AI Agents in Investment Teams: From Copilots to Coworkers

From Co-Pilots to Coworkers: What Changes When AI Starts Taking Action

McKinsey's recent work on agentic AI highlights the key transition: older generative systems mainly answered, summarized, or suggested. Agentic systems can do things, execute multistep tasks, trigger workflows, interact with other systems, and collaborate with other agents. That shift from response to action is more significant than many organizations realize.

A Co-Pilot helps a human do work. A coworker changes how the work itself is organized. In CRE, that distinction is profound.

A Co-Pilot might summarize an OM or draft a memo section. A coworker-level agent might ingest the email, classify the opportunity, extract structured data, check for duplicates against an existing pipeline, surface relevant comparables, flag missing evidence, and route the case into the next workflow stage, all before a human reviews the result.

McKinsey's framing of the shift from 'thought to action' captures this precisely. Once AI can act, not just advise, the surrounding workflow, governance, and accountability model must be rethought.

The distinctions matter for practical planning. A Co-Pilot operates inside a human's workflow: it assists with the task the human is already doing. An agent can pursue a multi step goal semi-autonomously, within defined boundaries. An agent system is a network of agents that can collaborate, hand off tasks to each other, and coordinate across workflow stages.

Most CRE firms today are at the Co-Pilot stage. They use AI to summarize, extract, or draft within a single step. The transition to agent systems represents a qualitative change, not just in what the technology can do, but in how the organization must manage it.

This is why McKinsey describes the agentic shift as an operating model change, not a tool upgrade. The management requirements are fundamentally different.

The Future of CRE Teams Is Agentic, Not Just Automated

For years, 'AI in real estate' mostly meant faster analysis, better dashboards, and more efficient document processing. That framing is already outdated. The more consequential shift is not from manual work to automation, but from human-only workflows to mixed human agent systems, a transition that HBR and McKinsey now describe as the rise of the 'agentic organization.'

When most real estate firms talk about AI adoption, they mean point solutions: a tool that reads PDFs, another that generates summaries, perhaps a chatbot for internal queries. Each tool automates a specific task. The underlying workflow, who does what, in what order, with what authority, remains unchanged.

This is understandable. Task-level automation is the easiest to implement and the least disruptive. But it also has diminishing returns. Once you have automated the extraction of a rent roll, you still face the same coordination challenges: routing output to the right analyst, ensuring consistency with prior analyses, tracking which version was used in the IC memo, and deciding when the data is reliable enough to act on.

The real transformation happens not when individual tasks get faster, but when the operating model itself changes, when AI agents become persistent participants in workflows, not just tools that humans invoke and dismiss. Recent work from HBR and McKinsey points in the same direction: organizations are beginning to treat AI agents less like tools and more like forms of digital labor that participate in work, not just support it.

The term 'agentic AI' describes systems that can pursue multi step goals with a degree of autonomy: gathering information, making intermediate decisions, escalating when confidence is low, and completing workflows that previously required a chain of human actions.

In CRE, most high value work is not one task. It is a chain: sourcing, intake, review, memo creation, diligence, approvals, handover, reporting, and action. Once AI can carry work across steps, instead of just assisting inside one step, the design question changes. The issue is no longer 'Where can we save time?' It becomes 'How should teams be structured when parts of execution are handled by agents, while humans supervise, decide, and intervene?'

McKinsey's research on agentic organizations suggests this is not a future scenario but an emerging reality. The firms that adapt first will not simply buy better software. They will redesign how work gets done.

That means investment professionals spending less time moving information across tools and more time defining decision thresholds, evaluating trade offs, and handling exceptions. Managers will be judged less on how much process they personally control and more on how effectively they design systems that combine people, data, and agents.

Why Every CRE Firm Will Need an Onboarding Plan for AI Agents

A recent HBR article made an important observation: the main challenge in adopting agentic AI is not primarily technological. It is about managing work. That is a critical reframe for CRE firms, many of which still approach AI as a software procurement problem rather than an operating model redesign.

Most firms introduce AI the way they introduce software: they buy it, configure it, train users on the interface, and hope for adoption. That model worked for static tools. It does not work for agents, because agents do not just respond to commands. They participate in workflows.

HBR's recent writing on onboarding AI agents and rethinking workplaces around them makes the case that agentic AI requires a different introduction model. If agents are going to carry work across steps, interact with data, and produce outputs that feed into decisions, they need to enter the organization with the same rigor applied to new team members.

That means defining scope, permissions, interfaces, escalation paths, and performance expectations, before the agent goes live.

