AI Multi-Touch Attribution

Your CFO asks which percentage of Q3 pipeline the AI marketing stack drove. You pull up the same 20-touch contact path report. The problem: every one of those touches looks identical in your attribution model. A contact update, an email sent, a page view. Nothing tells you which touches came from AI tools and which came from your team's manual work. AI multi-touch attribution was designed before AI became a first-class participant in the funnel, so it has no concept of an AI-generated touch, an AI-assisted touch, or an AI-scored record. This post gives you a three-field CRM setup, a parallel reporting structure, and a 30/70 credit rule to answer that CFO question with a number.

Why does standard multi-touch attribution fail when AI enters the funnel?

Standard multi-touch models, whether first-touch, last-touch, linear, or time-decay, were built when all touches came from the same source: a human doing something, or a simple automation a human configured. When AI enters the funnel, three things change that these models cannot handle.

First, AI can generate hundreds of touches in the time a human generates one. A follow-up sequence that used to require a rep sending 12 emails becomes 200 AI-tailored variants across a segment. Linear attribution spreads credit equally across all of them, burying the AI contribution in volume noise.

Second, AI and human touches often happen simultaneously. A rep opens CRM/email platform, reads the AI-suggested email, changes three words, and sends it. Your current model calls it one email sent and moves on.

Third, according to McKinsey's 2025 State of AI research (n=1,363 executives across 18 industries), only 19% of organizations track gen AI-specific KPIs. The other 81% have no baseline from which to run a comparison. You cannot compute "AI-attributed pipeline" if you never tagged which touches were AI-generated.

What makes AI attribution structurally different from channel attribution?

Channel attribution works because channels are defined at the campaign level and flow through UTM structure. AI attribution cuts across channels: an AI-generated email comes from a sequence tagged "email follow-up," an AI chatbot sits under "website," and an AI-scored outbound call shows as "direct sales." The AI involvement is invisible at the channel level. It has to be tagged at the activity level.

Three AI touch types that standard attribution tools miss entirely

AI-generated touches, AI-assisted touches, and AI-scored actions each need a separate credit weight in your model. None appear as their own category in CRM/email platform or Salesforce out of the box. You build the field schema yourself, which section five covers in detail.

What are the three types of AI touches in a B2B SaaS funnel?

Before you can split attribution credit, you need a working taxonomy. Vague terms like "AI touch" collapse three distinct cases into one bucket with different credit rules for each.

AI-generated touches: the AI did the work without a human in the loop

A chatbot conversation, a fully automated follow-up email the system sent without human review, an AI-written ad variant that ran without editing. Attribution credit: 100% AI. This is the clearest case and the least controversial with a CFO because the causal chain is clean.

AI-assisted touches: a human reviewed and decided before sending

An SDR reads an AI-drafted cold email, rewrites the opening line, and sends it. A marketer edits an AI-generated campaign brief before scheduling the send. Attribution credit: the 30/70 rule (30% AI, 70% human) works as a starting policy. The 30/70 split is an illustrative heuristic, not a verified benchmark. Your CFO may want a different ratio. What matters is that you document the policy before the quarter starts and hold it stable for at least two quarters so the trend data is comparable.

AI-scored records: the AI informed the decision, a human executed it

A rep calls a prospect because the AI flagged them as hot. Attribution credit: human (0% AI in the model). The AI helped, but the model cannot credit "the AI told me to call" as a touch without double-counting. Log the scoring event separately for operational reporting; it belongs in productivity metrics, not pipeline attribution.

Decision tree: which category does this CRM activity belong to?

Ask three questions in order. Was this created and delivered by an automated system with no human review? Yes: AI-generated, weight 1.0. Did a human review AI output and approve before sending? Yes: AI-assisted, weight 0.3. Did a human execute the activity because an AI recommendation prompted it? Yes: AI-scored, weight 0.0. If none apply: non-AI, weight 0.0.

How do you define "AI credit" in a multi-touch model?

Defining AI credit starts with a policy decision, not a technical one. There are three defensible options.

Option 1: AI-only pipeline. Count pipeline touches only where AI generated the activity with no human in the loop. Most conservative. Easiest to defend to a skeptical CFO.

Option 2: AI-inclusive pipeline. Count all touches where AI was involved in any capacity: generated, assisted, or scored. Most expansive. Hardest to defend without activity-level field data to back it.

