L
Lomit Patel
Guest
Will AI Fix B2B Attribution?
Not on its own. An AI agent answering questions from your CRM can only be as complete as the record underneath it. If calls, email threads, and relationships never made it into the CRM, the agent can't credit them. And if the record doesn't separate what was logged from what was guessed, the agent will present both as fact. Fix the record first.
Here is a deal as most CRMs see it.
A webinar. A discovery call. A proposal.
Three touches. Clean, easy to report, and missing part of the story. The peer who recommended you, the deck the champion forwarded to the CFO, the call where the CFO showed up: none of that is a touchpoint, so none of it gets credit.
Image 1: "Same deal. Two records." diagram
For years that was a hunch every B2B marketer shared and nobody could measure. Now there's data on how big the gap is.
What 521 deals show
Mada Seghete, co-founder of the revenue data company Upside and previously a co-founder of Branch, has published a study called "The Untracked Half". Upside ran 521 new-business deals at 13 B2B companies through its AI attribution model, with at least three independent AI analyses per deal reading every email, call transcript, note, and record behind it.
The headline finding: only 51% of the credit came from touches a system had recorded. The other 49% came from unstructured data. Twenty-five percent was extracted, meaning it was stated in an email, call, note, or form fill but never logged as its own touchpoint. Twenty-four percent was inferred from the evidence, like a prior relationship, brand awareness, or a champion pushing internally.
A few other findings stood out to me
- The buying group is mostly invisible. On won deals, the CRM showed 2.7 people. The analysis found 7.9 involved.
- What gets logged most matters least. In the year before a deal opened, marketing email made up 45% of everything logged on the account, but earned 3.6% of the credit.
- Better data changes the answer, not just the precision. Using the same linear multi-touch model, marketing's share of credit went from 20% on raw CRM data to 37% on the cleaned-up, joined data.
- The untracked touches carry more weight. Most meaningful touches on a deal were tracked, but a touch found only in unstructured data carried about twice the credit of one recorded in a system.
Seghete puts it plainly: "Half of what matters in getting a new-business deal sits in unstructured data."
Read this as a strong signal, not settled science. The paper is clear about its limits. The deals are a convenience sample from companies that bought Upside. The AI's judgments cite evidence but weren't tested against an outside ground truth. There's no control group. And the author co-founded the company that sells the attribution product used in the study, which the paper itself says readers should weigh.
Still, the direction matches what I've seen as an operator, and it matches what independent research says about how B2B buying works.
The record is the problem, not the model
Every attribution model splits credit across the touches it can see. If half the evidence never made it into the system, the smartest model in the world is still dividing up an incomplete pie.
The buying process makes this worse every year. In 6sense's 2025 B2B Buyer Experience Report, based on more than 4,000 buyers, buying groups average more than ten people, and buyers report an average of 16 interactions per person with the winning vendor. Buyers don't reach out to sellers until about 61% of the way through their journey. Kerry Cunningham, 6sense's head of research and thought leadership, said in the report's release: "Buyers are choosing a preliminary winner much earlier than they have in the past."
Camille Ricketts, who was the first marketing leader at Notion, named the structural problem in a September 2026 post, "Finally a better way to measure and manage ABM", on The Upside Down Funnel, a newsletter Upside publishes. CRMs are built around leads, contacts, and opportunities, so "your strategy depends on collective decisions, but runs on top of a measurement layer that only counts individual behaviors."
What I learned when our own dashboard told the wrong story
I've run into this from the consumer side, and it taught me to distrust a clean number.
At a community-led consumer platform I led, a large share of our traffic was labeled "Direct" in the dashboard. On paper, people were simply typing in our URL and showing up.
We didn't believe it. So we tested it the only honest way I know. We turned off paid media and creator spend entirely for a stretch, so we could see what growth looked like with none of the usual shortcuts running underneath it.
What we found didn't match the dashboard. A real portion of what had been counted as "Direct" was word of mouth and community influence that no tracking pixel was built to see.
The dashboard wasn't lying on purpose. It was reporting the touches it could observe. The mistake was ours, for treating an incomplete record as a complete one.
B2B has the same problem with higher stakes. The deals are bigger, the buying groups are larger, and the missing touches are in places a pixel will never reach.
Why AI agents raise the stakes
This used to be an analyst's problem. Someone pulled a report, someone else questioned it, and the gaps got argued about in a budget meeting. A person reading a report knows the CRM is messy and adjusts in their head.
An agent doesn't. GTM teams are putting AI agents directly on top of their revenue data. Ask which channel drove pipeline last quarter, and an agent answers in seconds, in confident prose, with no hedging. It can't tell a touch logged in Salesforce from one pulled out of a call transcript or one guessed from a pattern unless the record tells it.
Seghete's paper names the cost: "once an agent acts on it, the result is a wrong decision, made fast and repeated at scale."
That's why the most useful idea in her study isn't the 49%. It's the labeling. Every touch is graded by how you know it.
- Tracked: it has a system record, like a logged email, meeting, call, campaign membership, page view, or form.
- Extracted: it was found in unstructured data, like an email, transcript, note, or form fill, but never logged as its own touchpoint.
- Inferred: no record holds it, but the evidence implies it, like a buying window, a prior relationship, brand awareness, or a champion.
