Causal AI and governed action
Part 5 of the Marketer Graph series, the finale. Start with the overview, or read Part 4: trust, rails and verification first.
The structure, the rails and the verification add up to one thing: an agent that can reason from cause to effect, and then act on it. This is where the Marketer Graph is taking us, and where it still falls short.
A step towards causal AI - what should a marketer do next?
Step back from the mechanics and there is a larger reason we built it this way. A dashboard shows that spend went up and leads went up, and leaves the marketer to guess whether one caused the other or whether something else moved at the same time. The questions a marketer actually needs answered are causal, and they come in three kinds: what happened and why, what will happen if we change something, and what we should do next given what happened and what we think will happen next.
A graph is the natural home for those questions, because cause and effect are relationships, and relationships are what a graph stores. When the work a team did sits in the same structure as the outcomes it produced, with the links between them recorded, the agent can reason from a budget change to a move in CPL to a shift in the premium retained, instead of just noticing that two lines on a chart bent at the same time and then trying to investigate in context. The caveats and the confounders, the things that quietly break a naive read, are nodes in that same graph, so they are pulled into the reasoning rather than forgotten.
That is what lets Quinley work across all three questions on one substrate. It understands what happened, tracing a drop back through the chain of causes rather than stopping at the symptom. It anticipates what will happen, running those same relationships forward to show the likely impact of a change before it is committed to. And it decides what to do next, turning that into an action it can carry out and then carrying it through. Understand, predict, act, on the same graph.
Recorded relationships get a marketer a long way past a correlation on a chart, but an honest causal claim, that a change actually caused an outcome and was not merely followed by it, needs more than observation. For this we need experiments. This is also where marketing has struggled, because designing a clean holdout, tracking who was in and out, waiting out the lag from spend to bind, and reading the lift against everything else that moved is painstaking work that rarely gets done well. For ad platforms with billions of impressions it's feasible, for advertisers with limits this is hard to do.
We are not claiming to have solved causal inference. That is genuinely hard, and it is the frontier the market is working towards. What has changed is that the structures needed to do it properly are now in one place. The graph already records every decision, its scope, the time it happened and the outcomes that followed. The action layer can deliberately withhold a change from a slice of the market rather than rolling it everywhere. And the verification loop can measure the treated group against the held-out one and report the lift with the same discipline it grades any other number. Designed experiments, holdouts and incrementality testing are the next thing this foundation makes possible.
From answering questions to taking governed actions
Knowing why the leads were down is only half of it. The other half is the forward question, the one the marketer is really being asked: what is the best thing to do with the next dollar given what happened, and how do we get back to goal? Once an agent can see the full picture and trust its own numbers, it is a short step from explaining a problem to working on it, and that's what we're currently working on.
Taking decisions and actions on multi-million dollar ad budgets is incredibly complex. The governance, trust, review structures, sign-offs, partner relationships, data integrations and regulations make the programmatic execution, agents and graphs look straightforward.
The Marketer Graph is the reasoning layer underneath a set of marketing actions Quinley will carry out directly. It can move budget, shifting spend away from a fatiguing concept towards one that is converting based not on the ad platform but the full online/offline attribution graph, scoped to the right channel and the right ad account. It will build campaigns, taking a recommendation such as funding the strongest specialty insurance concept on Meta where it is currently underfunded and turning it into a campaign that is ready for review. And it can both make a recommendation and implement it once a marketer approves, which is the real change, because the old loop always ended at the insight and the next step was a person, a login and an afternoon.
Before it changes anything, it shows what the change is likely to do. Because the same definitions that diagnose the problem can be run forwards, Quinley can lay out the scenarios, how much to move and where, what it does to CPL and CAC, and what it is likely to mean for policies, retained premium and risk targets by the end of the quarter, so the marketer decides with the projected impact in front of them rather than on a hunch.
Even though this is, in a lot of cases, more evidence than a marketer would have gathered in the past before making changes, there is an exceptionally high bar to trust and implementation. For this reason, many data-driven AI marketing implementations stop short of this stage, because there is no accountable human to talk to about the changes the system will make, or, when things go wrong, the changes it made. Or the industry shows the complete opposite: AI tools that promise in flashy demos that a marketer can create and launch granular campaigns in ten minutes, except they lack all of the nuance, painstaking detail and considerations that actually drive performance for an advertiser, and they optimize on surface-level outcomes.
