How to trust automation in real estate workflows

I built an AI agent workflow that automatically underwrites real estate investment pitch decks. Here's how it works, the current pitfalls, and how I'd improve it: ➡️Trigger: PDF file is uploaded to Google Drive folder. ➡️Data Extraction: PDF file is passed to MistralOCR, which is one of the better AI tools for reading pdf files and returning accurate text. ➡️Analyzing: Markdown text is passed from Mistral to an openAI model, along with instructions on: -Who I want the openAI model to be (senior investment analyst) -What input I am giving them (pitch deck, formatted in markup text language) -What the ai's job is (read the text, underwrite the deal, and determine if it passes initial screening). -How I want it to output the data (High level summary in text format) ➡️Outputs: It identified the following takeaways: Key Strengths: (attractive price, well‐documented sources/uses of funds, strong projected returns) Key Risks: The absence of onsite management, aging physical assets with significant vacant units, and the need to overcome sub‑market rents ➡️How I'd improve the workflow: -Create a trigger and filter in Gmail, so this workflow runs automatically when I'm sent deals -Include a formatting step to organize the output from Mistral before sending this to openAI, to improve accuracy and lower usage costs -Plugging in a more powerful model (o3-mini vs 4o-mini), or adding on another AI to check the work of the first AI -Better prompting -Connect a database that the AI can refer to for context, and add future outputs to

The hardest part of building Agentic SaaS isn’t the model, the agent, or the workflow. It’s trust! LLMs and agent frameworks will commoditize. Workflow design will stop being the differentiator. What will separate the Agentic SaaS winners from the rest? Trust. Adoption won’t come from capability. It will come from belief that the software will act - consistently (e.g., a reconciliation agent that always balances invoices correctly), invisibly (e.g., a support agent that escalates only when necessary) and ethically (e.g., a recruiting agent that never screens out candidates based on gender or ethnicity). Think of it like autopilot in aviation. Pilots don’t trust it because it’s a clever software - they trust it because it’s reliable, bounded, and accountable. That’s the bar for Agentic SaaS! For Agentic SaaS leaders, designing software that inspires confidence at every step means focusing on three pillars of trust: 1\ Accountability before action: Users don’t need agents to be always right. They need to know that when in doubt, the agent will surface decisions for review instead of bluffing through. How: Flag uncertainty, escalate edge cases for human review, and leave an auditable trail. 2\ Transparency after action: Trust comes from knowing you could intervene if needed. How: Make actions reviewable, explainable, and interruptible. Offer clear “undo/override” options and maintain an audit log. 3\ Set clear boundaries: Trust is earned more by what an agent refuses to do than by what it can do. How: Show upfront what data the agent can access, block sensitive information by default, and explicitly define actions the agent will never take (e.g., sending messages or making changes without approval). Designing for trust isn’t easy, and even the best teams are still figuring it out. But it’s worth prioritizing. Models will get cheaper and workflows will get replicated, but the strategic moat for breakout Agentic SaaS companies will be trust! Features can be copied, but trust can’t. Disclaimer: Views personal.

A procurement officer told me, “Automation means losing control.” ↳ What happened next changed her mind forever. I smiled when she said it. ↳ Because years ago, I believed the same thing. Back at Henkel, I feared automation. I thought agents meant giving up judgment. I thought speed meant losing precision. I thought machines would dilute negotiation craft. But then reality punched harder than fear. ↳ Cycle times kept dragging. ↳ Compliance checks stayed painfully manual. ↳ Risk signals hid inside broken spreadsheets. Nothing felt in control. ↳ Everything felt reactive. So we deployed agentic workflows. And everything flipped overnight. Sourcing cycles collapsed from weeks to days. ↳ Supplier participation shot up instantly. Compliance signals surfaced automatically. Risk markers lit up in real time. I didn’t lose control. ↳ I gained command. For the first time, I saw the whole field: → every supplier → every compliance flag → every negotiation lever → every risk trend → every demand shift All in one stream. ↳ No blind spots. ↳ No lag. ↳ No guesswork. The real breakthrough came later — governance. When ethics and guardrails were embedded, ↳ automation stopped being a black box and became a glass box. AI ethics audits delivered 2x ROI. Every decision had a traceable trail. Every supplier had compliance benchmarks. Every risk had quantified impact. Speed became trust. Trust became leverage. That’s when she saw what I had learned the hard way: ↳ automation doesn’t remove control. Automation removes noise. Automation removes blind spots. Automation removes slow decisions. And judgment? It doesn’t disappear. ↳ It gets amplified. Procurement’s next era isn’t digital for show. ↳ It’s decisive by design. She walked in fearing she’d lose control. She walked out realizing she finally had it. Follow Martin B. for no-fluff, real-world insights on AI, Procurement, and IT collision.

