Masterclass: How FDEs make $1M/yr deploying AI agents
Voss argues AI can't be applied like a coat of paint: forward deployed engineers earn top pay by mapping real processes, re-engineering them, and deploying agents inside existing systems of record.
Watch on YouTube ↗Charlie's summary
- ▶ 5:05
Don't apply AI like a coat of paint: it takes process re-engineering, which is why AI roll-ups need forward deployed engineers
Voss says AI transformation companies buy people-heavy firms like accounting, IT and law practices, then embed engineers who map processes, systems of record and exceptions, and rebuild them with agents. Thrive has 35 engineers across 70 firms. The payoff is valuation: buy at $1 billion, lift margins, sell for 2, 4 or 8 billion. Voss notes it is far more involved than people assumed.
- ▶ 9:40
Map the process first with three sources: interviews, systems-of-record mining, and existing documents
Voss says companies rarely have a clean document to point an agent at. Information lives in people's heads, like the person who has been there 20 years. Interview department heads, then spend 3 to 4 weeks mining Salesforce, NetSuite or Workday to infer the real SOP, then read the documentation. Interviews alone miss half the picture. Greg calls the first step a human API. The split varies: large firms have years of data, SMBs mostly have people's heads.
- ▶ 13:53
The documented process is never the real one: a $5B company's simple quote flow was actually 20 steps with seven loops
A public software company's Deloitte-mapped process was build, submit, approve, send, negotiate, sign. Voss's team of four engineers found through process mining that it was a 20-step process with seven loops. 61% of requests looped back, legal sent 12% back, and 30% needed a new quote approval. Nobody had seen it so cleanly; the CRO and CFO said Voss's team understood their department better than they did. Mapping answers AI on what?
- ▶ 17:55
Sort every step into four buckets: delete, plain code, agentic, or human in the loop
Voss says each step of a mapped process belongs in one bucket. Delete steps that shouldn't exist in a post-AI world. Use plain code where rules apply and no judgment is needed. Use agents where judgment is needed and there is enough historical data, like building a quote. Keep humans on risky actions such as approval, negotiation, signing and payments. A human on payments guards against phishing invoices that an agent might pay.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.