← Research library

Attribution in an agentic funnel: credit decisions, not just channels

AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.

Marketing attribution has always been imperfect, and agents are now adding a new layer to an old problem. When agents research accounts, draft content, choose audiences, and time follow-ups, a single path to a closed deal can include dozens of automated decisions.

Channel reports were never built to see those decisions. This post describes how we think attribution should change when agents touch every step, and what a team can put in place before it needs a new model.

Why channel attribution falls short

Traditional attribution asks which channels and campaigns contributed to a conversion. First touch, last touch, and weighted multi-touch models all answer that question differently, and teams have argued about the weighting for years. The argument assumes a stable cast: a person designed the campaign, and the channel delivered it.

In an agentic funnel, the channel is often the least interesting part. The same email channel can carry a message whose audience, subject line, timing, and follow-up were each chosen by a different agent. If performance improves, the channel report credits email. It cannot tell you which of those choices made the difference, or whether a person or an agent made it.

That is why we suggest a different unit. A more useful question is which decisions contributed, and who or what made them.

From channels to decisions

A decision is any choice that changes what a customer or prospect sees. In a B2B program, typical agent decisions include which accounts enter a sequence, which message variant each contact receives, and when a follow-up goes out. Each one is a candidate for credit or blame.

Decisions are only measurable if they are recorded. An agent that picks a send time and moves on leaves no trace that attribution can use later. This makes decision logging the foundation of attribution in agentic systems. It also overlaps heavily with the run histories described in our post on observability for marketing agents, so many teams can extend logs they already keep.

What to log

For every agent action that touches a customer or prospect, we recommend recording a small, consistent set of fields.

  • Agent and version: which agent made the decision, and which version of its instructions and model it was running.
  • Inputs: the segment, brief, and data snapshot the agent used.
  • Decision: what it chose, such as the variant, the audience, or the timing.
  • Alternatives: the options it considered and rejected, when the agent exposes them.
  • Output link: a pointer to the resulting asset or action.
  • Join keys: the identifiers needed to connect the decision to later outcomes.

The join keys are where most logging efforts fail. If a decision log records an agent's choice but not the contact or account ID that the CRM uses, it cannot be joined to pipeline data later. Our post on identity resolution covers why consistent identifiers matter more once agents act across systems.

We also suggest writing decisions into the same timeline as human actions. A unified event stream lets you see that a person approved a segment, an agent wrote three variants, and a second agent scheduled the send, all in one sequence.

A worked example

Consider a hypothetical B2B software team running an account-based program for 300 target accounts. A research agent writes account briefs, a content agent drafts outreach, and a scheduling agent chooses send times. Over one quarter, the program produces 24 opportunities, up from 15 in the previous quarter.

The channel report credits outbound email for most of that gain. The decision log tells a richer story. Opportunities cluster in accounts where the research agent flagged a recent leadership change, and they are spread evenly across send times. That suggests the research agent's judgments mattered more than scheduling did.

The log does not prove that the research agent caused the gain. It tells you where to look next. A sensible next step is a controlled test: route some flagged accounts through the old manual research process and compare.

Comparing agent and human decisions

One practical use of decision logs is direct comparison. When agents and people make similar decisions, such as choosing subject lines or prioritizing accounts, you can compare their outcomes on comparable work.

Holdout designs make these comparisons more reliable. A share of decisions stays with the existing process, assigned at random, so differences in outcomes are less likely to reflect differences in the work itself. We suggest starting with a holdout of 10% to 20% for any decision you plan to report on. We cover design details in our post on incrementality testing for agent-driven campaigns.

Random assignment matters more than size. If reps hand their best accounts to the agent and keep the hard ones, the comparison will flatter the agent no matter how large the sample.

Avoid crediting the agent for the market

When results improve after introducing agents, it is tempting to give the agents full credit. Budget cycles, market shifts, product launches, pricing changes, and other campaigns all affect outcomes. Without a comparison group, an improvement may simply reflect a good quarter.

Before-and-after claims deserve particular care in reports to leadership. We recommend that any claim about agent impact name its comparison: a holdout, a matched period, or a prior baseline with known differences listed. If no comparison exists, say so plainly and present the result as a pattern worth testing.

There is a quieter failure too. Agents can move credit around without creating value. A follow-up agent that reaches every warm lead first may collect last-touch credit for deals that would have closed anyway. Watch for agents whose attributed share rises while total pipeline stays flat.

What to measure

Three measures give a useful starting picture of where agents help. Review them monthly, by agent and by decision type.

  • Decision coverage: the share of customer-facing agent actions with a complete log, including join keys. Aim for full coverage before trusting anything else.
  • Lift against holdout: the difference in outcome between agent decisions and the holdout process for the same decision type.
  • Credit shift: how attributed credit moves between agents, people, and channels over time, compared with movement in total pipeline.

If coverage is low, the other two are unreliable. Fix logging first.

A pragmatic approach

Most teams do not need a new attribution model on day one. We suggest three steps, in order.

  1. Add decision logs for every customer-touching agent action, with join keys to your CRM.
  2. Run a small number of controlled comparisons for the decisions you care about most.
  3. Enrich your existing attribution reporting with agent decision data rather than replacing it.

Over time, this builds a picture of where agents help, where they do not, and where they simply move credit around. It also gives finance and sales a shared record to argue from.

We are now studying how to attribute credit when several agents shape one decision together, such as a research agent's brief that changes a content agent's draft. Logging each step is straightforward. Deciding how much credit each step deserves is still an open question, and we expect it to need controlled tests rather than a formula.

How to apply this: make sure every customer-facing agent action is logged with the agent version, its inputs, and the CRM identifiers needed to join it to outcomes. Then pick one decision type and run a randomized holdout before you report on agent impact.

Related research

Put this research to work

Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.