Copilot Cowork is the most useful Copilot feature. I used it for a portfolio (20 assets) re-underwriting - used to take weeks of analyst time! That’s not hype - Cowork did it well enough that I think every investment team should pay attention. The workflow was very simple: 1) I uploaded 20 DCF spreadsheets for 20 buildings representing assets supposedly bought in 2024 (not a real portfolio). 2) I asked Cowork to update the models to the current date and project returns and cash flows for the next 10 years. I instructed Copilot to assume tenants had extended at prevailing market rents and to extract any required data from market reports. 3) Once the spreadsheets were updated, I asked to summarise the portfolio and its key characteristics. 4) Then I asked for a report suitable for an investment committee. 5) Finally, we iterated until the output had good-looking charts and graphs. The whole process took about 15min. The report it produced and a sample DCF I used as input are in the comments below. What makes it more interesting is that this was not just spreadsheet automation. The agent pulled market evidence from reports, applied it, updated the models, summarised the results and turned them into a presentation-style output. The key point is this: Copilot is beginning to handle long, mechanical, multi-step workflows of the kind that sit right at the centre of real estate analysis. Importantly, the human still stays in control (and needs to audit) but the time saving is considerable. You can inspect the assumptions, review the outputs, question the process and decide what is good enough to use. That is exactly how this technology should be used. Not “replace the analyst”. More like: let AI do the heavy lifting, and let the human do the judging. For real estate, that is a big shift. #AI #RealEstate #DCF #PortfolioManagement #InvestmentCommittee #PropTech #AssetManagement #Automation

Everyone’s calling their 𝗔𝗜 𝘁𝗼𝗼𝗹𝘀 “𝗮𝗴𝗲𝗻𝘁𝘀” 𝗶𝗻 𝟮𝟬𝟮𝟱. 98% of them? Just glorified workflows. The difference isn’t just semantic, it’s the key to building real trust in automation. Let’s take a concrete example: 𝗦𝗮𝗹𝗲𝘀 𝗙𝗼𝗹𝗹𝗼𝘄-𝗨𝗽 𝟭. 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄 → When a lead fills out a form, automatically send a thank-you email and assign an SDR. Reliable. Rule-based. 𝟮. 𝗔𝗜 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄 → Same process, but the AI personalizes the thank-you message. Still a predictable sequence; AI adds polish. 𝟯. 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄 → AI analyzes lead data, drafts a personalized outreach, posts the draft in Slack for human review, and updates the CRM. Mix of reasoning plus structured oversight. 𝟰. 𝗔𝗴𝗲𝗻𝘁 → AI researches the lead, writes outreach, follows up, books meetings, and updates the CRM on its own. Full autonomy with minimal human hand-holding. Let’s clear the air: agents make judgments and adapt on the fly. Workflows? They’re reliable, predictable, and follow a set path. Most so-called “agents” out there? Workflows in disguise. Solid, but inflexible. Real agents are rare. They think and adapt, and they’re not always predictable. The sweet spot? “𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀.” Combine the reliability of workflows with the smarts of true agents. Know the difference. It’s not just jargon; it’s about setting the right expectations for AI. Build trust in automation by calling things what they really are.

Is messy OCR quietly burning your agent ROI? Your agents are not “acting weird.”   They are acting on what they see.  If your OCR scrapes the page like a blunt tool, small misreads slip into core systems. At scale, that turns smart agents into fast, confident interns who move money and update records on shaky ground. The cost is not just rework. It is brand, trust, and operational drag.  Three levers to pull now:  1) Treat OCR as a core system, not plumbing.   Evaluate it like you would a CRM. Match engines to your real document mix, layout complexity, languages, and tolerance for mistakes. Search-grade text is rarely safe for action-grade workflows.  2) Wire confidence scores into your logic.   Modern OCR knows how sure it is per field. Use that. Auto execute only on high confidence. Send mid confidence to lightweight checks. Route low confidence into human review or safer paths. Speed where you can, brakes where you must.  3) Structure content before decisions.   Do not feed agents raw text blobs. Normalize into clear fields like “supplier,” “amount,” “due date,” and classify document types before any decision. Agents then operate inside a system, not a pile of notes.  If you looked at your current agent stack today, would you still blame the model, or finally inspect what your agents are actually seeing? #AI, #IntelligentAutomation, #DocumentProcessing, #RefreshWithRyza, #RiskManagement

“How do I trust an AI agent to run part of my business?” Great question. Founders aren’t short on ideas. They’re short on confidence that the assistant will actually do what they meant. Not just what they said. That’s why I’ve been studying Anthropic’s latest framework on building safe, trustworthy agents. Here’s how we apply it inside the Lab: 1. Start with control before autonomy Don’t give your agent the keys on Day 1. - Let it observe. - Let it assist. Then, layer in autonomy where the risk is low and the logic is proven. 2. Make the agent explain itself If your AI assistant recommends canceling a subscription or skipping a step, it should tell you why. We build in: - To-do lists. - Reasoning prompts. - Checkpoints. ... before any action is taken. 3. Align with your actual values An agent will do exactly what you ask. Which is the problem. So, we give Claude your company principles, red lines, and definitions of success ... BEFORE we ask it to decide. 4. Guard your context Claude can remember a lot (sometimes, too much). That’s why we use MCP to: - Restrict access. - Separate workflows. And prevent spillover between tools and projects. 5. Protect against manipulation AI agents can be tricked into misbehaving. (especially if they’re reading inputs from email, chat, or the web) Anthropic uses layered threat detection. We mimic that inside the Lab, with scoped tasks and read/write boundaries from the start. Bottom line: An AI agent that works fast but makes the wrong decisions is a liability. An AI agent that understands your: - Logic. - Your values. - And your boundaries. ... is an asset. That’s what we’re building inside the Lab. Agents you can trust to: Think clearly. Act carefully. And reflect on how your business works. Because until you trust the assistant, you won’t let it do the work. Thoughts?

