, and which ones quietly conflict with each other.

“Agents backed by an iPaaS give you full governance and control of where execution happens. For teams drowning in vendor AI chaos, the third option — iPaaS-backed agents that naturally integrate with everything — provides centralized governance, unified control points, and consistent execution environments.”

The answer is not to let every team turn on whatever they want. It is also not to create a central AI police force that blocks everything until innovation dies in a committee.

The middle path is an operating model: a small central group that sets standards, reviews risk, creates reusable patterns, and lets the business move quickly without turning the stack into a haunted appliance store

Funny enough, I spoke with someone who actually did volunteer for that AI referee job – her title is a bit fancier though. Here’s Lindsay Rothlisberger, Director of GTM Innovation at Zapier and also a member of their AI Center of Excellence. 

At Zapier, she sits at the spoke of a hub-and-spoke AI Center of Excellence, a small central team led by a Chief AI Officer, with Lindsay as the GTM representative responsible for translating governance into go-to-market practice. Her work covers deciding which AI tools get activated, setting the standards for what a shared skill has to pass before it enters the internal library, and keeping track of what people are building before it fragments into 100 unsanctioned experiments.

“We have a skill that reviews our skills for data and security compliance. It gives you a red, yellow, green. Red — here’s what you definitely need to fix. And in most cases, Claude can just say, ‘I can go fix that for you. Would you like me to do it?’ And just do it. So it’s a pretty simple process. But it is a standard review skill that runs through all of these different things.”

LINDSAY ROTHLISBERGER, Episode 223

Every shared skill at Zapier has to have a named owner. If it’s used across a team, someone is accountable for maintaining it. Skills covering go-to-market operations are owned by RevOps. New builders who want to ship a skill that runs through a cross-functional workflow pair with a RevOps partner to build it. The governance operates as a co-authorship model, designed to make sure what gets built can be maintained and trusted.

“We’re becoming more of enablers and guardrails and setting the systems and the infrastructure, but letting people who are really close to the problems figure out how to solve them.”

I’m betting that we’ll see a lot more roles like Lindsay’s pop up over the next few years… probably still not at the pace of new AI agents but hey… I can dream… It’s almost like every vendor wants to be the AI brain… but nobody wants to be the adult in the room.

Why Too Many Tools Break AI Agent Performance

The structural version of this problem is worse than the org chart makes it look. Enterprise AI deployments regularly expose 20 to 30 tools to a single agent simultaneously. Multiple sources say some version of  agent performance starts to degrade beyond 5 to 10 tools per agent (LangGraph, AWS’s Agentic AI Lens and Tianpan.co).

The failure is architectural. If a human cannot say definitively which tool handles a given task, the agent cannot either. 30 vendors each wanting you to turn on their AI creates an organizational headache and a configuration that breaks the reasoning layer before a single campaign runs.

Keith Jones has lived this from the operator’s chair:

“You’re playing translator between vendor A and vendor B and the agents or orchestration layers within them. You cannot automate what you do not understand. You cannot orchestrate across tools unless you know exactly how each one works under the hood.”

KEITH JONES, Episode 170

Olga Andrienko, who left a VP role at Semrush to build AI for marketing ops, adds the procurement reality that most architecture discussions politely sidestep:

“Large SaaS vendors will likely win orchestration in many cases because of security and procurement requirements. The alternative is building centralized AI systems on APIs so teams retain internal control.”

OLGA ANDRIENKO, Episode 186

Neither path is wrong. Both require an explicit decision. The teams that drift into vendor-led orchestration without thinking about it are the ones who end up most surprised when something breaks.

What the Internal-Build Path Actually Looks Like

Most architecture discussions describe the internal option in the abstract. Keith Jones runs GTM Systems at OpenAI and his team built and open-sourced their answer.

The foundation is harness engineering: building the environment a coding agent needs to operate within; the patterns in place, the standards it must follow, what it’s allowed to touch. At OpenAI, that harness is Salesforce-specific. Layered on top is Symphony, a generative code orchestration framework that connects a well-scoped ticket to agentic code deployment without requiring a developer to write every line.

