Autonomous Research Agents for Due Diligence

Key facts, with sources

• METR found the length of tasks frontier AI agents can complete autonomously with 50 percent reliability has been doubling roughly every 7 months since 2019. (METR)
• Continuations of METR's time-horizon tracking show frontier models in 2026 completing tasks that take human experts around 12 hours at 50 percent reliability, up from about 50 minutes for early-2025 models. (AI Digest)
• About 88 percent of AI agent pilots never reach production, with integration, reliability, latency, and security named as the main blockers rather than model quality. (Institute of Project Management)
• Gartner predicts at least 15 percent of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0 percent in 2024. (Gartner)
• The global AI agents market was valued at about $7.6 billion in 2025 and is projected to reach roughly $183 billion by 2033, a compound annual growth rate near 50 percent. (Azumo)

Agents beat analysts on breadth, not judgment

A research agent's real advantage is that it never gets tired on document forty. It applies the same extraction template to every contract in the data room, checks every named officer against registries, and reads the footnotes with the same attention as the headline numbers. Human analysts are demonstrably better at knowing which anomaly matters; agents are better at making sure no anomaly goes unread.

So I scope agents to first-pass coverage: enumerate the material, extract structured facts, flag inconsistencies between sources, and surface items that deserve senior attention. The output is a map of the territory, not an opinion about the deal.

The two failure modes that matter

Fabricated citations and stale data are the failures that can genuinely damage a transaction. A model asked to summarize beyond its retrieved context will produce plausible claims with invented attribution, and a claim about litigation status or ownership that was true two years ago can be materially false today.

My rule is blunt: a claim without a pinned source and a retrieval date is treated as false and either re-verified or dropped. That single policy, enforced mechanically in the pipeline rather than requested in the prompt, eliminates most of the risk that makes lawyers nervous about agent-assisted diligence.

Architecture: fan-out, grounded synthesis, verification pass

The pipeline has three stages. First, decomposition: the diligence question breaks into dozens of narrow sub-questions — corporate structure, encumbrances, key-person dependencies, contract change-of-control clauses — each dispatched with its own retrieval scope. Second, grounded synthesis: each sub-agent may only assert what its retrieved text supports, and every sentence in its output carries a document reference.

Third, and this is the stage most teams skip, an independent verification pass: a separate process reopens each cited source and confirms the quoted passage actually supports the claim attached to it. Claims that fail verification get flagged, not silently removed — a citation that didn't hold up is itself information the analyst should see.

The analyst owns the conclusion

I structure the final brief as a table of claims, each with its source, retrieval date, and a confidence marker, plus an explicit section for conflicts the agent found between sources. What the brief never contains is a recommendation. The moment an agent output says "we see no material risk," people stop reading the claims and start trusting the sentence — and no current system deserves that trust in a transaction context.

This division holds up well in practice: the agent compresses collection and cross-referencing dramatically, the human spends their hours on the judgment calls the fees are actually paying for, and accountability stays exactly where regulators and clients expect it.

The audit trail is a feature, not overhead

In diligence, you may later need to demonstrate why a conclusion was reached and what was examined — sometimes to a counterparty, sometimes in a dispute. So the system logs every sub-question issued, every document retrieved, every extraction made, and every claim's verification result, all tied to a run identifier.

This has a second benefit nobody anticipates: reusability. When a follow-up question arrives weeks later, the logged corpus and extractions mean the agent re-answers from already-verified material in minutes rather than re-crawling from scratch. The audit trail you built for defensibility becomes the cache that makes the second engagement fast.

When to hire senior help

Senior help matters most for the safety and reliability envelope, meaning sandboxing, permissions, rollback paths, and evaluation, which determines whether autonomy is an asset or a liability. If pilots keep failing on reliability rather than capability, an experienced agent engineer can usually diagnose whether the problem is tooling, prompts, or architecture within days. If your stack includes React Native + Python + AI, a senior engineer who owns the full product beats coordinating multiple juniors.

Bottom line

Dhairya Senjaliya ships AI — Autonomous Agents projects worldwide — book a scoping call to discuss your specific situation.

Common pitfalls to avoid

• ✕Ignoring compounding error rates; an agent that is 85 percent reliable per step succeeds only about 20 percent of the time across a 10-step workflow unless you add checkpoints and recovery
• ✕Granting write access to email, payments, or deletion without approval gates or sandboxing, turning a single hallucination into an irreversible action
• ✕Running long-lived loops with no budget cap, timeout, or kill switch, so a stuck agent burns tokens for hours before anyone notices
• ✕Evaluating on single runs when agent pass rates drop sharply under repeated-run consistency testing, making one good demo a misleading signal

Frequently asked questions

Sources

• METR
• AI Digest
• Institute of Project Management
• Gartner
• Azumo