Can the Buyer Operate This AI Asset on Day One?

Technical due diligence on AI companies still wastes deal-room hours grading prompt templates or debating whether the target fine-tunes its own weights. For an acquiring CTO, private-equity operating partner, or investment committee, the load-bearing question is much narrower: if the transaction closes on Friday afternoon, what happens to gross margin, legal title to the evaluation and training corpus, and production uptime on Monday morning when the asset separates from the seller’s environment and founder heroics? Evaluating the technical architecture itself requires verifying gateway taint controls , per-task pass^k consistency , and retrieval permission integrity . Yet even when the code passes those checks, AI acquisitions routinely miss their first-year EBITDA model because of three transaction-specific liabilities that never appear in a product demo.

The first liability sits inside the unit-economics spreadsheet, where blended average gross margin hides both heavy-tail usage exposure and money-losing enterprise contracts. Never accept a deal-room COGS summary that divides monthly model-provider invoices by total active seats. Demand thirty days of raw inference billing exports, orchestrator retry traces, background compaction logs, and human-in-the-loop review queue records joined by customer tenant_id. Two buried costs almost always surface in that join: targets frequently book human review queues (operators fixing failed extractions or approving escalated tool calls) under R&D or customer-support payroll rather than cost of revenue, inflating reported software gross margin by fifteen to thirty points, and inference bills are often artificially depressed by non-transferable cloud startup credits or introductory model pricing that steps up on a fixed calendar date.

To see the real margin structure, re-price every tenant’s thirty-day trace volume under the acquirer’s unsubsidized post-close rate card, add the fully loaded cost of human review minutes per tenant, and compute two separate distributions. First, calculate tail cost-to-serve (p50 versus p95 cost-to-serve per tenant) to bound worst-case usage spikes on uncapped agentic workflows. Second, calculate tenant contribution margin (p5 customer margin, joining fully loaded cost-to-serve against actual contract revenue) to expose flat-fee or underpriced enterprise accounts that are deeply gross-margin negative even at moderate request volume.

The second liability is a broken title chain on the target’s pitched “data and evaluation moat.” Nearly every AI seller claims a compounding flywheel built from production traces, user thumbs-up/down signals, human-reviewer corrections, and curated regression fixtures. In diligence, treat that corpus like unlicensed intellectual property until proven otherwise. Audit the customer Master Services Agreements, Data Processing Addenda, and upstream model-vendor terms clause by clause against the actual telemetry schema in the database. In enterprise B2B deals, the largest customers routinely strike the vendor’s right to use customer inputs or outputs for service improvement, or require deletion of all derived artifacts upon contract termination or change of control. If the target pooled raw customer prompts, human edits, and fine-tuning examples into a single shared evaluation dataset without tagging each row by tenant_id and contract-rights provenance (and without tenant-keyed encryption that lets you crypto-shred a single customer’s records cleanly), a single enterprise opt-out or change-of-control notice after closing can force the buyer to quarantine or delete the entire evaluation suite.

The third liability is Day-One operational coupling, which belongs in a priced Separation Register before the purchase agreement is signed. In a carve-out or tuck-in acquisition, AI systems fail on Monday morning over mundane boundary dependencies: enterprise model rate tiers and committed-throughput reservations (PTUs or provisioned concurrency) that belong to the parent company’s master cloud account and cannot be assigned; hardcoded pins to deprecated model snapshots whose shutdown window falls inside the first sixty days post-close; internal MCP servers or SaaS connectors authenticated via a departing founder’s personal OAuth grant; and undocumented Slack channels where two senior engineers manually unstick wedged agent runs twenty times a day.

Turn those three audits into numbers that go straight into the deal model and the integration plan. Every row in the Separation Register gets a named owner, a Day-One cutover requirement, and a priced 30/60/90-day remediation cost; every negative-margin contract at the p5 customer margin tail gets a repricing or usage-cap schedule at renewal; and any evaluation corpus lacking tenant-level provenance tags gets discounted to zero in the valuation model until a clean-room rebuild is priced. Before your investment committee gives final approval to sign the purchase agreement, ask the deal team one question: have we reconstructed p95 cost-to-serve and p5 customer margin from raw per-tenant traces at post-close vendor rates, including off-book human review minutes, or are we buying a blended gross-margin slide?