The workflow: a TPM writes a ticket with enough specificity to tell the agent what changes and where. The ticket gets a label. Symphony picks it up, uses the harness to understand the environment, writes a PR, and passes it to an engineer for review.

Keith’s frame for how the roles work together:

“The entity, if you will, is an intern, a junior developer. What’s the best way to make a junior developer successful? Give them a senior developer to teach them. They are the ones looking at the PRs as they’re being drafted — ensuring the integrity of our systems and that things are being held to a certain standard in terms of code quality.”

KEITH JONES, Episode 224

Two human roles. An army of junior developer agents in between. When the engineer catches a mistake in a PR, they write a new skill: when you see this, do that instead. The agent doesn’t repeat that mistake. The skill accumulates. The next ticket that requires the same specificity benefits from it automatically.

The productivity impact Keith puts a number on:

“I’ve got contractors who within less than 30 days of being with the business are shipping the same changes as someone who’s been here for two years. That is an acceleration in productivity that has not been heard of in a technical landscape before.”

The harness makes agent authority specific rather than general. The orchestration layer converts a well-scoped ticket into deployed code. The engineer is the senior developer who makes the junior better over time. The model only works if DevOps discipline existed before anyone added the agent.

3 factors should drive that decision. 

1. Technical capacity: does your team have engineers who can build and maintain custom integrations? If not, vendor-led is probably your realistic path for now. 
2. Data sensitivity: how much of your customer data are you comfortable routing through a third-party model? The more sensitive the data, the stronger the case for API-based control. 
3. Stack stability: if you’re adding or replacing tools every 6 months, vendor-native integrations break constantly; an iPaaS or custom orchestration layer insulates you from that churn. 

Most teams will land somewhere in the middle: vendor-led for low-sensitivity, high-change workflows; API-based for anything touching sensitive data or requiring stable governance.

BOSS BATTLE: The Agent Avalanche

The boss on this floor is The Agent Avalanche: a torrential downpour of agents operating on default assumptions, the implicit belief that someone else made the governance decisions, that the vendor’s interface is just a UX choice, that coordination happens automatically once agents are running.

It shows up as the governance gap (the space between what agents can do and what they should do) and as the default interface (the vendor chat box that quietly became your dispatch layer without anyone deciding it). Defeating both requires making each decision deliberately, before the consequences make the decision for you.

*Party assembled. Equipment loaded. *

[LEGENDARY ARMOR] Agent Governance Framework: 3-piece set (centralized knowledge graph, decision hierarchy, unified customer context layer). Equipped together they cut conflicting agent actions by 87 percent. Missing any piece and the agents go back to contradicting each other.

[EPIC SPELL] Observability Stack: Continuous anomaly detection across every data source and agent output. Converts “we found out from the CEO” into “we found it before it shipped.” The difference between a quality control system and a hope system.

[RARE UPGRADE] Deterministic Layer: Handles dates, math, and field selection so the LLM never has to guess. Removes probabilistic risk from the parts of the stack that cannot tolerate error. The LLM interprets intent. Deterministic systems execute it.

[COMPANION UNLOCKED] The Dispatch Team: Data engineers, privacy specialists, and traffic cops who govern routing across the entire marketing org. The first team built specifically to manage agents, not run campaigns. Assign owners, run governance drills, control the interface before a vendor does it for you.

Let’s put these new items to work.

Why You Need a Central Referee, With Agent Guardrails

Guardrails are an organizational question before a technical one, and they belong in the architecture from the start. The right frame is what AI should do in each context, scoped before the agent runs.

That’s Tiankai Feng, Data & AI Strategy Director at Thoughtworks and Author of Humanizing Data Strategy.

“This is a service design problem. The question is not how many AI features you can launch. The question is whether the experience creates value and loyalty for the customer.”

TIANKAI FENG, Episode 179

The practical version of that principle is governance drills: running scenarios before something breaks in production. Chris O’Neill, who built the 2025 AI and Marketing Performance Index with GrowthLoop, borrows the concept from cybersecurity:

“You need to simulate things going wrong. Because something will go wrong. And in that moment, everyone should already know their role.”