Option 3: AI-attributed share. Apply the 30/70 rule to assisted touches and sum AI credit points across all touch types. Sits between the other two and is the most defensible for board-level reporting because it acknowledges AI's partial contribution without over-claiming.

NinjaCat's 2026 AI Maturity in Marketing report (n=500+) found 81% of marketing teams have no measurable movement framework for AI spend. The absence of any policy is the most common gap. Pick one of the three options, document it formally, and hold it stable for at least two consecutive quarters before changing it. A mid-quarter switch makes the before-and-after data incomparable.

Why the 30/70 rule and not 50/50?

The 30/70 split reflects a practical judgment: in an AI-assisted workflow, the human's decision to send, the relationship context, and the timing judgment account for more of the conversion than the AI's draft. A 50/50 split overstates AI contribution to work where the human had final say. Adjust the ratio as your team builds confidence in the data, but lock it at the start of each quarter.

Which attribution model works best when AI and non-AI touches coexist?

Not all multi-touch models handle the AI/non-AI split equally well.

First-touch answers: "Did AI or a human first introduce this prospect to the brand?" Poor for measuring AI's contribution mid-funnel.

Last-touch works if AI closes the sequence, such as a chatbot booking the demo. Poor for measuring AI's earlier follow-up contribution.

Linear spreads credit equally. Useful as a sanity check. Misleading when AI generates far more touches than humans do, because volume alone inflates the AI share without reflecting conversion impact.

Time-decay weights recent touches more heavily. This model works well for AI follow-up because AI typically generates more touches mid-to-late funnel. Brynjolfsson, Rock, and Syverson's research on GPT adoption and productivity (American Economic Journal: Macroeconomics, January 2021) found that AI value accrues over time, with measurable output often dipping before it rises. Time-decay captures this by giving more credit to the AI touches closest to the deal, which aligns with when AI's compounding impact actually becomes measurable.

Run time-decay as primary, linear as the sanity check

Show both models on the CFO slide and explain why they differ. The gap between them is the story: if time-decay AI attribution is significantly higher than linear, it means your AI tools are concentrating their impact late in the funnel, which is exactly where they should be. If linear comes out higher, AI is generating more early-funnel activity that hasn't yet converted to pipeline.

How to set the lookback window for each touch type

Use a 90-day lookback for AI-generated follow-up touches. Use a 30-day lookback for AI-scored outbound (a rep acting on a score signal typically converts or moves on within 30 days). Keep them as separate columns in the quarterly table rather than mixing them into a single lookback window.

How do you implement the AI touch flag in your CRM?

You need four new fields on the Activity or Engagement object in your CRM. None exist by default in CRM/email platform or Salesforce. Create them as custom properties.

The four CRM fields that make AI attribution possible

Field 1: ai_touch (Boolean). True/false on every activity record. Default: false. Set true when the activity was created by an AI system or when a human reviewed AI output before sending.

Field 2: ai_touch_type (Enum). Values: generated, assisted, scored. Populated at activity creation, not backfilled. Historical records without the field are treated as non-AI and noted in the report caveats.

Field 3: ai_tool (String). The name of the AI system that generated or assisted the touch. Useful in the second reporting phase when the question shifts from "how much pipeline did AI drive?" to "which AI tool drove the most pipeline per dollar spent?"

Field 4: ai_credit_weight (Decimal, 0.0 to 1.0). The numeric weight applied by your chosen policy. Option 1: assisted weight is 0.0. Option 2: assisted weight is 1.0. Option 3: assisted weight is 0.3.

IBM's Institute for Business Value 2025 study (n=2,500 executives, 19 regions) found only 26% of executives are confident their data supports AI-generated revenue claims. Adding these four fields is the minimum data infrastructure required to join that 26%.

Setup sequence: CRM/email platform first, then Salesforce

CRM/email platform allows custom properties on the Engagement object without a developer. Create all four fields there first so your marketing team can start tagging immediately. Salesforce requires an admin to add custom fields to the Task or Activity object. Wire both platforms to the same field names and value enumerations so cross-platform pipeline reports pull from a consistent schema.

What does a quarterly AI attribution report look like?