What makes actions driven by Quinley safe is that an action runs through the same vetted definitions the diagnosis came from. Increasing budget on the specialty insurance concept means the same concept, the same channel and the same numbers that were on screen a moment earlier. It is enforced programmatically in the graph, not in agent context. The recommendation and the implementation are not two systems joined together after the fact. They are the same graph, read forwards.
None of this happens without a person and a guardrail. An action is always proposed before it is taken, scoped to a permissioned account and a specific change, and held until someone approves it, and the larger the money in play the more that matters. Every proposal carries its projected impact, every approval is recorded, and every change the agent makes is logged back into the same graph, so there is a full trail of what was done, on what evidence, and what it was meant to achieve. Nothing runs blind and nothing runs unbounded, and anything that was done can be traced and, if it needs to be, walked back. For insurance, where a single budget move can swing millions of dollars of spend and the premium that rides on it, governance and guardrails need to be built in.
What this buys the marketer is not only speed. The spend becomes defensible, because every move is tied to the definitions and the evidence sitting behind it. The next dollar goes further, because it is pointed at the concept and the channel the graph can actually show are working. And the trust travels upward, because the marketer can explain the decision and the person they report to can believe the number underneath it.
Where we go from here
That, underneath a marketer's simple question, is the whole idea. A structured, private memory of how a business's marketing data actually fits together, of what causes what, searched by meaning, executed on rails, checked against its own warehouse, and getting a little sharper every week it runs. We built Quinley Base so that marketers could ask their data anything. We built the Marketer Graph so that the data could answer back, and then help do something about it.
From here we are forging ahead with two complementary steps, further developing this causal understanding in a variety of ways building on this foundation and help our agents take trusted governed actions based on the graph. Causality is a long, complicated road and it will take a significant amount of time to solve well, the amount of unseen data, the interaction of various upstream models and their drivers, confounding effects, biases and more will take a significant amount of work but we've made a significant step forward to unlock trusted insight and actions towards the agentic insurance acquisition future.
Frequently asked questions
Why is a graph the right home for causal questions?
Because cause and effect are relationships, and relationships are what a graph stores. When the work a team did sits in the same structure as the outcomes it produced, with the links between them recorded, the agent can reason from a budget change to a move in CPL to a shift in retained premium, instead of just noticing that two lines on a chart bent at the same time. The caveats and confounders that quietly break a naive read are nodes in that same graph, so they get pulled into the reasoning rather than forgotten.
Is Fount claiming to have solved causal inference?
No. Causal inference is genuinely hard and it is the frontier the market is working towards. What has changed is that the structures needed to do it properly are now in one place: the graph records every decision, its scope, the time it happened and the outcomes that followed; the action layer can withhold a change from a slice of the market rather than rolling it everywhere; and the verification loop can measure the treated group against the held-out one. Designed experiments, holdouts and incrementality testing are the next thing this foundation makes possible.
What does 'governed action' actually mean?
An action is always proposed before it is taken, scoped to a permissioned account and a specific change, and held until someone approves it. Every proposal carries its projected impact, every approval is recorded, and every change is logged back into the same graph, so there is a full trail of what was done, on what evidence, and what it was meant to achieve. Nothing runs blind or unbounded, and anything done can be traced and, if needed, walked back. The larger the money in play, the more that matters.
What can Quinley actually do, versus just recommend?
The Marketer Graph is the reasoning layer underneath a set of actions Quinley will carry out directly: moving budget away from a fatiguing concept towards one that is converting on the full online and offline attribution graph, building a campaign that is ready for review, and implementing a recommendation once a marketer approves it. Before it changes anything, it shows what the change is likely to do - the scenarios, the effect on CPL and CAC, and the likely impact on policies, retained premium and risk targets - because the same definitions that diagnose the problem can be run forwards.
What makes an agent taking budget actions safe in insurance?
An action runs through the same vetted definitions the diagnosis came from, enforced programmatically in the graph rather than in agent context, so the recommendation and the implementation are the same graph read forwards rather than two systems joined after the fact. Combined with a human approval step, scoped permissions, projected-impact on every proposal and a full audit trail, the spend becomes defensible: every move is tied to the evidence behind it, and the person a marketer reports to can believe the number underneath it.