CHRIS O’NEILL, Episode 177

Red team drills for AI deployments: simulate hallucinated product copy, misrouted personalization, sensitive data exposure. Walk the actual scenarios before they happen. Assign the owners now, not when the incident is live. Running the scenario is what prepares the team; the policy is reference material, not a rehearsal.

The teams that do this are distinguishable from the ones that don’t, specifically because governance drills force clarity about authority. Who can shut off the agent? Who gets paged when the targeting logic behaves unexpectedly? Who owns the dataset the agent read from? In 2026, every one of those is an operational requirement.

The 2026 State of Martech report found that marketers are pursuing an average of 70 distinct AI use cases across their organizations. That number sounds like progress. It also describes a governance problem. Most of those 70 use cases are being stood up faster than the foundational infrastructure that would make them safe to run. Data lineage, compliance, privacy, and consent are not advancing at the same rate as the more visible, more celebrated use cases at the top of the list. 

That’s not good. 

The glamorous work (AI-generated copy, autonomous audiences, predictive lead scoring) moves fast. The unglamorous work (knowing what data the agent used, whether it had permission to use it, and what it would do with a bad record) moves slowly, if it moves at all. Every AI initiative eventually runs into the same question: do we actually trust the data, context, and permissions this agent is acting on? The teams that answer that question before they need to are the ones who don’t find out the answer from a customer complaint.

Privacy Compliance and Data Minimization

Privacy and compliance are where this stops being an abstract governance conversation. If an agent can read customer data, generate campaign copy, enrich a profile, suppress an audience, or send instructions to a downstream platform, it is now operating inside the same risk surface as the rest of your martech stack. You need enforceable permissions, consent awareness, lineage, approval paths, and a clear answer to who owns the decision when the agent acts.

This is where privacy becomes architecture. Consent cannot live in a PDF, a checkbox, or a policy doc the agent never reads. It has to become operational context. 

Michele Nieberding made this point in our episode on customer data infrastructure and server-side data processing: privacy compliance depends on proactive consent management and responsible data collection, not cleanup after the fact.

“If you’re not enforcing that consent as soon as the data is collected, that’s a problem. Because then what happens is you’re collecting data, you’re not sure that those consent preferences are actually being enforced. So then your data goes downstream, and you’re like, oh crap, I’ve just identified a compliance risk. There’s a problem by the time you identify it. Like, it is a problem, you are in triage.”

MICHELE NIEBERDING, Episode 125

One thing worth mentioning here is the uncomfortable tension of context vs data minimization. AI systems get better when they have more context, but privacy-first marketing depends on restraint. The right context is not all the context. It is the minimum necessary context for the decision the agent is allowed to make. That means consent, data minimization, and privacy rules have to become part of the context layer itself, not a separate review step after the campaign is already built.

“I always say that the data protection bit needs to be intertwined. If you build your data strategy completely independent, they don’t work together, and then ultimately none of them will work towards the goal of the business. Don’t wait until you build your great strategy and then bring in legal or the compliance officer. It’s too late. Bring them in in the beginning. Just in case data is not always just in case data, and just in time data is usually too late. So there is a balance that needs to be found.”

SIOBHAN SOLBERG, Episode 131

Observability: Finding Problems Before Your Users Do

Even a well-governed and compliant stack breaks. The question is whether you find out before your users do.

Kevin Hu, CEO of Metaplane (acquired by Datadog) describes the design objective:

“The question is, do you want to know about it before your users do? And do you want to get better at it over time? The critical differentiator for any martech or data team is the ability to identify and address these issues proactively — before they impact the end-user experience.”

KEVIN HU, Episode 116

Observability means being the first to find the problem. That requires a quality control system, not optimism about having zero problems.

Data quality is a continuous monitoring problem, because every new source, every new transformation, every new tool in the stack opens another door. The teams that build observability as a system (not a one-time audit) are the ones who stop finding out about broken data from the CEO.

