AI agents promise to transform marketing operations. The reality is messier. Some teams ship autonomous workflows in weeks. Others spend months evaluating vendors and produce nothing. The difference comes down to foundations, feedback loops, and knowing which problems to hand to machines. We pulled from conversations with practitioners across 207 episodes of Humans of Martech, from OpenAI’s Head of GTM Systems Keith Jones to GrowthLoop CEO Chris O’Neill and operators building agentic infrastructure at Canva, Semrush, Tray.io, and Civic Technologies, to map what actually works, what breaks, and where the sharpest operators are placing their bets.
How to start with AI when your stack is already messy
Start with 1 repetitive task your team already avoids, clean the data fields it touches, and automate that before reaching for anything strategic. Agents amplify messy data rather than fixing it. Without consistent tagging, unambiguous field definitions, and clean syntax, agents will fail confidently at scale.
Every marketing ops leader has heard the pitch: deploy AI agents across your stack and watch efficiency compound. The pitch skips the part where your field definitions contradict each other across 3 systems, your data dictionary lives in someone’s head, and half your automations were built by people who left 2 years ago.
Keith Jones, now running GTM systems at OpenAI, puts it directly. His team used OpenAI’s own models to build their internal data dictionary, using AI to accelerate the standardization work that makes everything else possible.
“You do not get to skip the boring prep work. If your data model is a mess, agents will magnify it.”
Keith Jones, Ep 170
Olga Andrienko spent nearly 12 years at Semrush before leaving to build AI for marketing ops. She watched teams reach for strategic AI projects first and stall. Her advice runs in the opposite direction: start with the tasks your team actively avoids. Her influencer team automated manual tracking of brand mentions across 40 influencers and 7 competitors, a job that involved spreadsheet copy-pasting and basic math. First month result: 2 hours saved per team member, per month.
She built a risk and reward matrix to make these tradeoff decisions visible, ranking projects on 2 axes: implementation risk and realistic time savings. Color-coded by department. Green for go-ahead tasks, red for compliance-heavy work requiring legal preparation. The pattern holds across every team she advises: start with repetitive work, measure time saved in concrete hours, build reputation through wins nobody can argue with.
| AI project type | Risk | Time saved | Start here? |
|---|---|---|---|
| Manual data entry and copy-pasting | Low | 2-5 hrs/person/month | Yes |
| Report generation and formatting | Low | 30 hrs/month team-wide | Yes |
| Analytics queries from spreadsheets | Low | 5 hrs/week | Yes |
| Campaign QA and compliance checks | Medium | 3-8 hrs/campaign | After first wins |
| Content generation with brand voice | Medium | Varies | After guardrails |
| Autonomous campaign execution | High | Significant | After orchestration layer |
Aboli Gangreddiwar, Senior Director at Credible, frames the foundation differently. Effective AI agent systems need 5 interconnected layers: a unified data warehouse, orchestration across agents, execution in production environments, feedback loops so agents learn from outcomes, and human oversight protecting brand and compliance standards. Most teams skip straight to execution and wonder why nothing sticks.
“If it had the right metadata, the right dictionary, or if I had access to the documentation, I could have navigated it better and corrected the tables it was looking at.”
Aboli Gangreddiwar, Ep 191
Anthony Rotio at GrowthLoop traces this problem to organizational incentives. Quarterly reporting cycles and compressed executive tenures create pressure for visible short-term gains. Leaders want big bumps now, even when compound growth requires patience and reinforcement.
“When you think about compound interest in finance, the early part looks almost linear. People want big bumps now, even if those bumps never build momentum.”
Anthony Rotio, Ep 208
Documenting workflows before automating them
You cannot automate invisible labor. The fastest path to a working AI agent is a workflow you already run manually, documented step by step, with every decision point made explicit. Skip this and agents will automate your confusion at scale.
Keith Jones learned this when he unveiled a new lead score to skeptical sales reps. One rep pushed back, then described exactly 4 traits he already evaluated on every call. The score automated his instinct. Jones uses this as a design principle: document the unwritten rules your team follows, validate them against outcomes, then train agents to replicate that logic.
“Let’s replace something I already do, automatically or subconsciously, and do it better, at scale, with an agent.”
