What should be on an AI agent due diligence checklist before deploying agents in a business?

An AI agent due diligence checklist is the structured set of technical, legal, financial, and operational questions you must answer before letting an autonomous software agent touch your data, your customers, or your money. Unlike traditional vendor due diligence, which mostly verifies security certifications and financial stability, agent due diligence has to account for systems that take actions on their own, chain together multiple tools, and fail in ways that are probabilistic rather than deterministic. As of August 2026, most enterprises have moved past pilot projects and into production deployments, and regulators, insurers, and acquirers now expect documented diligence on any agentic system. This guide walks through what a serious checklist covers, why each item matters, where teams commonly go wrong, and how the process differs depending on whether you are buying an agent platform, building one internally, or evaluating a startup whose core product is an agent.

Why Agent Due Diligence Is Different From Software Due Diligence

Also worth reading: What Should Founders and Operators Look for in an AI Buyer Diligence Checklist in 2026? · Which AI Diligence Pilot Metrics Actually Prove Business Value in 2026? · How Can Founders Use a Private AI Agent Deal Flow Network to Accelerate Financial Due Diligence?

Traditional SaaS due diligence assumes a tool that does what a human instructs it to do, when instructed. Agents break that assumption. An agent plans multi-step workflows, calls external APIs, writes and executes code, and adapts its behavior based on context. That means the failure modes are different: a hallucinated step in a legal document review workflow does not just produce a wrong answer, it can file the wrong document, send the wrong email, or commit the company to terms nobody approved. IBM's published case work with Fogel Law Group showed document review dropping from hours to minutes with watsonx-assisted review, but that speed is only valuable if accuracy thresholds and human review gates were verified before deployment — which is exactly what diligence establishes.

The second difference is accountability diffusion. When an agent chains five tools together and something breaks at step four, responsibility may sit with the model provider, the orchestration layer, the tool vendor, or the integrator who wired them together. Your checklist needs to map that chain explicitly and identify who carries contractual liability at each link. The third difference is drift: models get updated, prompts get edited, retrieval indexes change, and behavior that was validated in March may not hold in September. Diligence for agents is therefore not a one-time gate but a recurring audit cycle — most mature operators re-run the full checklist every six months, with lightweight quarterly reviews of anything touching regulated data.

The Core Checklist: Ten Items Every Review Should Cover

A practical checklist starts with ten items. First, model provenance: which foundation model powers the agent, what version, what training-data disclosures exist, and what happens when the provider deprecates that version. Second, evaluation results: request benchmark scores on tasks matching your use case, not generic leaderboards — ask for task-specific pass rates, error taxonomies, and regression test suites. Third, guardrails: what prevents the agent from taking destructive actions, and are those guardrails tested adversarially? Fourth, data handling: what the agent stores, where, for how long, whether your data trains anyone's models, and how retrieval sources are permissioned. Fifth, human-in-the-loop design: which action classes require approval, and is that enforced in code rather than in policy documents.

Sixth, observability: full trace logs of every reasoning step, tool call, and output, retained long enough for incident forensics — 90 days minimum, 12 months for regulated functions. Seventh, security posture: SOC 2 Type II, penetration testing of the agent's tool integrations specifically (agents expand attack surface through the APIs they can call), and prompt-injection testing results. Eighth, liability and indemnification: contract language covering damages caused by autonomous actions, which remains one of the least standardized areas in 2026 vendor agreements. Ninth, cost transparency: token consumption per task, inference pricing trends, and exit costs if you switch providers. Tenth, rollback and kill switches: the ability to disable the agent instantly and revert to manual process without data loss. Any vendor that cannot answer all ten within two weeks of asking is telling you something important about their operational maturity.

Evaluating the Vendor Stack: Model Provider vs. Agent Platform

Most buyers face a build-versus-buy decision between working directly with a foundation-model provider and buying a packaged agent platform. The trade-offs are real on both sides, and the right answer depends on your engineering capacity and risk tolerance. Direct model access gives you control over prompts, evals, and data flows, but you own every failure. Packaged platforms move faster but insert another party between you and the model, complicating both liability and debugging. The table below summarizes how the diligence emphasis shifts between the two paths:

Neither column is automatically safer. A well-run internal deployment with rigorous evals can outperform a sloppy platform deployment, and vice versa. What matters is that the diligence items above get answered somewhere in the stack — the mistake is assuming the platform has already done the work for you.