Elizabeth Dobbs built observability directly into the agent layer at Databricks, as a structural feature of how Marge operates, not an afterthought. Every answer surfaces the SQL it ran. There’s a feedback loop with thumbs up and thumbs down. The team reviews all of it weekly and traces bad answers back to their source:

“We don’t release her to the wild and say good luck everyone. She has a feedback loop — thumbs up, thumbs down. We do benchmarking, we look at all of the data on a very regular basis. If we have issues, we will reach out to the marketer and say: we can see your whole query. What felt wrong? So it’s not just an agent that’s out there by itself. It is constantly governed and reviewed.”

ELIZABETH DOBBS, Episode 219

An AI deployment with a quality control system built into it. The gap between those 2 approaches compounds fast when you’re running agents across millions of customer records.

The governance makes their velocity sustainable. The observability makes the governance honest.

But there’s still a question underneath both of those. Observability tells you when something breaks. The dispatch layer decides what was supposed to happen in the first place. And there’s a layer above both of those that most architecture discussions quietly skip.

Your buyers are already asking ChatGPT about your brand. Do you know what it’s saying?

Every day, ChatGPT, Claude, Gemini, Perplexity decide which brands get named, which pages get cited, and which products gets recommended—often before a buyer ever lands on your site.

Scrunch makes your brand the answer. We’ll show you how AI talks about your brand today, where the citation gaps are, and the page-level fixes that put you back in the answer.

Curious what it says about your brand?

→ See how AI reads my site

The Dispatch Layer that Controls Routing Logic

Most discussions of AI orchestration focus on what happens behind the interface: which agents are running, how they hand work to each other, what the coordination layer looks like.

Most orchestration discussions skip the question of who built the door.

Every marketer on your team opens some interface every morning. A chat box, a command bar, a conversational AI feature in one of their tools. They type what they want, and something happens. But between the question and the answer sits a routing agent (a master agent) that decides which tools, which sub-agents, and which workflows actually get invoked.

Whoever controls that routing agent controls what gets used downstream.

That’s Aboli Gangreddiwar, General Manager, Lifetime Value at Credible and someone who has been building agentic infrastructure for marketing operations, on what it takes to turn individual agents into something that actually runs end-to-end:

“If I am sending out an email campaign, I could have a copy agent, a Figma agent, and a coding agent. Right now, teams are building those individually, but at some point you need orchestration so they can pass work back and forth. Agentic infrastructure depends on layers that work together instead of one-off experiments. If your data is fragmented, agents will fail before they even start.”

ABOLI GANGREDDIWAR, Episode 191

Individual agents are proofs of concept; orchestrated agents that hand work to each other are the production system.

What Is the Dispatch Layer and Who Should Own It

Rebecca Corliss calls this team the dispatch layer, a new hub between marketing leadership and execution pods, staffed with data engineers, privacy specialists, and what she calls traffic cops:

“I think there’s gonna be a new team forming — a new part of this organization that’s gonna expand from marketing ops and MarTech. GrowthLoop is affectionately calling it the dispatch team. Imagine this new dispatch layer that is the group that’s thinking about the systems, the data, the AI, the architecture, and campaign activation for the entire marketing org holistically. The reason we call it dispatch is because this group is going to be accountable for how all marketing communications go out in a way that’s really effective for all of the marketing objectives being fulfilled.”

REBECCA CORLISS, Episode 188

The dispatch team is the right answer to orchestration. For most teams, it starts as a single person with a documented set of standing decisions: which agents are active, what the suppression rules are across overlapping campaigns, who gets paged when something breaks.

Whatever the title, someone needs to own the routing logic explicitly, maintain a written governance doc, and run a quarterly review of which agents are running and why. That’s the minimum viable version of this function. The team grows once the complexity demands it.

The Layer Above the Dispatch is the Interface

But there’s a layer above the dispatch team that most organizations haven’t thought about yet: the interface itself.

Florian Delval, who writes the Foreign Key newsletter on data infrastructure, wrote about this in a piece published April 2026. Behind every conversational interface, every chat box a marketer types into, sits a routing agent that decides what gets invoked. Whoever controls that agent controls what gets used downstream.

