Our mission at Conversion is to remove the ceiling on what a marketing team can do. We build one context layer over CRM, warehouse and product data, and agents that plan, build and ship the work on top of it, from segmentation and outbound through campaign execution and reporting.

Coding, support and legal agents are already maturing and the productivity gains there are no longer in dispute. We believe marketing agents will do the same, with one difference. Most agent categories defend the bottom line. Marketing is one of the few that moves the top line directly, which is why the bar for getting it right is higher and the payoff is larger.

The goal of our research program is to push the frontier on autonomous marketing execution. Model intelligence is not what holds this back. Marketing work lives across systems built for reporting rather than reasoning, so most teams point frontier models at a substrate that cannot answer basic questions about their own customers.

The beliefs underneath the work:

1. 01 The leverage of a marketing team is rising fast. The best product has never won by default. What decides outcomes is whether a company can reach the specific person with the matching problem, and that has always been bounded by how much one marketer could personally carry. Every platform shift so far has added work without adding hours. That bound is what is finally moving, and Conversion exists to move it further.
2. 02 Marketing is far from solved. It asks for creative judgment and mechanical rigor at once, and neither the problem nor the answer is ever fully specified. That makes the work harder to train against than code or support, where correctness is usually knowable. Nothing here is scored by a test suite, and whether the segment was right or the copy was good gets judged weeks later through an attribution model nobody fully agrees on. The human stays in the loop longer than in other mediums, not as a safety net but because taste is part of the work.
3. 03 The hardest problems live in the enterprise. No two companies model leads, accounts and opportunities the same way, and admin decisions made years ago define what is possible today. An agent that assumes a standard object graph is wrong at every account worth having. Building here also means revenue is not the only measure of a good decision, because explainability, auditability and permissioning become requirements rather than features.
4. 04 Frontier intelligence needs a capable harness. Model intelligence does not survive a bad substrate. Marketing work runs on vast integrations that have to execute exactly, and every action is a one way door. Code reverts and a support reply can be apologised for, but a send lands in every inbox on the list at once, cannot be recalled, and spends domain reputation that took years to build.
5. 05 Data is paramount. It is how agents improve, and how they gather the context a task actually requires. It is also the hardest part of the job. What a good decision needs is scattered across systems that were never designed to be joined, and the marketer is usually the one person in the building who cannot join any of it without filing a ticket. Conversion's data platform exists to close that gap.

Innovation here means rebuilding how humans and agents work together across all three surfaces at once: the data platform, the agent harness, and the interface a marketer actually uses. Whoever directs an agent has to be able to see what it decided, understand why, and stop it, without reading a query. We are working on all three, against real customer orgs.

We write up what we learn here, including the experiments that went nowhere.