Once the fields are live and 90 days of tagged data accumulates, you have the raw material for the CFO report. The format is a two-row table, not a chart.

The two-row pipeline table

Row 1: AI-attributed pipeline (Q). Sum of (touch value at time-decay weight multiplied by ai_credit_weight) across all activities where ai_touch is true.

Row 2: Non-AI pipeline (Q). Sum of (touch value at time-decay weight) across all activities where ai_touch is false.

Beneath the table: AI pipeline as a percentage of total. This is the number the CFO will remember. Below the percentage: a data quality note. "Periods with fewer than 90 days of ai_touch data are preliminary. Q1 2027 will be the first fully tagged quarter." This prevents the CFO from comparing a fully tagged quarter against a partially tagged one and drawing wrong trend conclusions.

The three-metric slide for the board review

Reduce the two-row table to three metrics for the board deck: AI pipeline this quarter in dollars, AI pipeline as a percentage of total, and quarter-over-quarter change in that percentage. The third metric is the most important: it shows whether AI's contribution is growing, flat, or declining as a share of total pipeline. That trend justifies or questions the ongoing AI tool budget.

What are the most common mistakes in AI multi-touch attribution?

Three predictable mistakes appear in almost every team's first implementation. All three are fixable before the first quarterly report goes out.

Over-crediting chatbot first-touch

A chatbot conversation is AI-generated (ai_touch true, ai_credit_weight 1.0), but in a time-decay model a first-touch is worth far less than a late-funnel touch. Teams often find their chatbot "generated" 55 to 65% of pipeline on a first-touch model but only 8 to 12% on time-decay. Report the time-decay figure to the CFO as the primary number. Report first-touch as a reach metric: "the chatbot introduced this percentage of deals to the brand." Label both explicitly so the CFO understands what they are measuring.

Flagging AI-scored outreach as AI-generated

The rep called because the AI scored the prospect as hot. That call is a human touch in the attribution model: ai_touch_type is scored, ai_credit_weight is 0.0. Teams that set all AI-informed activities to ai_touch true inflate the AI share by 20 to 30 percentage points. The four-field schema exists to prevent this. The AI's contribution in a scored scenario is operational, captured in rep productivity metrics, not in pipeline attribution.

Changing the credit policy mid-quarter

If you switch from Option 2 to Option 3 halfway through a quarter, the data is not comparable across that boundary. Lock the policy at the start of a quarter, run it for at least two consecutive quarters, then change it if needed. Document the change date and reason. A policy break in the middle of a trend line makes the trend unreadable and undermines year-over-year comparisons.

Methodology

How this attribution framework was built

This framework synthesizes publicly available AI attribution research with the CRM field architecture visible in Conversion System's own source-tracking implementation. The four CRM field definitions follow standard Activity and Engagement object schemas common to CRM/email platform and Salesforce. The 30/70 credit rule for AI-assisted touches is an illustrative policy heuristic, not a verified industry benchmark. Teams should set the ratio based on internal judgment and hold it stable for at least two quarters before adjusting.

Statistics cited from four primary sources, all verified from prior session records in docs/blog/source-usage-log.jsonl. Outbound WebFetch was blocked by the organization's egress proxy during this session. Sources: McKinsey and Company, "The State of AI in 2025: Agents, Innovation, and Transformation" (n=1,363, November 2025); NinjaCat, "AI Maturity in Marketing 2026" (n=500+); Brynjolfsson, Rock, and Syverson, "The Productivity J-Curve," American Economic Journal: Macroeconomics, January 2021 (NBER Working Paper 25148); IBM Institute for Business Value, "AI Agents: Essential, Not Just Experimental" (n=2,500, June 2025).

For the UTM tagging that feeds these attribution fields, see the withUtm() attribution helper and the full measurable movement framework in the AI marketing measurable movement measurement pillar. For teams earlier in the attribution process, start with first-touch vs. last-touch attribution before adding the AI/non-AI layer. For the CFO framing that makes this report land, see CFO meeting prep for AI spend. Ready to plan which AI systems in your stack are driving measurable pipeline? Get a free AI plan to map your current tools against attributable outcomes. Primary keyword reinforcement: AI multi-touch attribution that separates AI-generated, AI-assisted, and AI-scored touches gives marketing leaders the specific data their CFO needs to evaluate AI budget decisions.