“If you don’t control the interface,” he writes, “it doesn’t matter how much orchestration you own. Users may never reach your agentic network. It’s like building the most advanced city in the world, but controlling none of the roads that lead to it.”

The grocery store analogy is instructive. The store is a commodity interface, unremarkable and interchangeable. But the store employees decide which products are visible, which ones get surfaced, which ones get picked. The interface is a commodity, but it controls the outcome. D2C brands spent the last decade trying to bypass the retailer specifically because of this dynamic. 

In the agentic marketing stack, the same logic applies: the vendor whose chat interface your marketers open every morning is, functionally, your dispatch layer. Whether you intended that or not.

“In theory, nobody should care too much about the interface. In practice, it determines control.”

FLORIAN DELVAL, Foreign Key

Lindsay Rothlisberger lives this from the inside. At Zapier, the AI Center of Excellence is functionally the routing layer for the organization, the team that decides what agent harnesses get provisioned, what context gets connected, and what governance applies before anyone sits down to use it. The marketers at Zapier open whatever interface the CoE set up. The routing decisions were already made upstream:

“At Zapier, we’ve got access to several agent harnesses that folks can decide based on preference. Some are living and breathing in Cursor or Claude Code. We leave that up to the end user. But what we did do is form an AI center of excellence. They set up the resources and the frameworks and the guidelines and the enablement. And I am a part of that AI center of excellence as a guild member dedicated to go-to-market. There’s the central team, and then I’m sort of the spoke out to go-to-market. The two of us work on making sure that those things translate into go-to-market — making sure we’re boots on the ground about how people are building and what they need, and being that feedback loop back.”

LINDSAY ROTHLISBERGER, Episode 223

Whether Cursor connects to Databricks MCP or Zapier MCP, what context loads by default, what rules govern safe use: all of it was decided before the end user sat down. The interface feels like autonomy. The defaults are architecture.

Teams that build their own conversational layer, like Elizabeth Dobbs did at Databricks with Marge, retain that routing control. Teams that default to a vendor’s AI interface hand it over. The orchestration you built is still there. Your marketers just may never reach it.

What Production AI Agents Look Like on a Governed Data Foundation

At Databricks, Elizabeth Dobbs and her team built 3 production agents sitting entirely on top of their marketing lakehouse:

• Marge handles conversational data access, trained on the specific vocabulary and definitions of every marketing discipline in the organization, with trusted answer pairs so any marketer can see the SQL behind a result and verify it.
• Tagatha automates content tagging continuously and retroactively, eliminating the taxonomy debt that accumulates every time a product or ICP changes.
• Atlas handles segmentation, combining rules-based and intent-based logic so audiences can be defined once and updated when new ones surface.

Every agent reads from the same governed source, with no copies and no sync jobs.

“The idea of the one chat: you have a chat interface very similar to ChatGPT. But on the backend we have all of our agents trained with their specializations. The marketer doesn’t know there are all of these agents they have to navigate between — we’re trying to obfuscate all the complexity — but if a marketer has a 101 question, Marge is the place. If they’re going to go deep, they end up in the right room with more narrow scope so she can be much more effective in her answers.”

ELIZABETH DOBBS, Episode 219

The routing, the specialization, and the orchestration are all absorbed by the architecture. The marketer just asks a question.

The 2026 State of Martech report describes the vendor stakes clearly. In the last era, martech platforms competed to be where marketers worked, the place people logged into every morning. In the next era, they will compete to be what agents can work with. The race has moved from desktop real estate to MCP integration, API surface area, and agent-readiness. 

Platforms that position themselves as substrates for agents, rather than destinations for humans, are making a different kind of bet about where value gets captured next. That is the SaaS-to-substrate shift, and it is already determining which vendors are building toward the center of your stack and which are quietly being routed around.

The Deterministic Layer: Why the Interface Can’t Run on Guesses

Controlling the interface is only half the problem. The other half is what happens when the interface actually tries to answer a question.