The technical capabilities of AI systems improve rapidly. The organizational capabilities to absorb them do not. That asymmetry is the primary reason why many AI projects stall after the pilot phase.

The gap is not usually model quality. It is that the surrounding workflow was not redesigned to accommodate the agent. Data inputs are not structured. Review points are not defined. Handoff protocols do not exist. The agent can perform its task, but the organization cannot integrate the output.

This is a management problem, not a technology problem. And it requires a management solution: deliberate onboarding.

AI Agents Won't Replace Investment Teams, They'll Change What Good Teams Look Like

The public debate still frames AI as a replacement story: which jobs disappear, which functions shrink. HBR's recent writing on digital labor and McKinsey's work on the 'agentic organization' suggest a more useful lens, the real question is not whether teams survive, but how the definition of an effective team changes when part of the workforce is digital.

The replacement framing, 'Will AI take my job?', is emotionally understandable but strategically unhelpful. It reduces a structural transformation to a headcount question and prevents leaders from engaging with the more important problem: how to redesign teams so that humans and agents each work where they are strongest.

In CRE, good teams have historically been those that can absorb messy information, coordinate across functions, and still move with conviction. But once agents can handle large parts of information intake, standard synthesis, and process follow up, the premium shifts. Good teams become the ones that know how to divide labor intelligently between humans and machines.

HBR's writing on digital labor and talent redesign reinforces this: the organizations that will perform best are not those that eliminate roles fastest, but those that redesign roles most thoughtfully.

Replacement thinking leads to two equally poor outcomes. In cautious firms, it creates resistance: teams protect existing processes because they fear redundancy. In aggressive firms, it creates reckless automation: steps get handed to machines without adequate governance, and quality problems emerge months later.

The better framing is redesign. Which parts of a role are repetitive and well-structured enough for agents? Which parts require human judgment, contextual knowledge, or relationship intelligence? And how should handoffs between human and machine be governed?

McKinsey's work on the agentic organization makes this concrete: the future workforce is mixed, and the management challenge is orchestrating that mix, not eliminating one side of it.

How AI Agents Run 24/7 for Investment Teams, and What That Actually Means

MIT Sloan reports 35 percent of organizations have adopted AI agents, with another 44 percent planning to. But 'agents running 24/7' remains vague in most pitches. The concrete version: software that watches inboxes, datarooms, covenants and markets continuously, does the preparatory work autonomously, and escalates to humans at defined checkpoints. The value is in the watching and preparing. The discipline is in the escalating.

Every vendor now promises agents that 'work around the clock for you.' The phrase deserves suspicion, and then a precise answer, because behind the marketing sits a real architectural shift. Anthropic's widely used definition draws the line cleanly: workflows are systems where models and tools follow predefined paths, while agents are systems where the model 'dynamically directs its own processes and tool usage.' The practical difference for an investment team is between software that runs when you click, and software that notices things while you sleep.

The adoption numbers say this has left the lab: in the MIT Sloan Management Review and BCG survey cited by MIT Sloan, 35 percent of organizations had adopted AI agents and 44 percent planned to deploy them soon. What has not left the lab in most firms is a clear answer to the operational questions: what exactly runs continuously, what may it do on its own, and when must a human be woken.

This article gives the concrete version for an investment context: what a 24/7 agent layer actually does hour by hour, where its autonomy should end, and how to manage it as the piece of operations it now is.

Take the most common case: deal intake. A broker sends a teaser at 22:40. An intake agent watches the mailbox, recognizes the email as a deal, extracts the asset, location, area, rent and price into structured fields, checks the sender and property against the firm's history, has this seller shown us this asset before, at what price, and runs the result against the acquisition profile. By 22:43 the deal exists in the pipeline, pre-screened, with a profile match verdict and its sourcing history attached. Nobody was working. The Monday meeting starts from a prepared pipeline instead of an unread inbox.

The same pattern generalizes across the lifecycle. In due diligence, an agent watches the dataroom: new documents are indexed and extracted on arrival, the completeness checklist updates itself, and a new document that contradicts an already-reviewed one becomes a flagged finding rather than a buried surprise. In the hold phase, agents monitor rent rolls, covenant headroom and market signals continuously, so quarter-end reporting summarizes what the system already knew. The common thread: agents compress the time between something happening and the team knowing about it, from days to minutes.

Notice what is absent from these examples: the agent never bids, never answers the broker, never signs anything. It reads, structures, checks, drafts and flags. That boundary is not a temporary technical limitation. It is the design.