Keith Jones, Ep 170
Olga Andrienko takes the same principle and strips it down to daily motion. She approaches every task by asking whether AI can accelerate it, with no backlog and no sandbox testing phase. She uses AI for first drafts, decision support, comparisons, translations, summarization, and research. Building automation into daily motion rather than scheduling separate testing blocks creates compound benefits.
“There’s no backlog. I just have this task. How can I make it quicker?”
Olga Andrienko, Ep 186
Josh Hill focuses on the unsexy but critical layer beneath AI glamour. He advocates for identifying and optimizing key processes before layering intelligence on top, particularly the ad hoc nature of certain tasks like re-engaging lost deals. Impromptu approaches create inconsistent outcomes that agents will replicate faithfully. Standardize the process first, then automate it.
- Pick 1 frequent, consistent task your team repeats weekly.
- Write down every step, including the decisions people make without thinking about them.
- Identify the data sources each step touches and verify field definitions match across systems.
- Build the agent to replicate the documented process.
- Review the output against what a human would have done.
- Only then expand to the next workflow.
Keith Jones recommends starting with a “champion” approach: appoint 1 person to explore AI applications while others focus on day-to-day tasks. This dedicated role gradually builds interest across the organization, especially when others notice tangible improvements. Over time, curiosity sparks a ripple effect across departments.
“The failures in the beginning are smaller. The risk is lower. But the education is just as valuable.”
Keith Jones, Ep 170
Converting workflows into reusable agent skills
Agent skills are documented workflows converted into reusable capabilities that any agent can call on demand. Instead of building single-purpose automations, you build a library of skills that agents compose into complex sequences without human intervention.
Rich Waldron at Tray.io pushes the workflow-first approach further. His team converts visual workflows into agent “skills.” A workflow that enriches contact data through Clearbit, cross-references organizations against HubSpot records, and checks campaign enrollment status becomes a reusable capability any agent can call. When someone requests “set up our quarterly campaign,” the agent identifies which skills it needs and executes them in sequence without human intervention.
“Marketing ops professionals have the unique skill where they understand the business problem but also have the super skill of almost being an engineer. You can bring these ideas to life extremely quickly.”
Rich Waldron, Ep 162
Waldron described a customer who built a single workflow that identified site visitors by IP, enriched contacts, analyzed page context, and generated signals when patterns aligned. It accomplished what billion-dollar ABM platforms promise, built in hours. The agent applies reasoning to determine which skills to use when, turning pre-built workflows into a flexible, on-demand digital workforce.
Stephen Stouffer at Tray leverages a combination of APIs, OpenAI, Google, LinkedIn, and more, to build out complex automation processes. Each AI agent is designed to use the best tool for the task at hand. Given the right context and instructions, these agents gather relevant data from press releases, Crunchbase, LinkedIn profiles, and blog posts to craft personalized outreach that a human would take hours to produce.
- Identify your most repeated multi-step workflow (lead enrichment, campaign setup, QA checks).
- Break it into discrete skills, each handling 1 step with clear inputs and outputs.
- Build each skill as a standalone workflow that can be triggered independently.
- Connect skills through an orchestration layer so agents can compose them on demand.
- Test by asking the agent to handle a real request and compare output to manual execution.
Aboli Gangreddiwar sees orchestration as the next gap forming. Right now most teams build agents individually: a copy agent here, a Figma agent there, a coding agent somewhere else. The missing piece is coordination.
“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.”
Aboli Gangreddiwar, Ep 191
How AI agents differ from traditional automation
Traditional automation tools accelerate individual tasks and require manual input for each use. AI agents operate autonomously, initiate actions independently, and orchestrate end-to-end processes without constant human checkpoints. The distinction determines whether you are speeding up existing work or restructuring how work happens.
Olga Andrienko draws a clean line between the 2. She realized after a weekend course on autonomous systems that agent-level thinking was structurally different from prompt engineering. Tools help you move faster. Agents help you step away entirely. The structural difference changes what can be automated and who is needed to do it.
| Traditional automation | AI agents | |
|---|---|---|
| Input | Manual trigger for each use | Autonomous initiation |
| Logic | Fixed if-then rules | Reasoning over context |
| Scope | Single task acceleration | End-to-end process orchestration |
| Learning | Static until manually updated | Adapts from feedback loops |
| Human role | Operator running the tool | Designer setting guardrails |
Rich Waldron describes the shift as converting workflows from fixed trigger sequences into flexible capabilities where the agent decides which skill to use based on the prompt it receives. A traditional automation fires when a lead fills out a form. An agent evaluates the lead, selects the appropriate enrichment workflow, decides whether the lead qualifies for a sales touch or a nurture sequence, and executes the right path without asking.
