Manage the agents, not the prompts

The asset manager with a pitch document due in three weeks doesn’t need a prompt library. They need management skills for AI: the ability to brief an agent the way they’d brief a graduate hire, check the output the way they’d check a draft investor report, and know when to pull work back in-house. Most of the training offered to property teams right now is technical. Here’s the model, here’s the prompt structure, here’s the temperature setting. But the bottleneck isn’t knowing how to talk to the machine. It’s knowing what to ask for, how to evaluate what comes back, and when to override it.

Ethan Mollick at Wharton frames this clearly in Management as AI superpower: the people who’ll get the most from AI are the ones who already know how to delegate, review, and course-correct. The technology is the easy bit. The hard bit is recognising when the agent has drifted off brief, the same way you’d spot when a consultant has misread the asset.

What management skills for AI actually mean in practice

Management skills for AI are the same ones you’d use with a junior analyst. You need to write a tight brief, not a vague instruction. You need to check the work against the brief, not just accept it because it looks polished. You need to know when the output is good enough to ship and when it needs another pass. And you need to recognise when the task requires judgement the agent doesn’t have.

When I run capability sessions with portfolio teams, I don’t start with the tools. I start with a real task: a brochure for an empty unit on the second floor, a briefing note for a planning committee, a positioning deck for a partnership launch. Then I work through how you’d brief it, what you’d check, and where you’d intervene. The model matters less than the discipline.

Why do property teams struggle with delegation to agents?

Most property professionals haven’t managed people at scale. An asset manager might run a portfolio worth £200 million but only directly manage two people. They’re used to doing the work themselves or outsourcing whole chunks to a consultant. AI asks for something in between: supervision without micromanagement. That’s a learned skill, and most people don’t have the reps.

BCG’s research on AI and productivity found that the highest performers weren’t the most technical users. They were the ones who knew when to trust the output and when to step in. In property, that means recognising when the agent has written serviceable copy for a hoarding but missed the planning constraint that matters, or when it’s drafted a valuation summary that’s technically accurate but tonally wrong for the investor.

What should a brief for an agent look like?

A good brief for an agent looks like a good brief for a person. It includes context (this is a stalled scheme, the project is going to public consultation, the opposition is organised), constraints (two sides of A4, aimed at residents not planners, avoids jargon), and a clear outcome (the audience needs to understand the economic case without feeling talked down to). Vague instructions produce vague output. Tight briefs produce usable drafts.

I document this as part of every handover. If an asset team is going to run their own content rhythm after I’ve left, they need to know how I brief the tools, what I check, and where I override. That’s not a prompt library. It’s a quality control process, written down so the next associate can follow it without guessing.

How do you evaluate output when you’re not the expert?

The hardest part is evaluating work outside your domain. If you’re an asset manager reviewing agent-generated copy, you’re not a copywriter. If you’re a finance director reviewing a draft investor memo, you might not know what good looks like. The skill is knowing what questions to ask. Does this match the brief? Does it contradict anything known to be true? Does it sound like something the business would actually say?

Stanford HAI’s work on human-AI collaboration emphasises this: effective oversight doesn’t require deep technical knowledge. It requires a clear mental model of what the task is for, who the audience is, and what failure looks like. In property, that means checking whether the agent has understood the asset, not just whether it’s written grammatically correct sentences.

When should you pull work back in-house?

Sometimes the agent can’t do the task, and the skill is recognising that early. If the output keeps missing the mark after three rounds, it’s not a prompting problem. It’s a judgement problem. The task requires lived experience the agent doesn’t have: reading a room at a planning committee, knowing which investor cares about ESG and which cares about yield, sensing when a seller is bluffing.

I see this with positioning work. An agent can draft ten versions of a tagline, but it can’t tell you which one will land with a particular buyer. That’s pattern recognition from years of investor calls and completed transactions. AI arms teams, it doesn’t replace them. The people who thrive are the ones who know where the line is.

What does good capability transfer look like?

When I hand over a brand system or a content platform, the training isn’t about the content management system. It’s about the decisions embedded in the system. Why this tone, not that one. Why I brief photography this way. Why the signage hierarchy works like this. The asset team needs to be able to make the next decision without me, and that means understanding the logic, not just the files.

I engineer the handover from day one. Documentation sits alongside delivery: file sources, fonts, content rules, and the reasoning behind them. If the engagement is a pitch document, I don’t just write the deck. I show the team how I’d brief an agent to draft the next one, what I’d check, and when I’d rewrite it myself. That’s capability transfer, not just delivery.

Why AI literacy isn’t about the tools

AI literacy for asset managers isn’t learning which model to use. It’s understanding what the models are good at and where they fail. They’re excellent at structured tasks with clear constraints: summarising lease terms, drafting templated content, reformatting data. They’re poor at tasks requiring context they don’t have: reading political risk in a planning process, knowing which comparable is actually comparable, sensing when a brief is incomplete.

The property teams getting value from AI in 2025 are the ones treating it like a capable but inexperienced hire. They’re delegating real tasks, not just experimenting. They’re building feedback loops, not hoarding prompts. And they’re training their people in management skills for AI, not technical skills they’ll never use. The bottleneck isn’t the technology. It’s knowing how to manage it.

If you’re running a portfolio team or an investment office and want to build this capability without the noise, send a paragraph. I’ll write back.