Regulatory and Legal Diligence in the Post-Executive-Order Environment

The regulatory picture changed materially after the federal AI executive order and subsequent agency guidance began filtering down into procurement requirements. TechTarget's analysis of the executive order noted that vendor management strategies had to shift from point-in-time certification toward continuous assurance: agencies and their contractors increasingly require ongoing monitoring commitments, incident reporting timelines measured in days rather than quarters, and documentation of model provenance. If you sell into government or regulated industries, your checklist should verify that any agent vendor can support these continuous-assurance obligations, because a vendor that cannot will become your compliance problem.

Trade secret protection deserves separate attention. Reuters' coverage of AI-era trade secret frameworks highlighted a specific gap: employees feeding proprietary information into third-party agents can inadvertently destroy trade secret status by disclosing secrets outside confidentiality controls. Your checklist should confirm contractual confidentiality from the vendor, technical controls preventing training on your inputs, and internal policies defining which categories of information may never enter an agent context window. Financial services buyers should also review sector-specific guidance — Anthropic's published materials on agents for financial services, for example, describe deployment patterns with explicit human-approval gates for transactions, which aligns with supervisory expectations at most banking regulators. The general rule: if a regulator would require a licensed human to approve an action, your agent workflow needs a named human approval step, and your diligence should verify that step exists in code.

Pricing Diligence: The Seat-Based Model Is Breaking

Pricing diligence used to mean checking per-seat costs against headcount. Agents break that math. When an agent performs the work that previously justified five seats, seat-based pricing either collapses or vendors migrate to outcome-based and consumption-based models. Startup Fortune's analysis of SaaS pricing under agent adoption describes the transition: vendors are experimenting with per-task pricing, per-outcome pricing, and hybrid models, and early numbers suggest consumption-based agent pricing can run anywhere from 2x to 15x the equivalent seat cost depending on task volume. Buttondown's writing on cheaper inference and supply-chain disruption adds the supply-side context: inference costs have fallen dramatically year over year, so a vendor quoting high per-task prices today may be defending margins that erode within 12–18 months.

For your checklist, this means three concrete asks. First, demand a cost model tied to your projected task volumes, not a rate card — a well-modeled deployment typically lands between $0.50 and $5.00 per completed agent task for knowledge-work use cases as of mid-2026, though complex multi-step workflows can exceed that. Second, negotiate price-adjustment clauses tied to published inference price indices, since falling input costs should flow through to you. Third, model the total cost including evaluation infrastructure, monitoring tooling, and the human review time agents still require; organizations routinely underestimate these adjacent costs by 40–60% in year one. A deal that looks cheap per task can be expensive per completed, verified outcome.

Common Mistakes and How to Avoid Them

The most frequent mistake is treating the demo as evidence. Demos are curated runs on favorable inputs; diligence requires running the agent against your own worst-case data, edge cases, and adversarial prompts before signing. Budget two to four weeks for a structured proof-of-concept with pre-agreed success thresholds — for example, 95% task completion with zero critical errors on a 500-case sample. The second mistake is skipping the tool-permission audit. Agents inherit whatever permissions the integration account holds, and an agent with broad API scopes becomes a lateral-movement vector during a breach. Scope every integration to the minimum permissions the workflow actually needs, and verify that scoping in a staging environment.

Third, teams conflate model quality with system quality. A strong foundation model wrapped in weak retrieval, sloppy prompt design, and missing guardrails produces a weak agent. Evaluate the assembled system, not the underlying model. Fourth, buyers neglect the exit plan. Ask how you export conversation logs, evaluation datasets, and configuration before you need to leave; vendors answer this question honestly only when asked upfront. Fifth, organizations assign diligence to IT alone. Because agents change business processes, the review should include the process owner, security, legal, and finance — a cross-functional review typically takes three to five weeks end to end, versus eight or more when routed sequentially through departments. Finally, do not over-index on brand names. A recognizable model name says nothing about the orchestration quality, data practices, or support responsiveness of the company selling you the finished product.