“AI agents have leapt beyond basic LLM integration to become autonomous workers performing complex marketing tasks without human guidance.”
Rich Waldron, Ep 162
Wyatt Bales adds the nuance that matters for ops teams: automation handles the batch-and-blast tasks but does not eliminate the need for human involvement in whiteboarding and strategy planning. The power of automation extends beyond sending a batch of emails. It leverages higher-quality data to create micro-segments and tailor individualized emails based on factors like when a user created their account or initiated a free trial. This opens the door to comprehensive personalization, but only at the scale where the investment pays for itself.
“At the very basic, anybody should be able to feed your QA checklist, your campaign doc, your email proofs or SMS copy, and have it do that initial review.”
Aboli Gangreddiwar, Ep 191
How agentic AI changes marketing org design
Agentic AI shifts marketing organizations from channel-specialist structures to a 3-layer model: leadership with technical fluency, a dispatch layer managing systems and data architecture, and outcome-focused pods at the edge. Channel specialists evolve from tactical execution to journey orchestration.
Rebecca Corliss envisions a restructured marketing organization built on 3 layers. Leadership requires new technical fluency. CMOs must understand data systems, architecture, and AI operations. A new dispatch layer functions as a control tower for campaigns, staffed with data engineers, privacy specialists, and traffic cops who manage which campaigns reach customers when multiple business units compete for attention. Pods operate at the edge, each focused on a specific objective like repeat purchases or product recommendations.
“Imagine this new dispatch layer, the group that is thinking about the systems, the data, the AI, the architecture, and campaign activation for the entire marketing org holistically.”
Rebecca Corliss, Ep 188
This reshapes what it means to be a channel specialist. AI agents now handle mechanical optimization tasks like bidding, placements, and targeting. Rebecca argues the remaining value lies in deciding when a channel should activate for a specific person at a specific journey stage.
“The game is to decide Bill is going to get communication on LinkedIn only in these circumstances, only when he is in this cycle of his journey, only when he needs this offer.”
Rebecca Corliss, Ep 188
Chris O’Neill at GrowthLoop sees marketing teams resembling software teams: small, agile pods with shared ownership of experiments and outcomes. He pushes for AI skill ladders defining fluency expectations by role and level. A mid-level lifecycle marketer should test multiple GPT workflows, benchmark performance against campaign goals, and collaborate with data teams to refine prompts. That is the floor, not the ceiling. Someone must own the judgment layer. O’Neill sees a clear need for an internal AI regulator with authority to halt production and review output for tone, bias, brand safety, and legal risk.
Scott Brinker frames the human side of this shift. Organizations fund tools while skipping the slow work of converting new systems into living workflows. He regrets dismissing a course on organizational politics during MIT, preferring advanced technical subjects. He later recognized that understanding informal networks and decision patterns proved more useful than any system architecture. Change leadership beats technical skills when it comes to getting teams to actually adopt AI.
Olga Andrienko adds a sharper edge. She watched agents begin to eliminate entire teams. She contrasts Shopify’s internal AI memo positioning AI as employee partners against Duolingo’s memo about contractor replacement. Same action, opposite reception. Company framing matters. The organizations getting this right treat AI as a force multiplier for existing talent.
Replacing BI dashboards with AI-powered data access
Teams are replacing traditional BI tools with LLM clients that route warehouse data through natural language, combining sales data, product usage, web traffic, and CRM activity in a single conversation. Work that previously required multiple tools and a data engineer bottleneck now happens in minutes, but only with clean schema documentation and role-based guardrails.
Anna Aubuchon, VP of Operations at Civic Technologies, replaced her team’s central BI tool with an LLM client. GTM teams combine CRM, usage events, and website activity to ship campaigns grounded in actual user behavior. Product teams explore patterns separating casual from power users. Finance monitors AI spending in real time.
“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?”
