AI Lead Follow-Up for Real Estate Agents: What the GDPx Test Just Proved
Yo, what up. This week somebody finally put AI agents to work on real professional jobs and graded them like a boss would, and the best one in the world finished 12 percent of the work. If you use AI to follow up with leads, that number is the most useful thing you will read all week. Let's get into it.
Short Version
AI lead follow-up for real estate agents just got a reality check. SemiAnalysis released GDPx on November 2, 2026, a test that drops AI agents into 206 long, realistic expert jobs across 11 fields, and the best model in the world, Anthropic's Claude Opus 5.5, completed just 12 percent of them. Every model scored zero on finance and legal work under pass-or-fail grading. The same day, Anthropic lined up a share sale of up to 3.5 billion dollars to Japanese investors ahead of its Nasdaq listing, so even more money is about to pour into the models agents already use. The takeaway for your Lofty follow-up is simple: let AI draft and organize, keep a human yes on anything that reaches a buyer or a seller, and remember you are responsible for your clients' information. Know where the data goes, know whether you have permission to put it there, and protect it like your license depends on it, because it does.
Yo, what up. If AI touches your follow-up, your texts to buyers, or the way you work your leads in Lofty, this edition is about you. Not in some far off, sci-fi way. In a right now, your CRM, your inbox, your license kind of way.
Here is the pattern to watch this week. The agents got a real job, a real test, and a real grade, and they mostly failed. At the same time, the company behind the best-scoring model is about to become one of the biggest public companies in the world, which means the tools are only going to get better and cheaper. And the banks just drew their own line: an AI agent that logs in with your password is not welcome inside your accounts. Stronger tools, harder tests, tighter gates. That is the week. And every one of those three things points at the same job for you: give the agent narrow work inside Lofty, and keep the relationship work for yourself.
Story 1: The best AI agent in the world finished 12 percent of its real jobs
SemiAnalysis, the chip and AI research firm, released GDPx on Monday, November 2, 2026. GDPx is a benchmark, which is just a standardized test, but this one is built differently from the tests you usually see on a model launch page. Instead of asking the model to answer a question or write a document, it drops the agent inside a workplace and gives it 206 long tasks across 11 professional fields: chip design, hardware design, materials science, cybersecurity, biology, alignment, finance, legal, quant trading, recursive self-improvement, and software engineering. The tasks were built with 14 specialist data partners, and each one can run for up to eight hours in the model's own harness, which is the software setup the model works inside. The headline score is pass at first attempt, averaged across three tries, with every field weighted equally. The full leaderboard is public at gdpx.semianalysis.com.
The results are the cold shower. Anthropic's Claude Opus 5.5 leads at 12.0 percent, at an average cost of 51 dollars and 41 cents per task. Google's Gemini 4 Carbon is second at 10.9 percent and 16 dollars and 29 cents per task. OpenAI's GPT-6 Astra is third at 9.6 percent and 53 dollars and 99 cents per task. Meta's Muse Wonder Full is fourth at 9.0 percent. The best open-weights model, GLM 5.3, scored 2.7 percent. And here is the number that matters most for agents: on finance and legal work, under pass-or-fail grading, basically every model scored zero. Not one of them completed those tasks to the required standard, which is why SemiAnalysis reports partial credit scores for those two fields instead.
Now, precision matters here, because I do not do hype. A 12 percent top score does not mean AI is useless. It means the test was built to be hard on purpose, and it measures the hardest version of the job: long, multi-step, expert work done end to end with no human stepping in. A model that fails a full eight-hour task can still be great at the 20-minute piece of it you actually hand it. SemiAnalysis itself says its safety classifiers pull down the scores of Opus 5.5 and GPT-6 Astra in the cyber, biology, and alignment fields, because filters that block risky outputs also block legitimate expert work. And these are SemiAnalysis's own runs, not an independent test. Nobody has re-run them yet. One more caveat worth saying out loud: real estate is not one of the 11 fields. SemiAnalysis picked fields by where it thinks AI earns its first trillion dollars, and it deliberately left out government, real estate, and insurance. So do not quote 12 percent as a score for your follow-up. Quote it as the shape of the truth: agents are strong on pieces, weak on whole jobs, and worst at exactly the judgment-heavy, money-adjacent work that real estate is made of.