When to Act and How to Sequence the Work

Timing matters more than most teams assume. Run full diligence before any production deployment, before renewing an existing agent contract, and before any acquisition or investment in a company whose product depends on agents — in M&A contexts, agent-related liabilities around data handling and unapproved actions have begun appearing in representation-and-warranty negotiations. For startups raising capital, expect sophisticated investors to run exactly this checklist; preparing the artifacts in advance (eval reports, security documentation, data-flow diagrams) shortens fundraising by weeks. Founders and operators building deal-flow networks, private-market platforms, or operator communities increasingly share diligence templates peer-to-peer, since the fastest way to harden a checklist is seeing which questions tripped up someone else's vendor.

Sequence the work in four phases. Weeks one and two: gather documentation and run the ten-item core checklist against written responses. Weeks two through four: hands-on testing in a sandbox with your data, adversarial prompts, and failure-injection exercises. Weeks four to six: legal and commercial negotiation covering liability, indemnification, data rights, and pricing adjustments. Week six onward: staged rollout starting with low-risk internal workflows, expanding to customer-facing actions only after 30 days of clean incident logs. Total elapsed time for a disciplined process runs six to ten weeks. Skipping phases does not save time; it converts diligence failures into production incidents, which cost an order of magnitude more to remediate than the weeks you saved.

The Bottom Line

Agent due diligence is not bureaucracy; it is the mechanism that lets you capture genuine productivity gains — document review compressed from hours to minutes, research cycles cut by half, back-office workflows running overnight — without importing unpriced risk. The checklist above reflects what leading operators, regulators, and investors converged on by mid-2026: verify the system, not the marketing; scope the permissions; contract the liability; price the outcomes; and re-audit on a fixed cadence. Companies that institutionalize this process deploy agents faster over time, not slower, because they stop relitigating the same questions from scratch with every new vendor.", "faq": [ { "q": "How long does AI agent due diligence take?", "a": "A disciplined process runs six to ten weeks end to end: two weeks of documentation review, two to four weeks of sandboxed testing with your own data, then legal and commercial negotiation. Cross-functional parallel review keeps it under eight weeks; sequential departmental routing can stretch past twelve." }, { "q": "What is the biggest red flag when evaluating an AI agent vendor?", "a": "Inability to provide task-specific evaluation results on data resembling yours within two weeks. Vendors who only offer generic benchmarks or curated demos are hiding weak performance on realistic inputs. Refusal to scope tool permissions narrowly or explain data retention is a close second." }, { "q": "How much does it cost to run AI agents in production?", "a": "As of mid-2026, knowledge-work agent tasks typically cost $0.50 to $5.00 per completed task in direct inference costs, while packaged platforms charge 3–10x raw inference via per-seat or per-task pricing. Budget an additional 40–60% of year-one spend for evaluation infrastructure, monitoring, and human review time." }, { "q": "Do I need human approval steps in my agent workflows?", "a": "Yes, for any action class a regulator would require a licensed human to approve — payments, legal filings, medical decisions, and contract commitments. The approval gate should be enforced in code, not just policy documents, and your due diligence should verify it exists in the deployed system." }, { "q": "Should I buy an agent platform or build directly on a foundation model?", "a": "Buy a platform if you need deployment in 2–6 weeks and accept less control and higher switching costs; build directly if you have ML engineering staff and need custom workflows, accepting 8–16 weeks to launch. Either way, the same ten-item diligence checklist applies — the platform does not exempt you from verifying evals, data handling, and liability terms." } ], "quick_facts": [ { "label": "Category", "value": "Vendor risk management / AI governance" }, { "label": "Timeline", "value": "6–10 weeks for full diligence; re-audit every 6 months" }, { "label": "Cost", "value": "$0.50–$5.00 per agent task (inference); platforms charge 3–10x raw inference" }, { "label": "Best for", "value": "Operators deploying agents in production, buyers of agent platforms, and investors evaluating AI-native startups" }, { "label": "Core checklist size", "value": "10 items: provenance, evals, guardrails, data, HITL, observability, security, liability, pricing, rollback" } ], "sources": [ "https://startupfortune.com/how-to-price-saas-when-ai-agents-replace-seats", "https://buttondown.com/cheaper-inference-and-supply-chain-disruption", "https://www.ibm.com/case-studies/fogel-law-group-watsonx", "https://www.techtarget.com/ai-executive-order-vendor-management", "https://www.anthropic.com/agents-for-financial-services", "https://www.reuters.com/legal/ai-era-framework-protecting-trade-secrets" ], "follow_up_keyword": "AI agent vendor evaluation scorecard"