Anna Aubuchon, Ep 199
The shift sounds liberating until you consider accuracy. Aubuchon watched models latch onto mismatched naming conventions, producing answers that felt precise but pointed wrong directions. One team called everything “users” while the warehouse called them “accounts.” The LLM followed warehouse terminology, and marketers thought an entirely new dataset had been invented.
“If you do not have good schema documentation, it will infer what it can out of whatever you have set up.”
Anna Aubuchon, Ep 199
Her rule: engineers and analysts write business-critical prompts because they understand warehouse mechanics and metric behavior deeply. Everyone else layers natural language exploration on top of those definitions. She uses chain-of-thought prompting that shows how the model reached its answer so teams can audit accuracy before acting on results.
Matthew Castino at Canva hit a similar bottleneck from the measurement side. His team had 15 people pinging 1 specialist to run geo experiments. That person was drowning. They built a natural language layer converting technical notebooks into accessible tools. Marketers now request tests in specific countries, define daily spend, and set duration. Snowflake’s Cortex layer handles routine data questions without burdening data scientists.
“We were getting fifteen people pinging one person to run an experiment. That person was drowning.”
Matthew Castino, Ep 200
Anna built role-based guardrails across every data source. Each MCP server gets its own rules: blocking PII at field levels, hiding email addresses for specific roles, limiting actions in Postgres without restricting Jira. Her team rejects nearly all outside AI services because vendors fail compliance filters. Building internally provided movement freedom while guardrails enabled trust.
- Document your schema in business language, not technical jargon. Define what “user,” “account,” “customer,” and “lead” mean in your warehouse.
- Have engineers write the foundational prompts that define metric calculations and table relationships.
- Layer natural language access on top for non-technical teams.
- Set up role-based guardrails per data source, not 1-size-fits-all permissions.
- Use chain-of-thought prompting so you can audit how the model reached its answer.
- Test naming convention alignment across teams before trusting AI-generated reports.
Build versus buy decisions for AI tools
Evaluate AI build vs buy decisions quarterly using 2 criteria: whether the capability touches a core competency deserving internal ownership, and the level of domain complexity involved. Build whenever the work influences differentiation, customer experience, or long-term strategy. Buy when capabilities are non-core and commoditized.
Anna Aubuchon recommends this quarterly cadence because the AI landscape moves too fast for annual vendor reviews. Modern integration layers like MCP, n8n, and direct LLM access have compressed the technical overhead that previously justified expensive vendor solutions.
“A one year agreement might as well be a decade in AI right now.”
Anna Aubuchon, Ep 199
Her team at Civic replaced expensive BPO level-1 support with internal AI, then took ownership of the intelligence layer. Operators controlled knowledge quality, reviewed patterns directly, and kept insights in-house. The economics shift when you own your intelligence: vendors refine systems from thousands of user interactions, but customers keep only what they supply. In-house models allow retraining when customer behavior shifts.
| Build when… | Buy when… |
|---|---|
| The workflow influences differentiation or customer experience | The capability is commoditized and non-core |
| You need retraining as customer behavior shifts | Implementation time exceeds internal capacity |
| Vendor tools fail your compliance filters | The tool has strong integration with your existing stack |
| You want to own the intelligence layer long-term | Rapid deployment matters more than customization |
Olga Andrienko bypasses procurement delays entirely by prioritizing tools already approved in her organization’s stack. At Semrush running G Suite, any Google feature like Gemini or Vertex AI required zero procurement. She saved $5,000 on a single project by building in Gemini instead of purchasing a new vendor tool. For public-facing workflows, she engages freelancers and agencies using their own AI tools, securing IP ownership through contracts with buyout clauses.
“Do not be inhibited by what the AI tool can offer you. Your imagination is the limit.”
Anna Aubuchon, Ep 199
Chris O’Neill dismisses tool maturity as an excuse and filters vendors by build velocity, iteration speed, and transparency about actual AI capabilities. He critiques legacy players for iterating at a glacial pace while smaller vendors ship weekly.
Why feedback loops determine whether AI helps or multiplies noise
AI amplifies whatever feedback loop it sits inside. When feedback is loose or implied, AI accelerates activity without direction, creating volume without progress. When feedback is tight, AI compounds learning in ways humans cannot match at scale. The distinction comes down to causality data and measurable definitions of done.
