Why Your AI Agents Are Stuck in Your Legacy Martech Stack

D

Deploying an AI agent into your current martech stack is like wiring a Tesla powertrain into a 2009 Accord. The car still drives like a 2009 Accord.

This week, OpenAI cofounder Greg Brockman said the quiet part out loud: ChatGPT's plugins, heavily marketed in 2023, failed "because the models weren't ready." He's partially right. But the bigger reason plugins flopped wasn't the model. It was the architecture underneath. The plugins were trying to extend software workflows designed for humans into something that required machine-to-machine coordination. Same mistake most marketing teams are making right now with agents.

Most operators have already heard the pitch for agentic AI. Some have installed agents. A few have connected them to HubSpot, or piped them into Slack, or pointed them at a Google Sheet. And most are quietly disappointed with the results. Not because the agent is bad. Because the workflow it lives inside was never built for it.

Legacy martech stacks were designed around a specific assumption: a human is always in the loop, making judgment calls between steps. A rep pulls a lead from Salesforce. A strategist reads the report in GA4. A media buyer checks Advantage+ performance and adjusts the budget. Every tool was built to present information to a person, wait for that person to decide, and then accept an instruction.

That architecture works fine for humans. It's catastrophic for agents. Agents don't wait. They plan a sequence, execute steps, check outputs, and loop. When they hit a system that requires a human authentication step, a manual export, or a platform that doesn't expose a proper API, they stall. Every handoff point designed for a human becomes a chokepoint for a machine.

The result is an agent that technically runs but practically crawls. It completes 40% of its intended workflow, then waits for a human to unstick it. That's not operator AI. That's a very expensive Zapier.

There are three specific layers where legacy martech kills agent performance. Most teams only diagnose one of them.

• Data access. The agent needs a clean, structured data source it can read without a human pulling a CSV first. Most CRMs and ad platforms were built for dashboards, not programmatic reads. HubSpot has an API; most teams have never used it. GA4 exports are still manually triggered on most teams we audit.
• Decision logic. Human workflows encode decisions inside people's heads, not inside the system. The agent has no access to the rule: "if ROAS drops below 2.8 for three consecutive days, pause the campaign and alert the buyer." That rule lives in a senior media buyer's muscle memory. It was never written down, never systematized.
• Action execution. Even when an agent correctly decides what to do, most platforms require a human session to execute. Performance Max budget changes via API are possible but require proper OAuth setup most teams skipped when they onboarded. The agent reasons correctly and then hits a locked door.

This is why the model is only 10% of what makes an agent work. Shopping for a better model when your data access is broken is like upgrading the engine in a car with no fuel line.

Rearchitecting for agentic AI is not a technology project. It's a workflow audit with a specific question: where does a human currently touch data between two systems? Every one of those touches is a potential agent insertion point. Most of them are also the slowest parts of your marketing operation.

A concrete example. The marketing director at a 6-location med spa pulls weekly performance data from Meta Ads, drops it into a Google Sheet, formats it, writes a summary, and sends it to the owner. That process takes 90 minutes every Monday. An agent with proper API access to Meta, write access to Google Sheets, and a structured output template runs that in 4 minutes with no human in the loop. But that only works if the API credentials are set up, the Sheet has a stable schema, and someone defined what "summary" means in machine-readable terms.

That's the work most teams skip. They install Claude or a ChatGPT wrapper, point it at a task, and wonder why the output is inconsistent. The inconsistency isn't the model. It's the undefined inputs and the unstable schema underneath.

“The bottleneck is never the model. It's the workflow the model was asked to operate inside.”

Brockman's comment this week about an "almost no interface" future is directionally correct. But the path there runs through boring infrastructure work: clean APIs, defined schemas, machine-readable decision rules. The invisible agent he's imagining requires visible, well-structured plumbing underneath.

There's a second architectural risk worth naming this week. Mistral CEO Arthur Mensch warned publicly that proprietary AI model providers are accumulating data on your business processes. His claim is pointed: labs observe how you use their tools, and in some cases have used those observations competitively.

Whether or not you take Mensch's warning at face value, the underlying concern is real. If your agent workflow runs entirely inside a single vendor's ecosystem, that vendor sees your decision logic, your customer data patterns, and your operational priorities. That's not a reason to panic. It is a reason to think carefully about what data your agents are passing to which models, and to design your architecture with some intentional separation between layers.

The teams building durable agent systems treat the model layer as interchangeable. Claude today, Gemini tomorrow, a fine-tuned local model for specific tasks next year. The harness, meaning the orchestration logic, data connections, and decision rules, lives outside any single vendor. That's the operator AI architecture that compounds over time.

If your agents are underperforming, run this audit before you touch the model.

• 01List every recurring marketing task that involves moving data between two systems. Every one.
• 02For each task, identify whether a machine-readable API exists on both ends. Not whether an API exists. Whether your team has actually connected it.
• 03Identify the decision rule that governs what happens next. If you can't write it in plain English in under two sentences, the rule isn't systematized and the agent can't use it.
• 04Map where human authentication, manual exports, or one-off judgment calls currently block automation. These are your chokepoints.
• 05Prioritize the two or three tasks where fixing the plumbing would eliminate the most human time. Start there.

This audit is not glamorous. It doesn't involve a new model or a better prompt. It involves reading API documentation, talking to the person who currently does the manual task, and writing down rules that have never been written down before. Most teams skip it because it feels like IT work. It's actually the highest-leverage marketing work you can do in 2026.

Tools like n8n and Zapier handle simple connective tissue. Cursor and Claude Code handle the more complex orchestration logic where a sequence of decisions needs to be programmed, not just routed. The Marketing AI Institute's SmarterX case this week demonstrated exactly this: Claude Code, typically treated as a developer tool, handled one of the most common marketing tasks, making sense of messy data, without a single line of traditional code written by a developer. The bottleneck wasn't the model. It was that nobody had thought to point it at the data.

The teams that win with agentic AI in the next 18 months won't be the ones with the best model. They'll be the ones who did the unglamorous work of mapping their workflows, systematizing their decision logic, and building an architecture the agent can actually operate inside.

Bolt-on AI is the self-checkout at Walgreens. It looks like automation. It still requires a human to supervise every step. Operator AI is Amazon's warehouse fulfillment system: the human designed it, the machine runs it, and the throughput is structurally impossible to match manually.

If your current agents feel more like the former, the fix isn't a better model. It's building the harness they actually need to run. That work starts with your workflows, not your model selection.