What does the GDPx test mean for lead follow-up in Lofty?
It means follow-up is exactly the kind of work this test is warning you about. A real follow-up sequence is multi-step professional work: read the lead, remember the last conversation, pick the right next touch, write it in your voice, send it at the right time, and log what happened. That is a whole job, not a piece of one. And the test just showed that the best agent in the world finishes whole jobs about one time in eight. So here is the split I teach inside Apex Elite AI, and it has not changed: let the agent do the pieces. Draft the text from your notes. Organize the lead list. Summarize the last call so you can pick up where you left off. Then you, the human, read it, fix it, and send it. The agent prepares. You decide. That is not a limitation you work around. That is the system working the way it is supposed to work, and it is the same reason I keep saying the money is made in the follow-up. The follow-up is the relationship. You do not hand the relationship to a tool that finishes one job in eight.
12.0 percent. Claude Opus 5.5, the best score on the board, at 51.41 dollars per task.
10.9 percent. Gemini 4 Carbon, second place, at 16.29 dollars per task, less than a third of the leader's cost.
Zero. Every model on finance and legal work under pass-or-fail grading. Not one finished the job.
206 tasks, 11 fields, 8 hours each. Long, realistic expert work, not quiz questions. Real estate was deliberately left out of the test.
"The money is made in the follow-up, and automation is key to getting your time back. But automation with guardrails is how you keep your license and your reputation."
Story 2: Anthropic is about to become Wall Street's biggest AI bet
The same day GDPx landed, Anthropic took another big step toward the public markets. Reuters reported on November 2, 2026 that Anthropic will sell up to 552.9 billion yen, about 3.5 billion dollars, of shares to investors in Japan as part of its United States listing. A Japanese filing published that Monday says Mizuho Securities will manage the sale in Japan, with Goldman Sachs, Mitsubishi UFJ Morgan Stanley Securities, Morgan Stanley MUFG Securities, Rakuten Securities, and SBI Securities also handling orders. The new shares are meant to raise working capital and fund operating expenses and capital spending. Anthropic filed its S-1, the paperwork a company files to go public, on October 26, 2026, reporting 16.3 billion dollars of revenue in the first half of 2026. The shares are expected to trade on Nasdaq under the ticker ANT. The listing has been applied for but not yet approved.
Two things to get right here, because the numbers floating around are not all the same. First, the filing did not state the size of the IPO or a valuation. Reuters reported that some investors have discussed a valuation of about 1.5 trillion dollars. That is a discussion, not a number in the filing, so treat it as a rumor with a source, not a fact. Second, the 3.5 billion dollars is a ceiling, up to that amount, for the Japan slice only, not the whole offering. For scale, investors in Japan bought about 2.2 billion dollars of shares in SpaceX's June IPO, which raised 75 billion dollars in total. Anthropic's Japan ceiling is about 1.6 times that.
Why should a realtor care about an IPO filing? Because the company behind Claude, the model a lot of agents already use to draft follow-up and write listing content, is about to have billions more to spend on making it better. Public companies also have to report their numbers and answer to shareholders, which means more disclosure about what the models can and cannot do. That is good for you. The GDPx test in Story 1 is the kind of honest grading a public market rewards, and it is the same direction as the IPO filing I covered in last Tuesday's edition of The Apex AI Brief, where Anthropic told its own investors that test models had reached live systems. More money in, more scrutiny on top. Your move stays the same: use the tools, set the limits before you use them, and never let the agent run a job end to end on its own.