Anthony Rotio sees AI as an amplifier. Each customer interaction becomes a structured snapshot capturing customer state, delivered message, and measured outcome simultaneously. Over time, these snapshots form a record that supports counterfactual reasoning, where systems estimate what different messages would have produced under identical conditions.
“You need tight definitions of done and tests that actually matter. Agents can wander all night if the system knows how to evaluate the result.”
Anthony Rotio, Ep 208
Anthony’s concept of agent context graphs preserves experimental logic after experiments ship. Conditions making experiments succeed rarely remain static. Customer behavior shifts, competition increases, product changes accumulate. Instead of archiving what worked, the system continues running simulated cohorts from causal data, monitoring when response patterns drift. Teams observe assumption decay weekly and adjust before relevance erodes.
Chris O’Neill at GrowthLoop applies reinforcement learning to the same problem. Most AI-generated outbound sequences sound identical because teams generate content without measuring what actually works. Reinforcement learning changes the loop: it learns by doing, tracking real-world feedback and updating logic in motion.
“It would be silly to ignore all the data from previous experiments. Reinforcement learning gives us a way to build on it without starting over each time.”
Chris O’Neill, Ep 177
Matthew Castino at Canva brings the measurement discipline. His team treats experiments as correction mechanisms that push models toward observable reality. MMM models often rely on hundreds of daily observations while estimating hundreds of parameters, creating unstable signals. Experiments anchor the models to ground truth. But experiment results expire. Tests from 8 months ago no longer reflect current conditions because creative changes, budget pacing shifts, and competitive dynamics evolve constantly.
“Experiments are core. The model is the best possible thing when you do not have an experiment, but you need both feeding each other.”
Matthew Castino, Ep 200
Matt’s team uses modeling uncertainty to flag priorities, designs experiments isolating those gaps, then feeds learnings back before environmental changes render results obsolete. The feedback loop closes only when measurement, experimentation, and modeling all inform each other continuously.
Keeping AI outputs accurate and on-brand
Accuracy requires 3 habits: treat schema documentation as living artifacts that evolve with your data, use chain-of-thought prompting that shows how the model reached its answer, and review every output that influences downstream work. Without these, AI produces confident-sounding answers that quietly point teams in wrong directions.
Anna Aubuchon learned this through a naming convention mismatch. One team called records “users” while the warehouse called them “accounts.” The LLM followed warehouse terminology, producing reports that looked precise but confused marketers. Her fix: engineers define the foundational prompts, business teams explore on top.
Chris O’Neill recommends red team drills simulating hallucinations, regulatory breaches, and data leakage before agents enter live environments. Everyone should already know their role when something goes wrong.
“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, Ep 177
Aboli Gangreddiwar adds that human oversight remains essential for brand standards, compliance, and legal requirements. Campaign QA agents can automate the initial review: UTM parameters, spacing, brand alignment, rendering issues, and compliance screening. But the final call stays with humans.
“With compliance, it is either yes, no, or let’s evaluate the risk. AI can handle the yes and no cases, and that alone saves huge time.”
Aboli Gangreddiwar, Ep 191
Olga Andrienko built an in-document feedback layer where AI scores drafts using past campaign performance and hand-labeled examples of strong and weak copy. It flags weak headlines, vague CTAs, and bloated structure. She tracks editor edits to gradually improve generation quality over time. The system trains writers rather than replacing them.
- Document your schema in business language and keep it updated as tables change.
- Use chain-of-thought prompting so the model shows its reasoning before you trust the output.
- Run red team drills simulating hallucinations, data exposure, and compliance failures.
- Automate initial QA (UTM checks, brand alignment, rendering) with agents, but keep humans on final approval.
- Track every edit humans make to AI-generated content. Feed corrections back to improve future outputs.
- Appoint an internal AI regulator with authority to halt production when outputs fail quality, brand, or legal standards.
What skills marketing ops professionals need for AI
The skill requirement has shifted from “can you code” to “can you think systematically about information flow.” Operators need to understand data structure, model behavior basics, integration patterns, and prompt engineering fundamentals. Building AI skill ladders by role and level ensures teams know the floor, not just the ceiling.
