Key Takeaways
- • Vivan Agrawal, cofounder of Zarna, an AI agent vendor for private equity firms, argues that PE firms need a control plane outside the model to keep AI agents from acting on the wrong deal or source.
- • A 2025 Deloitte survey of 1,000 US corporate and private equity leaders found 86% said their organizations had integrated generative AI into M&A workflows.
- • EnterpriseBench researchers reported that even the most capable AI models completed only 41.8% of 500 enterprise tasks spanning software engineering, HR, finance and administrative work.
- • A Google Research preprint testing 260 agent setups found a central coordinator improved results by 80.8% on parallel financial-reasoning tasks, while every multi-agent setup tested did worse by 39% to 70% on a sequential planning benchmark.
- • Agrawal recommends firms ask vendors what share of output claims trace to a source and how many CRM writes were confirmed by read-back.
Summary
Vivan Agrawal, cofounder of Zarna, which builds AI agents for private equity firms, describes a failure mode in which agents that each perform their assigned task still produce a wrong result. In his example, an associate asked an agent to move "Project Falcon" to IOI submitted. The CRM agent matched an older, dead 2023 deal, while the email agent addressed the live auction's banker and attached a valuation range from the 2023 model tagged "Falcon IOI." Only the associate's manual review prevented a bid range from being sent to the wrong company's banker. Agrawal writes that no model misread anything; each agent independently decided which deal was meant.
Agrawal cites a 2025 Deloitte survey of 1,000 US corporate and private equity leaders, in which 86% said their organizations had integrated generative AI into M&A workflows. He says his firm's early agents performed well on single documents but broke down across documents, the CRM and email. The fix was to set a fixed firm, team and deal scope in ordinary code before the model runs, with delegated agents receiving narrower slices of that scope and widening requiring a policy change or human approval.
The article draws on research to support a central orchestrator that resolves the deal once, asks the user to disambiguate when records match, and assigns each agent the same deal ID. EnterpriseBench researchers reported that the most capable models completed only 41.8% of 500 tasks, with difficulty attributed to data split across systems, role-based access controls and cross-functional workflows. A Google Research preprint testing 260 agent setups found a central coordinator improved parallel financial-reasoning results by 80.8%, while multi-agent setups performed worse by 39% to 70% on a sequential planning benchmark; independent agents amplified errors 17.2 times compared with 4.4 times under central coordination. MAST researchers, reviewing more than 1,600 traces from seven multi-agent frameworks, identified 14 failure modes, including agents repeating steps and ending before verifying results.
Agrawal argues that firms should keep every source value alongside its origin, observation date and trust ranking for each field rather than merging figures, and that unsourced or conflicting material claims should block delivery. Writes to other systems should require approval for the exact record, field and value, carry unique keys to prevent duplicate application, and be read back for confirmation. He suggests PE firms measure the share of material claims traceable to a source and the share of attempted writes confirmed by read-back, and ask vendors those two questions before adoption.
Agrawal cites a 2025 Deloitte survey of 1,000 US corporate and private equity leaders, in which 86% said their organizations had integrated generative AI into M&A workflows. He says his firm's early agents performed well on single documents but broke down across documents, the CRM and email. The fix was to set a fixed firm, team and deal scope in ordinary code before the model runs, with delegated agents receiving narrower slices of that scope and widening requiring a policy change or human approval.
The article draws on research to support a central orchestrator that resolves the deal once, asks the user to disambiguate when records match, and assigns each agent the same deal ID. EnterpriseBench researchers reported that the most capable models completed only 41.8% of 500 tasks, with difficulty attributed to data split across systems, role-based access controls and cross-functional workflows. A Google Research preprint testing 260 agent setups found a central coordinator improved parallel financial-reasoning results by 80.8%, while multi-agent setups performed worse by 39% to 70% on a sequential planning benchmark; independent agents amplified errors 17.2 times compared with 4.4 times under central coordination. MAST researchers, reviewing more than 1,600 traces from seven multi-agent frameworks, identified 14 failure modes, including agents repeating steps and ending before verifying results.
Agrawal argues that firms should keep every source value alongside its origin, observation date and trust ranking for each field rather than merging figures, and that unsourced or conflicting material claims should block delivery. Writes to other systems should require approval for the exact record, field and value, carry unique keys to prevent duplicate application, and be read back for confirmation. He suggests PE firms measure the share of material claims traceable to a source and the share of attempted writes confirmed by read-back, and ask vendors those two questions before adoption.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.