AI News This Week
1. Anthropic Builds Claude for the Rest of Us On May 13 Anthropic launched Claude for Small Business: 15 ready-to-run agentic workflows wired into the tools small operators already live in, QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, and Microsoft 365.
2. Notion Turns Its Workspace Into a Hub for AI Agents On May 13 Notion launched its Developer Platform, letting external AI agents operate directly inside the workspace through a cloud runtime (Workers), an External Agent API, live database sync, and partner agents including Claude Code and Codex.
3. ICSC Brings Prop Tech Front & Center ICSC Las Vegas is underway this week, and for the first time the show carved out a dedicated proptech track, ICSC+PROPTECH and one of the highlighted vendors is an AI-powered video analytics firm. According to the Commercial Observer: retail owners aren't just browsing, they're buying, and security and logistics are the hottest categories.
4. A $4B Advisory Platform Just Launched an AI That Runs the CRE Back Office On May 14, Pegasus, a CRE advisory platform with $4B in AUM, and AI firm Fore Enterprise launched Fore Real, a platform that automates property tax appeals, insurance compliance, lease abstraction, and tenant communications. Pegasus trained it on data from 400+ properties across 37 states, the team claims 98-99% model accuracy, and they plan to open it to other firms.
This Week’s Studio Lesson: Learn to Prompt Smarter
This week we dropped an update to a foundational AI skill: Prompting. The way you prompt matters more the more you lean on AI, not less.
Everyone’s seen an output from a weak prompt: You tell AI to "research the Kansas City multifamily market," expect gold because the models are now so advanced, but you still get mush. The fix is a structure every strong prompt has: Task, persona, context and references.
The persona is one of the highest leverage components of prompting. One sentence, "You are an expert multifamily acquisitions analyst," narrows the vocabulary and logic the model uses, and the output quality jumps immediately. Specific instructions around the task are equally important - What do you actually want to see in this research? News? Stats? Actual deals in the market? The more you define what you’re actually looking for, the better the output you’re likely to get. Pair these components with context around how you’ll actually use the research, and references to research reports you’ve actually found useful, and the AI slop you would’ve received turns into the gold you were looking for.
But beyond the core prompting structure, there’s additional tactics you can use to improve your output from AI as you work with it as a thought partner. Here’s the tactic that got the biggest reaction from the crowd in this live lesson:
Once once you have the output you were looking for, tell AI "please reverse engineer a prompt I could have used at the start to get this same result." Test what comes back, and if it works, save it as a Skill or instructions in a Project. The next market analysis, deal review, or variance report starts from those instructions, and takes a fraction of the time compared to the first run.
None of this is rocket science. That's the point. For CRE professionals leaning on AI for underwriting, market research, or deal sourcing, the fundamentals are what separate a useful output from an hour of wasted time.
Watch the full session on learning to prompt in CRE AI Studio.
Confession: If someone says they “built an AI agent”, I immediately get skeptical. That might sound brash, but the more the hype builds around AI agents, the more it’s used to describe anything and everything AI-related.
But the reality is, we’re approaching the point where many off-the-shelf AI tools do legitimately have agentic capabilities, making the definition of an AI agent even murkier. So here’s a simplified explanation of what an AI agent actually is, without the hype, and minimal jargon.
You probably already understand a basic chatbot. You ask, it answers. Question in, answer out. The key change with an agent is that you hand it a goal and it works toward the finished result on its own, planning tasks, accessing tools, executing the tasks, checking its work in a loop until the job is done, and potentially moving on to the next job. Chat is question-to-answer. “An agent is goal-to-result” - Greg Isenberg.
AI agents typically have the same component parts. Once you wrap your head around this, it gets a bit less murky.
The model is the brain. Claude, GPT, Gemini, etc. You pick one, you probably aren’t building this.
The harness is the interface, it’s how you work with the agent. The harness could be Claude Code or Cowork, Manus, or even one that you’ve built yourself (though, most aren’t quite there yet.) It’s how you interact with the agent. For most people, the harness will determine the models you can use with it (you probably won’t use ChatGPT’s latest model in Claude Cowork.)
The tools are what it's allowed to access and use to get the job(s) done. Your inbox, your files, a spreadsheet, a database. You grant the access access to the tools so the agent can access information from them, and potentially execute new tasks within those tools.
And then there's context. What the agent knows about you, your firm, your deals, how you work, the tasks you want it to work on, what "good" looks like, what it should always do and should never do. The context is the instructions and materials you hand it to reference.
So the components are simple enough. But the capability is also important. In theory, an agent has all those components, but can also operate in an autonomous loop where it decides what to do next based on what just happened.
Now look at that list again. The model, the harness, the tools, most AI users select all three off a menu. If somebody says they’ve built an AI agent, that might just mean they gave some context files to Claude Code and labeled one of them “Agent Instructions”. Whether or not that means they’ve actually built an agent.. I’ll let you be the judge, because AI experts can’t seem to decide. Some say “yes”, some say that to build an agent, you actually need to build your own harness, but I’m not sure the distinction matters.
What does matter is the context. This is the component of AI agents that people typically author themselves. It’s likely what makes your agent different, and helps determine how it makes you more productive. It’s the instructions and references you provide to help the agent connect the dots in your business. None of that lives in the model.
So what is an agent? The answer might not be as impactful as you thought. What will actually be impactful is how you customize the agent to work for you.