Mauro Figueiredo highlights the unique position of marketing ops professionals in bridging technology and business. They possess deep understanding of both the technical systems and the strategic business processes necessary to leverage AI effectively. This dual focus makes them ideally suited to lead AI adoption within organizations.
Chris O’Neill pushes for defined expectations. A mid-level lifecycle marketer should test multiple GPT workflows, benchmark performance against campaign goals, and collaborate with data teams to refine prompts. That is the baseline, not the aspiration.
“Being able to articulate with precision how AI is going to drive and improve your profit and loss statement, that’s where it starts.”
Chris O’Neill, Ep 177
Scott Brinker values the T-shaped model: 1 specialty where you develop deep judgment, plus adjacent domains for collaboration literacy. He frames the future as human orchestration of AI capabilities rather than humans being replaced by them. The marketers who stay relevant are the ones who learn to direct AI systems rather than compete with them on output volume.
Anna Aubuchon sees natural language AI as removing gatekeeping entirely. Anyone who can articulate a workflow, describe an experience, and explain a purpose has the foundation to build.
“Think about all the cool things you want to do, then use your words and start explaining it.”
Anna Aubuchon, Ep 199
Rebecca Corliss pushes further, arguing that compute will become a budget line equal to media spend. CMOs need to learn what drives those costs and whether returns justify expenses. Ops teams need to measure AI ROI in language CFOs understand: concrete hours saved, cost per automated process, revenue influenced by AI-assisted campaigns.
“Compute is your agent’s salary. Are we actually seeing the ROI of that effort?”
Rebecca Corliss, Ep 188
Frequently asked questions
Where should a marketing ops team start with AI?
Start with 1 repetitive task your team already avoids, like manual data entry, report formatting, or spreadsheet-based analytics queries. Olga Andrienko’s team at Semrush began with influencer mention tracking (2 hours saved per person per month) and scaled from there. Measure time saved in concrete hours. Stephen Stouffer recommends appointing 1 “champion” to explore AI applications while others handle day-to-day work.
Should marketing ops teams build or buy AI tools?
Anna Aubuchon recommends evaluating quarterly using 2 criteria: whether the capability touches a core competency deserving internal ownership, and the level of domain complexity involved. Build whenever the work influences differentiation, customer experience, or long-term strategy. Buy when capabilities are non-core and commoditized. As Aubuchon puts it, “A 1 year agreement might as well be a decade in AI right now.”
How do I get AI tools approved by legal and procurement?
Olga Andrienko bypasses procurement delays by prioritizing tools already approved in her organization’s stack. At Semrush running G Suite, any Google feature like Gemini or Vertex AI required zero procurement. She saved $5,000 on a single project by building in Gemini instead of purchasing a new vendor tool. For public-facing workflows, she engages freelancers using their own AI tools, securing IP ownership through contracts with buyout clauses.
What skills do marketing ops professionals need for AI?
The requirement has shifted from coding to systematic thinking about information flow. Anna Aubuchon expects operators to understand data structure, model behavior basics, integration patterns, and prompt engineering fundamentals. Scott Brinker values the T-shaped model: 1 specialty plus adjacent domains for collaboration literacy. Chris O’Neill recommends building AI skill ladders that define fluency expectations by role and level.
Will AI agents replace marketing ops jobs?
The role changes shape, not size. Wyatt Bales argues automation handles batch-and-blast tasks but does not eliminate human involvement in strategy and planning. Rebecca Corliss sees ops evolving from maintenance staff into a growth-driving function with real accountability, building infrastructure that business units depend on to hit targets. Operators who treat their work as system design rather than ticket management become more valuable.
What is the difference between AI agents and regular automation?
Traditional tools require manual input for each use and accelerate individual work. Agent systems operate autonomously, initiate actions, and orchestrate end-to-end processes without constant human checkpoints. Rich Waldron describes this as converting workflows into agent “skills” where the agent decides which capability to use based on the prompt it receives, rather than following a fixed trigger sequence.
How do I measure ROI on AI investments in marketing ops?
Olga Andrienko measures time saved in concrete hours per team member per month. Her reporting workflow saves 30 hours monthly across her team. Chris O’Neill insists on precision: articulate exactly how AI drives improvement in your profit and loss statement using language your CFO would understand. Rebecca Corliss pushes further, arguing that compute will become a budget line equal to media spend, and CMOs need to learn what drives those costs.