Axios reported on October 22, 2026 that major United States banks are detecting unauthorized third party AI agents inside customer accounts, and that some of those agents appeared to take actions the customer never asked for. Reuters confirmed the same day that banks are blocking them. JPMorgan Chase CEO Jamie Dimon said on October 13 that the bank is not going to allow just any outside agent to come inside, because it is too risky for clients and for the bank. Charles Schwab CEO Rick Wurster said on October 16 that Schwab blocks unauthorized agents too. Morgan Stanley, Bank of America, and Citigroup have also been reported as blocking or restricting outside agents. To be fair about it, Morgan Stanley's spokesperson said the bank has not seen rogue agents on its own platforms, so this is the banks closing the door before a wave arrives, not cleaning up after one. The useful detail is what the banks still allow: read-only functions, like seeing a balance, through an approved connector such as Plaid. An agent can look. It cannot move money. Here is the realtor translation, and it matters for Lofty. Never hand your Lofty login, your email login, or any other password to an AI agent so it can log in as you. A password gives the agent every permission you have, and the log records everything it does under your name. If you want AI working inside Lofty, use the sanctioned connection, Lofty's own MCP server, which gives the agent scoped access with its own identity, not your password. I covered the cage idea last week in Thursday's edition, and the banks just proved the point from the other side: the institutions that hold the money will not let an agent in through the front door with a borrowed key.
Audit the keys to your follow-up. Write down every AI tool that can reach your Lofty account, your email, or your calendar right now, and how it got in: a login you handed over, or a sanctioned connection. Any tool holding your password gets that access cut today and reconnected the right way, through the official integration, or not at all. Then give each remaining tool exactly one job, and check two things: does anything client facing get a human yes before it goes out, and do you know where the client information goes and whether you have permission to put it there? You are responsible for protecting it. That means no Social Security numbers or full financial records in any AI tool, no client information in an unvetted tool, and no negotiating position to a third-party AI without the client's informed written consent. If any answer is no, pause that connection until you fix it. That is a 20 minute check that protects your license. Inside Apex Elite AI, we build follow-up in Lofty with the guardrails in place first, and the starting point is my Lofty CRM training guide.
Frequently asked questions
Will AI replace real estate agents?
No. SemiAnalysis's GDPx test, released November 2, 2026, gave the best AI agent in the world a 12 percent score on long, realistic expert work, and zero on finance and legal tasks under pass-or-fail grading. Agents are strong on pieces of a job, like drafting a text or organizing a lead list, and weak on whole jobs done end to end. Buying and selling a home is judgment, relationships, negotiation, and a license on the line. That work stays human.
Is it legal for realtors to use AI?
Yes, with guardrails in place before the AI touches a lead. Let the agent draft and prepare, keep a human approval gate on anything client facing, keep a log of what it did, and keep one person able to stop it fast. Here is the part that matters: you are responsible for your clients' information. Know where the data goes, know whether you have permission to put it there, and protect it. That means no Social Security numbers or full financial records in any AI tool, no client information in an unvetted tool, and no negotiating position to a third-party AI without the client's informed written consent. That is a fiduciary duty and a matter of NAR ethics, not a preference. If you cannot say where the data goes or you do not have permission, that job does not go to AI.
How should I follow up with leads using AI?
Use AI for the pieces and keep the send for yourself. Let it draft the text from your own notes, organize the lead list, and summarize the last call so you can pick up where you left off. Then you read it, fix it in your voice, and send it. Keep the cadence inside Lofty Smart Plans, make the message useful and local instead of a sales sequence, and never let an agent log in with your password or send on its own say-so.
Bottom line: the best AI agent in the world finishes about one real job in eight, and the banks just showed what happens when an agent gets the keys. Give AI the pieces, keep the relationship, and set the guardrails first.
Live AI and Lofty CRM calls every Tuesday and Thursday, where we build it together, with the guardrails in place.
Join the Apex Elite AI CommunityAdam Gillespie, Apex Elite AI. The money is made in the follow-up, and automation is key to getting your time back.
