Keeping control when tools can act

Artificial intelligence is already inside the companies investors own, the managers they select, the research tools they use and the operating processes that support investment decisions. The first visible wave of generative AI was largely about personal productivity: drafting, summarizing, translating, coding and answering questions.

With the evolution of genAI to agentic AI, the discussion changes because the latter can support multi-step processes. It can gather information, compare sources, use tools, preserve context, route outputs and help teams act on defined goals. That creates opportunity: wider research coverage, faster scenario analysis, sharper monitoring, better documentation and less repetitive preparation.

But the governance question also changes. When AI is used for a single task, the main concern is whether the output is accurate enough for that task. When AI becomes part of a workflow, the question becomes broader: what does the system influence, what data does it touch, what actions is it allowed to support, and who remains accountable when something goes wrong? Instead of simply avoiding AI, the right answer is to set the conditions under which it can be used safely, responsibly and effectively.

Most AI risks in an investment institution can be grouped into four practical categories. They are familiar governance risks in a faster wrapper:

A useful oversight framework can be built around four associated questions. The standard should rise with the consequence of the AI-assisted action; for example, internal summarization doesn’t need the same controls as investment recommendations, manager selection or client-facing communication.

1. Grounding: where does the answer come from?
Grounding asks whether an AI-supported output is based on real, checkable source material rather than fluent guesswork.

For a low-consequence task, grounding may mean citing sources and spot-checking them. For a higher-consequence task, important claims should trace back to original documents or approved data sources. A confident answer without traceable evidence should not influence a consequential decision.

2. Stability: does it hold up if assumptions change?
Stability asks whether the output remains reasonable when the inputs or framing change.

Does the conclusion survive a different time window, cost assumption, peer group, scenario or prompt wording? If an answer flips under small changes, it may be too fragile to rely on, even if it sounds convincing. For investment processes, stability does not mean eliminating uncertainty. It means making fragility visible before the output is escalated.

3. Permissioning: what is the system allowed to do?
Permissioning asks what the AI system is authorized to access, produce or trigger, and where human approval is required.

The same tool might be allowed to summarize internal documents, allowed to draft a recommendation for human review, but never allowed to send external communication, change a portfolio, access restricted data or trigger a trade-side action without approval. These rules should be written down, reviewed and, where possible, built into the system through access controls, workflow design and approval points.

4. Accountability: who owns and signs off?
Grounding, stability and permissioning help reduce risk, but they don’t remove the need for human accountability.

As AI becomes more capable, investors need to know who owns each AI-supported process: who approved the use case, who reviewed the output, who signed off, what record was kept and what happens if the system produces an error.

The more consequential the use, the clearer the evidence trail should be. A later reviewer should be able to reconstruct what AI was used for, what it produced, what sources were used, what human review took place and what decision followed. This evidence trail matters for investment discipline, client trust, internal control and regulatory defensibility. It also helps separate raw model output from firm-approved work.

Independent evaluation here is essential: an AI-supported workflow should not be assessed only by asking the system whether it performed well. For consequential uses, firms need independent checks on whether the output was grounded, stable, permissioned and useful. In practice, that means separating the workflow that produces the artifact from the process that reviews it, as highlighted in Figure 2.

The realistic near-term model is governed collaboration between humans and AI. AI can prepare, summarize, scan, compare, monitor, draft and escalate. Humans define the objective, interpret the evidence, approve material outputs and remain accountable for capital allocation, client outcomes and fiduciary responsibility. In other words, the most useful early applications of agentic AI may be relatively unglamorous, with the value coming from speed and process quality.

A thematic equity team, for example, might use an AI-supported process to gather filings, market commentary and policy documents around a long-duration investment theme. The system could deduplicate sources, tag relevant material, draft a preliminary memo with factual claims linked to source documents and flag inconsistencies. The analyst would then spend more time challenging the thesis than assembling the file. The portfolio decision would still belong to a human. The key phrase is ‘governed delegation’: AI supports the process, but does not own the judgment.

AI capability is becoming part of manager due diligence. But investors should avoid rewarding the manager with the most impressive AI story. The goal is to distinguish real, repeatable, governed capability from isolated experimentation or marketing language. An example question list might look as follows:

• Where is AI used in your investment and operating process?

• Is each use case in production, pilot or experimentation?

• What decisions does AI influence?

• Does it draft, summarize, flag, recommend or execute?

• Where is human approval required?

• What data does the system use, and how are data privacy, licensing, lineage and retention controlled?

• How are outputs validated before use and monitored after deployment?

• Can you reconstruct the data, assumptions, outputs and human approvals behind a material AI-assisted process?

• Who owns AI governance inside the firm?

• What safeguards are in place for hallucination, bias, inaccurate reporting, data misuse and vendor dependence?

• What is the failure protocol if an AI-assisted process produces an error?

A manager saying "we use AI" is not enough: trust in the capability must come from the governed process around the model.

AI adoption in investment institutions is likely to unfold in stages. Institutions do not need to jump immediately to the most advanced stage; strong governance at stages 1 and 2 is what makes later adoption safer.

Stage 1: AI as assistant
Individuals use AI to summarize, draft, search, code, translate and prepare analysis. This is already becoming standard office practice. The governance focus is mainly data handling, acceptable use and basic review.

Stage 2: AI as workflow infrastructure
AI becomes embedded into repeatable processes, such as research preparation, manager monitoring, risk reporting, due diligence, policy review, client reporting and documentation. The governance question shifts from ‘what tool are you using?’ to ‘what process has been redesigned, and what controls sit around it?’

Stage 3: AI as a network of specialized assistants
More specialized AI systems begin to monitor markets, portfolios, managers, operational risks or governance exceptions. Each has a narrow role and clear permissions. Each escalates to humans when limits are reached. This stage is plausible, but should be built gradually and governed tightly.

Investors do not need to become AI engineers. But they do need enough AI literacy to ask better questions. Where is AI used? What decision does it influence? What data does it touch? What evidence supports the output? How stable is the conclusion? What is the system allowed to do? Who is accountable? What happens when it fails? These questions make responsible innovation possible.

Agentic AI is likely to become part of the operating fabric of asset management: helping teams gather evidence, monitor risks, process information and document decisions. The institutions that benefit most will not necessarily be those that adopt AI the fastest. They will be those that make controlled experimentation easier without weakening governance standards. The future of asset management is unlikely to be fully autonomous, but the more AI can do, the more clearly humans need to define what it is allowed to do