01The Attribution Problem
Enterprise workflows blend human judgment with AI execution at every step. A customer calls a contact center about a disputed charge. An AI agent retrieves the account history, identifies the relevant transaction, and drafts a resolution. The human agent reviews the draft, corrects the refund amount, adds a personal note, and sends the response. The customer receives a resolution. The question: who did the work?
The vendor's dashboard credits this as an AI-resolved interaction. The workforce management system credits the human agent with a completed case. Neither is accurate. The AI retrieved data and drafted a response. The human corrected a factual error and personalized the communication. Labeling this as "AI work" or "human work" misrepresents reality. The question is what percentage of the outcome is attributable to each actor — and whether that attribution is defensible under scrutiny.
Without structured attribution, organizations cannot determine the actual value contributed by each actor, the cost associated with each contribution, or the liability exposure when outcomes are disputed. The attribution gap creates financial, operational, and regulatory risk simultaneously.
02Why Binary Attribution Fails
Most enterprise systems track work ownership, not work contribution. A ticket is assigned to an agent. A case is owned by a team. These are administrative labels, not attribution measurements.
Example: Support ticket attribution
A customer submits a warranty claim through chat. The AI agent classifies the claim, retrieves the warranty terms, and generates a denial response citing the expired warranty period. The human agent reviews the response, notices the customer purchased an extended warranty that the AI missed, reverses the decision, and approves the claim. The vendor reports this as an "AI-handled interaction." The workforce system shows the human agent spent 4 minutes on the ticket.
In reality, the AI's contribution was negative — it produced an incorrect decision that required human intervention to reverse. Attributing this interaction to the AI inflates automation metrics. Attributing it to the human undercounts the cost of AI error correction.
Effective attribution requires step-level analysis that accounts for:
03How Veratrace Attributes Work
Each Trusted Work Unit captures the full sequence of evidence events with actor identification at every step. The Attribution Engine then calculates contribution percentages using configurable policies that organizations can adjust to reflect their operational priorities.
Example: Attribution calculation
A TWU for a customer refund interaction contains four evidence events: (1) AI retrieves order history, (2) AI generates refund recommendation, (3) human agent modifies the refund amount from $45.00 to $67.50 and changes the reason code, (4) system processes the refund. The Attribution Engine evaluates: the AI's data retrieval was used as-is (credit: full), the AI's recommendation was substantively modified (credit: partial, weighted by edit significance), the human's modification changed the financial outcome (credit: decision authority), and the system executed mechanically (credit: none). Result: 34% AI, 66% human. Not a guess. A calculated percentage derived from sealed evidence.
The engine distinguishes between AI-initiated, AI-completed, human-completed after AI initiation, and fully human workflows. It flags rework — cases where a human agent substantially modified an AI-generated output before delivery — and adjusts attribution accordingly.
04Attribution and Billing
Accurate attribution directly impacts cost allocation.
Example: Contact center billing dispute
A 150-seat contact center pays an AI vendor $0.85 per resolved interaction. Monthly volume: 40,000 interactions. Monthly AI cost: $34,000. Independent attribution analysis reveals that 12,000 of those interactions (30%) required substantive human rework — agents spending 3-7 minutes correcting AI-generated responses before delivery. The enterprise is paying $10,200 per month for AI "resolutions" that human agents actually resolved. Over a year, that is $122,400 in charges for work the AI did not complete.
Attribution data makes this visible. It enables accurate vendor reconciliation by providing the evidence needed to distinguish between AI-completed work and AI-attempted work. It enables honest ROI measurement by accounting for the full cost of each outcome, including hidden rework costs.
05Regulatory Requirements
Frameworks like the EU AI Act require organizations to disclose when AI systems make consequential decisions. The Colorado AI Act requires notification when AI is used in high-risk decisions affecting consumers. NIST AI RMF calls for documented AI involvement in decision processes.
Example: Insurance claims compliance
A health insurance company uses AI to pre-screen claims for coverage eligibility. When a claim is denied, the regulatory framework requires disclosure of whether AI was involved in the decision. Without attribution data, the company cannot distinguish between claims denied by AI recommendation (which require disclosure) and claims denied by human reviewers who happened to use AI-generated summaries (which may not). The attribution record in each TWU resolves this: it identifies exactly what role the AI played — whether it made the recommendation, provided supporting data, or had no involvement in the denial decision.
All of these frameworks assume that organizations can answer a basic question: what did the AI do, and what did the human do? Attribution infrastructure provides the evidence to answer. Without it, compliance becomes a matter of assertion rather than demonstration — which is precisely what regulators are trying to prevent.
See how attribution fits into broader compliance infrastructure requirements.