Ed Campbell, CEO of Bright Analytics, spent months building an MCP server for marketing analytics and documented every way it goes wrong. He wrote about it on LinkedIn here. The first instinct of pointing an LLM at your raw data and starting asking questions “kind of works. Enough to be exciting. Not enough to trust.” “Campaign” in Google Ads and “Campaign” in Meta Ads are separate concepts in separate tables.

The AI picks whichever spend field looks most relevant, won’t flag the ambiguity, and returns a number with complete confidence. That number may be wrong.

The underlying issue is structural. An LLM is probabilistic: it guesses, confidently and with reasonable calibration, but it guesses. Dates, math, and field selection cannot be guesses. “Last quarter” might mean a calendar quarter or a financial one. “Year to date” might start in April. A business whose working week runs Friday to Thursday has a “last week” that is nobody else’s last week.

Without a deterministic layer handling these questions, the analyst has to re-explain their calendar every single time. That is just a more expensive way to do the same manual work.

Campbell’s team built a semantic layer between the raw data and the MCP, a structured layer that encodes domain knowledge once, permanently. The LLM only ever sees the metrics your team has explicitly defined, tested, and trusts. Dates get a dedicated API that resolves against each customer’s specific calendar. Arithmetic gets pre-calculated in the data objects, because LLMs cannot do reliable math. The LLM interprets intent. Deterministic systems execute it.

“Neglect the architecture behind your MCP and you have a very expensive, very confident liability. Get it right and you will have something transformational.”

ED CAMPBELL, Data Roadies, February 2026

System Stewardship: The Role That Emerges When You Clear This Floor

Anna Aubuchon, VP of Operations at Civic Technologies, rebuilt her entire analytics stack around this principle, routing warehouse data into an LLM client so any stakeholder can ask questions in plain language and get answers in minutes instead of waiting for the next morning’s dashboard:

“We’re shifting from being execution-focused to now being very design-focused. You’re not driven by tasks, but you’re driven by: how do I architect solutions now? I can go layer by layer into the data, ask the exact questions I care about, and get proactive nudges like ‘have you considered this pattern?’ That work used to require multiple tools. Now it happens inside one conversation. Don’t be inhibited by what the AI tool can offer you — your imagination is the limit. In an industry that is moving at breakneck pace, vendor lock-in is a kiss of death when it comes to competitive agility.”

ANNA AUBUCHON, Episode 199

And Matthew Castino, Marketing Measurement Science Lead at Canva, describes what the data layer looks like when it’s mature enough to stop bottlenecking the people who need it:

“We’re working with Snowflake and the Cortex product to introduce natural language querying of the data warehouse for stakeholders. Data science is creating a semantic layer in our warehouse which has the core context that the model needs to answer questions in a reliable fashion. We’re working towards a world where there are low-hanging questions that probably exist in a dashboard somewhere that someone can’t find — trying to create this natural language interface for our stakeholders so they can answer questions quickly without requiring a data scientist, allowing a data scientist to focus on more complex work.”

MATTHEW CASTINO, Episode 200

The 2026 State of Martech report describes what is happening to the roles on every team that clears this floor. Campaign managers are becoming agent operators, then value engineers (people whose job shifts from executing campaigns to defining what good outcomes look like and ensuring agents pursue them). System administrators are becoming stack wranglers, then context engineers (people whose job shifts from maintaining tools to designing the shared context layer those tools and agents read from). 

The report uses the chrysalis metaphor: a butterfly is a different thing assembled from the same material, built during a period that looked, from the outside, like nothing was happening. Teams that use this transition to rebuild their job definitions around higher-leverage work will run the next era; teams that wait for it to settle first will find the roles already defined without them.

Elizabeth Dobbs, whose team runs Marge, Tagatha, and Atlas as production infrastructure, says:

“I don’t think agents are yet ready to be fully human out of the loop, fully prime time on their own. But in 12, 18 months they will absolutely be there. Our job as a marketing technology team is to place really strategic bets and build the foundation so we’re not the team that’s been waiting for it to be perfect and then ready to start.”

ELIZABETH DOBBS, Episode 219

Build toward human-out-of-loop from where you actually are.