How do I keep AI outputs accurate and on-brand?
Anna Aubuchon recommends 3 habits: treat schema documentation as living artifacts, use chain-of-thought prompting that shows how the model reached its answer, and review outputs influencing downstream work. Chris O’Neill recommends red team drills simulating hallucinations, regulatory breaches, and data leakage before agents enter live environments. Aboli Gangreddiwar adds that human oversight remains essential for brand standards, compliance, and legal requirements.
How do I handle data quality issues before deploying AI agents?
Keith Jones recommends building your data dictionary first, using AI itself to accelerate the standardization work. Clean field definitions, consistent tagging, and unambiguous syntax across systems. Aboli Gangreddiwar adds that self-healing data agents can detect anomalies and resolve conflicts in real time, but only when metadata quality supports them. Without a clean foundation, agents amplify existing problems at scale.
Can AI agents handle campaign QA and compliance checks?
Yes, for the initial review. Aboli Gangreddiwar recommends feeding your QA checklist, campaign doc, email proofs, or SMS copy to an agent for automated screening of UTM parameters, spacing, brand alignment, rendering issues, and compliance. For compliance specifically, “AI can handle the yes and no cases, and that alone saves huge time.” The final approval stays with humans for edge cases and risk evaluation.
What is reinforcement learning and how does it apply to marketing?
Reinforcement learning is a method where AI learns by doing, tracking real-world feedback and updating its logic in motion rather than relying on static rules. Chris O’Neill at GrowthLoop applies it to campaign optimization: instead of generating content without measuring what works, reinforcement learning adjusts timing, sequencing, and delivery based on live behavior signals. It builds on data from previous experiments rather than starting over each time.
What is the dispatch layer model for marketing organizations?
Rebecca Corliss proposes restructuring marketing organizations into 3 layers: leadership (CMOs with technical fluency in data systems and AI), a dispatch layer (data engineers, privacy specialists, and traffic cops managing systems, data, AI architecture, and campaign activation holistically), and pods (teams focused on specific business outcomes like retention or product recommendations). The dispatch layer acts as a control tower preventing campaigns from competing for the same audience.
Explore the episodes
- 208: Anthony Rotio: Exploring causal context graphs and retail media — How feedback loops, causal data, and agent context graphs determine whether AI compounds learning or multiplies noise
- 199: Anna Aubuchon: Using AI to build what you used to buy — Replacing enterprise BI with LLMs, the build vs buy framework for AI, and role-based guardrails across MCP servers
- 200: Matthew Castino: How Canva measures marketing — Natural language layer for geo experiments, MMM scenario tools, and feedback loops between models and experiments
- 191: Aboli Gangreddiwar: AI agents in marketing operations — The 5-layer agentic infrastructure, self-healing data agents, campaign QA agents, and compliance automation
- 188: Rebecca Corliss: The future agentic marketing org — The dispatch layer model, lifecycle marketers at the center, and the shift from channel specialist to journey orchestrator
- 186: Olga Andrienko: Ex-VP at Semrush left brand to build AI for mops — Risk/reward matrix for AI projects, custom GPTs for daily work, procurement bypass strategy, and building an AI-first reflex
- 177: Chris O’Neill: AI agent swarms and reinforcement learning — Reinforcement learning for campaign optimization, AI skill ladders, red team drills, and compound marketing
- 170: Keith Jones: OpenAI’s Head of GTM systems on buying martech — Data dictionaries before agents, documenting invisible labor, lead scoring from sales rep instinct
- 162: Rich Waldron: How to build and manage AI agents — Converting visual workflows into agent skills, the 3-level intelligence layer, and why marketing ops professionals are uniquely positioned
- 201: Scott Brinker: If he reset his career today — Change leadership over technical skills, T-shaped specialization, and understanding organizational politics
- 148: Stephen Stouffer: AI’s role in customer journeys and messaging — AI agents for personalized outreach, smarter content selection, and the champion approach to AI adoption
- 80: Wyatt Bales: Redefining marketing with AI and automation — Why automation handles batch-and-blast but strategy stays human, and the potential of micro-segmentation at scale
Want more practitioner perspectives? Subscribe to Humans of Martech for weekly conversations with the operators building the future of marketing technology.