Oleg Malii, CEO of voice AI company Temvox, published an analysis in Forbes Tech Council on July 7 mapping where AI agents create genuine ROI within venture capital operations. The argument is specific: VC work splits cleanly into judgment-driven decisions (founder assessment, market timing, conviction) and operational plumbing (deal sourcing, document processing, CRM management). Agents belong in the plumbing. When they cross into judgment, they degrade signal quality rather than improve it.
The analysis matters beyond VC because it articulates a pattern visible across every professional services category attempting agent adoption: law, consulting, accounting, investment banking. The operational layer surrounding expert judgment is enormous, repetitive, and document-heavy. That layer is automatable now. The judgment itself is not.
The 60-70% Question
McKinsey estimated that generative AI could add $2.6 trillion to $4.4 trillion in annual value across industries, with current technologies capable of automating “60 to 70 percent of employees’ time.” The figure gets cited frequently. What gets ignored is the composition of that time.
In professional services, the 60-70% breaks down unevenly. A significant portion is reading, cross-referencing, summarizing, comparing, and following up. These are tasks that require context but not creativity, that demand accuracy but not conviction. They repeat across engagements with variations in detail but not structure.
The remaining 30-40% is where professional services firms earn their margins: reading a founder’s body language, sensing when a deal’s competitive dynamics have shifted, knowing when a client’s stated problem masks a deeper organizational failure. These tasks require pattern recognition built from years of domain-specific experience. No agent has that experience.
Malii frames this concretely for VC. Deal sourcing agents can monitor “public signals, funding news, hiring activity, founder content and sector-specific databases,” according to his Forbes analysis. A deck-review agent can extract “product, market, traction, revenue model, round size and stated milestones” and flag gaps: missing cohort data, vague IP claims, enterprise traction without customer specifics. A diligence agent can cross-reference founder claims against public data. All of this is the plumbing.
The Signal Degradation Problem
The analysis surfaces a risk that most agent deployment coverage ignores: agents can make signal quality worse, not better. Malii warns that “careless automation can flood a fund with mediocre deal flow,” suggesting that indiscriminate sourcing agents would increase volume without improving relevance.
This is the core tension in professional services agent deployment. A law firm deploying agents to review contracts faster might process three times the volume while missing the clause that matters. A consulting firm using agents to generate competitive analyses might produce polished documents that flatten the nuance a senior partner would have caught. The output looks better. The decision quality gets worse.
The risk is structural, not technical. Professional services firms charge for judgment, but they spend most of their time on operations. Agents are good at operations. The danger is assuming that operational speed compounds into better judgment. It does not. Faster document processing means nothing if the documents processed are the wrong ones.
One Fund Tested the Extreme Case
Business Insider reported in October 2025 that Davidovs Venture Collective launched a $75 million fund after removing five analyst roles. DVC replaced traditional analysts with AI agents and a network of 170 LPs who collectively handled sourcing, diligence, and portfolio monitoring. The agents run deal memos, match founders with experts, and generate portfolio follow-up summaries.
DVC represents the aggressive end of the spectrum. Most firms will not eliminate analyst positions. But the experiment tests a real hypothesis: can agents handle the operational layer well enough that a fund can redistribute analyst time toward higher-leverage activities? DVC’s structure suggests the answer is conditional. The agents handle structured tasks. The LP network handles the judgment and relationship layers that agents cannot replicate.
The distinction matters. DVC did not replace human judgment with AI. It replaced one form of human labor (junior analyst document processing) with AI, while expanding a different form of human labor (170 LPs providing domain expertise and network access). The total human involvement may not have decreased. Its composition shifted.
The Digital Due Diligence Gap
BCG reported that 73% of private equity firms now run digital due diligence on most deals. Only 22% said a company’s digital readiness influences their go/no-go investment decisions. That 51-percentage-point gap reveals the state of agent deployment in financial services more broadly: firms are generating more data-driven analysis than ever, but the analysis is not changing decisions.
This pattern repeats across professional services. Law firms run AI-assisted contract reviews but make the same deal terms they would have negotiated manually. Consulting firms produce AI-generated market sizing but adjust the figures based on partner intuition before presenting to clients. Investment banks use AI for comparable company analysis but rely on senior banker judgment for valuation recommendations.
The gap is not a failure of the technology. It reflects how professional services firms actually work: the operational layer exists to support expert judgment, not replace it. When agents make the operational layer faster and more comprehensive, the expert still reviews the output, adjusts it based on experience, and makes the decision. The agent saved time. It did not change the outcome.
Five Characteristics of Agent Workflows That Survive
Malii’s analysis identifies five characteristics of agent workflows that actually persist in professional services environments, rather than getting abandoned after initial deployment. Each addresses a specific failure mode.
First, successful workflows are narrow. A single agent handling sourcing, screening, and diligence creates accountability gaps that make review impossible. Separate agents for each function create clearer ownership and cleaner audit trails. When a sourcing agent produces a false positive, the error is identifiable. When a combined agent produces a flawed deal memo, isolating the failure point is harder.
Second, the agents connect to existing systems. Agents that require a new interface, dashboard, or workflow get abandoned within weeks. Agents embedded in the CRM, data room, and knowledge base the team already uses get adopted, because they reduce friction rather than adding a new tool to learn.
Third, every factual claim requires a citation. This is the rule that separates professional services agent deployment from consumer AI usage. In market sizing, competitor mapping, regulatory analysis, and customer research, an unsourced claim is worse than no claim. Agents that generate confident-sounding analysis without source trails actively damage decision quality, because professionals trust the format (a polished memo looks authoritative) even when the content is fabricated.
Fourth, human approval is preserved at every decision point. The agent prepares the analysis. The human makes the call. No exceptions, no “auto-approve for low-risk decisions” shortcuts. The moment a professional services firm allows an agent to make decisions without human review, the firm has accepted liability for agent judgment. No firm’s insurance covers that.
Fifth, the workflow handles failure gracefully. Agents fail. They hallucinate, misinterpret data, and occasionally produce output that is confidently wrong. Workflows that treat agent output as draft material, always subject to human review and correction, survive. Workflows that treat agent output as final product fail the first time the output is wrong and the client notices.
The Professional Services Pattern
The pattern emerging across VC, law, consulting, and financial services is consistent. Agents compress the operational layer. They reduce the time between a question being asked and the supporting analysis being assembled. They do not compress the judgment layer, and attempts to use them for judgment produce worse outcomes than manual processes.
For builders, this means the addressable market in professional services is large but specific. The operational plumbing around expert judgment is a multi-trillion-dollar category globally. Automating that plumbing is valuable. But the sales pitch cannot be “agents replace analysts.” It must be “agents make the time between question and answer shorter, so experts can spend more time on the decisions that earn the fees.”
Firms that understand this distinction will deploy agents effectively. Firms that treat agents as analyst replacements will discover what DVC’s experiment already suggests: the human judgment does not go away. It redistributes.