Camille adds a fourth label I'd keep: unresolved, for anything you can't trace at all. As she puts it, "Give uncertainty an honest place to sit so you can see how much of deals depend on the invisible."
You don't need any particular vendor to adopt this. Every data point in your revenue record should carry its own provenance: where it came from, and how sure you are.
Where this approach can go wrong
Rebuilding the record is not a free lunch. Three things deserve a hard look before you trust it.
Extraction can be wrong. Pulling a touch from a call transcript depends on the transcript and on the model reading it. Citations back to the exact line matter because they let a person check the claim in seconds.
Inference is still a guess. Labeling something as inferred is honest. It doesn't make the inference right. Be careful about putting real budget behind credit that rests mostly on inferred touches. Upside's approach, as Camille describes it, runs several independent AI analysts on each deal, each arguing for how credit should be split, with a judge that weighs the quality of the reasoning and escalates when they disagree badly. That's the right instinct, but it's still judgment, not proof.
No attribution model proves cause. The paper says so itself. A fuller record tells you where to look. It doesn't tell you what would have happened without a given channel.
None of this argues for staying with the three-touch CRM view. It argues for being clear about what each number rests on.
A provenance check before you trust an attribution number
Here's the discipline I'd put in place, whichever tool you use. I think of it as a provenance check, and it comes down to four steps.
1. Inventory where deals actually happen. List every place a buying conversation leaves a trace: CRM, marketing automation, call recordings, email, ad platforms, web analytics, event lists, community channels. Then mark which ones feed your attribution today. The gap between those two lists is your blind spot.
2. Label every touch by how you know it. Tracked. Extracted, with a link back to the source. Inferred. Unresolved, when you can't trace it. If your current system can't tell you which is which, that's the first thing to fix.
Image 2: "Label every touch by how you know it" diagram
3. Report credit by evidence grade. When you show leadership which channel drove pipeline, split it: how much rests on tracked touches, how much on extracted ones, how much on inference. A channel that wins mostly on inferred credit is a hypothesis, not a result.
4. Test the big calls against reality. Before you move real budget, run a holdout. Turn a channel off in one region or segment for a fixed window and watch what happens to pipeline. A good record tells you where to look. A test tells you whether it was right.
What to ask any attribution vendor
Upside isn't the only company working on this. Tools like HockeyStack and Dreamdata also pull B2B revenue data together from multiple systems, and each makes its own choices about how touches are captured and credited.
Ask every one of them the same questions
- Which sources does it read beyond the CRM?
- Can it show me where every touch came from?
- Does it separate tracked, extracted, and inferred data, or blend them?
- If an AI agent answers a question from this data, will the answer carry those labels?
The last question is the one most buyers aren't asking yet. It's the one that will matter most.
What to do now
This quarter. Run the inventory in step one. Count how many of the places your deals actually happen feed your attribution today. If it's less than half, that's your first project.
Next two quarters. Start labeling. Even a manual tag in a spreadsheet for your top 20 deals will show you how much of your current story rests on guesses.
Before you put an AI agent on your revenue data. Make sure the record carries provenance first. An agent on an unlabeled record doesn't fix attribution. It just makes the wrong answer faster and more convincing.
In Lean AI, I argued that small teams win by letting machines do the work they're good at. That still holds. But machines are only as good as what they're reading, and in B2B attribution, what they've been reading is about half the deal.
When Camille's post went up, I left a comment that still sums up how I see it. The biggest shift is moving from measuring activity to understanding decision-making. AI lets us rebuild that context at scale, without pretending every touchpoint can be reduced to a precise percentage.
The teams that win the next few years won't have the fanciest attribution model. They'll have the most honest record.
FAQ
How much does B2B attribution miss?
In Upside's study of 521 new-business deals at 13 B2B companies, 49% of the credit came from unstructured data, such as emails, call transcripts, notes, and relationships, that was never logged as a touchpoint. The sample comes from Upside customers, so treat it as a strong signal rather than a market benchmark.
What is the difference between tracked, extracted, and inferred touchpoints?
A tracked touch has a system record. An extracted touch is stated in unstructured data, like a call transcript or email, but was never logged as its own touchpoint. An inferred touch has no record, but the evidence implies it, like a prior relationship or brand awareness. Some teams add a fourth label, unresolved, for touches they can't trace at all.
Can AI agents fix marketing attribution?
Not by themselves. An agent can only work from the record it's given. If the record is incomplete or doesn't label what's known versus guessed, the agent will repeat those gaps with confidence. Fix the record and its labels first.
What is the best attribution model for B2B?
The model matters less than the data underneath it. In the same study, the same linear multi-touch model gave marketing 20% of the credit on raw CRM data and 37% on cleaned-up data. Rebuild the record first, label each touch, and confirm big budget calls with holdout tests.
How do you test whether attribution is right?
Run a holdout. Turn off a channel in one region, segment, or time window and compare pipeline against a similar group where it stayed on. If pipeline barely moves, the channel was getting credit for demand it didn't create.
Why is ABM so hard to measure?
Because CRMs and reporting are built around individual leads and actions, while account-based deals are decided by a group of people over months, often before an opportunity exists in the CRM. Measuring ABM well means one timeline per account, starting the day you decided to target it.