Most real estate agents I know are drowning in marketing tasks that don't move the needle. They manually post content across 5 platforms, spend hours in their CRM updating lead statuses, and customize the same follow-up emails repeatedly. All while wondering why their pipeline isn't growing. Here's what I tell my real estate clients: Your time is too valuable to waste on repetitive marketing tasks that could be automated. The difference between agents who scale and those who struggle isn't how hard they work, it's how intelligently they build systems that work for them. I've helped brokerages implement workflow automations (using tools like Make.com or Zapier) that completely transform their lead generation by handling three key areas: 1. Content multiplication: Build one workflow that takes a single market update or listing and automatically transforms it into multiple formats, Instagram carousel, LinkedIn post, email newsletter, and website blog. One creation, four channels, zero additional effort. 2. Lead qualification and routing: Create intelligent paths for new leads based on their behavior. When someone submits a form on your site, automation can instantly segment them based on price point, buying timeline, or neighborhood interest, then trigger the perfect follow-up sequence. 3. Client journey management: Set up workflows that track transaction milestones and automatically send updates, gather feedback, or request referrals at the perfect moment. This maintains the relationship without requiring your constant attention. I implemented these automations for an agent who saw their lead-to-appointment ratio improve by 37% in just 45 days, not because they generated more leads, but because no lead fell through the cracks. The real estate agents who win in today's market aren't always working 80-hour weeks (the work is still needed, don't get me wrong). They're building intelligent systems that handle the repetitive work, so they can focus on what truly matters: building relationships and closing deals. What marketing task is currently stealing too much of your time? I'd be curious to know what you're trying to automate first.

The AI arms race is about to produce the most expensive wave of technical debt the industry has ever seen. Most people won’t admit they’re adding to the problem. From the outside, it just looks like progress. Copilots shipping. Automations running. "AI-powered" on every landing page. But underneath, these are probabilistic systems built on top of workflows that are already fragile. The failure mode is different now too. In the past, broken systems caused problems inside the company. You would notice, fix them, and move on. Now they fail in front of the customer. Visibly. Trust is hard to rebuild, especially in established industries where it took years to earn. Sequencing is the whole game. Here's the actual order of operations: 𝟭. 𝗠𝗮𝗽 𝗮𝗻𝗱 𝗶𝗻𝘀𝘁𝗿𝘂𝗺𝗲𝗻𝘁 𝗼𝗻𝗲 𝗱𝗲𝘁𝗲𝗿𝗺𝗶𝗻𝗶𝘀𝘁𝗶𝗰 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄 𝗳𝗶𝗿𝘀𝘁. Pick something that must return the same output every time: a quote, invoice, compliance check, or scheduling rule. Get it right, reliably. That's your trust anchor. 𝟮. 𝗠𝗮𝗸𝗲 𝗱𝗮𝘁𝗮 𝗲𝘅𝗽𝗹𝗶𝗰𝗶𝘁 𝗯𝗲𝗳𝗼𝗿𝗲 𝗮𝗻𝘆𝘁𝗵𝗶𝗻𝗴 𝗲𝗹𝘀𝗲 𝘁𝗼𝘂𝗰𝗵𝗲𝘀 𝗶𝘁. Define your inputs, data owners, and validation rules clearly. Don’t make assumptions. This isn’t just for good practice; the data from your reliable workflows trains your AI. If you skip this, your AI will be built on bad data. 𝟯. 𝗔𝗱𝗱 𝗽𝗿𝗼𝗯𝗮𝗯𝗶𝗹𝗶𝘀𝘁𝗶𝗰 𝗔𝗜 𝗼𝗻𝗹𝘆 𝘄𝗵𝗲𝗿𝗲 𝗮 𝘄𝗿𝗼𝗻𝗴 𝗮𝗻𝘀𝘄𝗲𝗿 𝗶𝘀 𝗿𝗲𝗰𝗼𝘃𝗲𝗿𝗮𝗯𝗹𝗲. Use AI for things like search relevance, anomaly alerts, forecasting, or recommendations; these are safer areas. Avoid using AI first for billing, compliance, or scheduling, since mistakes there are harder to fix. 𝟰. 𝗨𝘀𝗲 𝗮 𝗵𝗮𝗿𝗱 𝘁𝗿𝘂𝘀𝘁 𝗴𝗮𝘁𝗲. If a feature loses customer trust even once, send it back for human review—no exceptions. Trust takes years to build, but one mistake can destroy it. Operators who get this right will pull ahead quietly while everyone else is still debugging their copilot. If you want to see how I apply this across acquisitions, I laid out the full framework here >> https://t2m.io/4aDyNCpu For the love of the game 🏴☠️ ⚡️