Now: David Chan‘s sentence:

Alignment, not integration, will be the ultimate rate limiter.

DAVID CHAN, Mar 27, 2026

We spent 2 decades learning to connect systems to each other. The pipes are largely solved. The next 2 decades are the semantic design problem: shared definitions, causal foundations, institutional memory, and the governance that keeps humans in the loop without bottlenecking the agents that act on their behalf.

The question running through every conversation in this series is the same one. When every system has its own version of the customer (its own definitions, its own assumptions, its own local truth), how does a company decide what’s real?

Solving the alignment problem requires a clearer theory of what the data is actually for.

It looks like a team that can answer 3 questions:

1. What did the agents do last week?
2. Why did they do it?
3. What would change if we wanted them to do something different next week?

If your team can answer all 3, you’ve cleared the dungeon.

FINAL ACHIEVEMENT: The Dungeon Is Cleared

4 floors, 4 bosses, all cleared.

The False Truth King in the CRM. The Export Hydra spreading it everywhere. The Hallucination Oracle spewing believable nonsense from agents acting on undefined meaning. The Correlation Boomerang Archer scaling the wrong behavior at speed. And the Agent Avalanche operating with default assumptions, the quiet belief that someone else already made the governance decisions.

What you have on the other side is a foundation that can learn. Audiences that belong to no single tool. Agents that interpret data with the right context. A causal record that grows more reliable with every experiment you run. A dispatch layer where the governance decisions were made on purpose, by a person with a name and a doc.

The architecture you just built is the thing the next version of this problem runs on top of. The vendors will keep sending emails. The bosses will respawn in new forms. The teams that cleared these 4 floors will recognize the pattern next time it shows up, because they’ve already seen how each one hides.

Episode Recap

Every vendor in your stack has an agent now, each one operating from its own model of the customer, each one capable of messaging, suppressing, scoring, or updating records, none of them coordinating. The governance gap is the space between what agents can do and what they should do, and most stacks have populated the first without addressing the second.

The tactical fix has 3 components. A centralized knowledge graph gives all agents a shared, real-time understanding of the customer. A decision hierarchy determines which agent’s recommendation takes precedence when 2 conflict. A unified customer context layer keeps every system working from the same current data.

The dispatch layer is the organizational answer. Rebecca Corliss describes it as a new team between marketing leadership and execution pods, staffed with data engineers, privacy specialists, and traffic cops who govern routing across the entire marketing org. The minimum viable version is 1 person with a documented set of standing decisions: which agents are active, what the suppression rules are across overlapping campaigns, who gets paged when something breaks. That document is the dispatch layer. The team grows once the complexity demands it.

The interface question is the one most architecture discussions skip. Behind every chat box a marketer types into sits a routing agent that decides which tools, sub-agents, and workflows get invoked. The team that controls that routing agent controls what gets used downstream. Florian Delval puts it clearly: teams that control the interface control the routing, and the vendor whose chat interface your marketers open every morning holds that position by default unless you’ve made an explicit decision otherwise.

The bigger-picture implication is what David Chan described as the work of the next 2 decades: alignment, not integration, will be the rate limiter. The pipes are largely solved. What comes next is the semantic design problem, the shared definitions, causal foundations, institutional memory, and governance that keep humans in the loop without bottlenecking the agents acting on their behalf.

The teams that clear all 4 floors are the ones who can answer 3 questions: what did the agents do last week, why did they do it, and what would change if we wanted them to do something different next week. Listen to the full series starting at episode 1 with the fall of CRM gravity.

Thanks for descending with me and huge thanks for all of our guests through the past 20 months that joined me on the podcast!

Full episode ⬇️ or Back to the top ⬆️

✌️

—
Cover art created with Midjourney (check out how)

Apple •Spotify• Pocket Casts •Youtube •Overcast •RSS

Related tags

See all episodes

Future-proofing the humans behind the tech

Apple •Pocket Casts•Google •Overcast •Spotify •Breaker •Castro •RSS

• LinkedIn
• Twitter
• Instagram
• YouTube
• RSS Feed