This is how we streamline AI implementation across 26 companies in a PE portfolio. Each company has 10 departments, and each department has 10 workflows. If you do them one at a time, you need "50,000 forward deployed engineers across three years." Our method is to group the companies by system of record. For finance, the 5 companies on NetSuite are one group. The 4 companies on Dynamics are a different group. Then you find what NetSuite can do and what it cannot do. You build only for the gaps. In one group, the processes are still different. The 5 NetSuite companies used 12, 9, 15, 18 and 13 steps for their finance process. The target is 6 steps, with agents at specific points. You will not often make all 5 exactly equal. You can come close if you map each process with the stakeholders before you change it. A single company or 26, sort the work by its system before you build.

AI agents should add to the systems your client already uses. One client spent several years and $10 million on one ERP migration. Offer that client a move to an AI-native CRM, and "you've lost them." So we build our agents inside the systems of record. They read the data in Salesforce and do the work in Salesforce. When a human must be in the loop, the agent sends a message in Slack. The staff need no new training, so they use the agents faster. SMB clients with no CRM are the exception. You can put them on one. If you want to work as a forward deployed engineer, build on what your client already uses.

I believe vibe manufacturing will create thousands of companies that make $1M or more a year. In 2024, Lovable and Bolt made anyone a vibe coder. Now the same shift happens for physical products, and OpenAI's Astra was the first model of this era. Before, a hardware product meant your team needed a full-time 3D designer, a mechanical engineer, and an electrical engineer. Now you pay $20 a month for a frontier model to walk you through all of it. The market for your product can be small. If it costs little to make, there is a business. Hardware looked too hard to me. I did not know it, so I did not try. Now I have a hardware desk on the way. Make one thing: a printable object, a CAD file, a Blender model, or a circuit board. You owe it to yourself.

How does an FDE audit a real workflow? We did one for a $5B public software company with 150+ products. Step 1: Read their process document. - Theirs had 7 steps: build the quote, submit, Deal Desk, approve, send, negotiate, sign. Step 2: Interview more than one person. - One person can tell you the wrong story. Step 3: Put process mining agents on the CRM. The agents found 20 steps and 7 loops. - 61% of requests go from step four or five back to step one. - Legal sends the request back 12% of the time. - 30% of the time, a new quote must get approval again. Step 4: Map the real workflow. - 4 of our deployed engineers did this. - The CRO and CFOs told us, "I feel like you understand our department better than we do." Before this map, no one in the department had seen the workflow this clearly. Do this audit before you use AI, so you know which process is broken.

How does an FDE map the processes to deploy AI in a real business? We use 3 sources. Source 1 is people. - Companies tell us: "this person's been at the firm for 20 years and they just handle it." - We hear this from Fortune 50 companies and from SMBs. - So in finance, you talk to the head of AP, AR, reconciliations, billing, banking, and FP&A. You make a "human API." Source 2 is the systems of record. - With real-time access to Salesforce for 3 to 4 weeks, you see what data comes in and how often people correct it. Source 3 is the documents in SharePoint, Drive, Notion, Slack, Teams, and Gmail. If you use the system data without the interviews, you miss half of the picture. The split is different for each company. A large company can have 10 years of Salesforce data. An SMB has most of it in people's heads. Choose your angle for each company before you build.