Overview
Summary
Greg Isenberg sits down with Vas of Varick Agents for a 54-minute walkthrough of how forward deployed engineers actually put AI to work inside a company. The pay claim is the hook. Some FDEs are clearing $1M a year because they drive $5M, $10M, or $25M of value and take a slice. The method underneath that number is not a new app. It is process reengineering.
They start by listening. Interviews with the people doing the work, plus agents quietly reading the systems of record for a few weeks, produce a map that is almost always about three times longer than the process on paper. Every step then goes into one of four buckets: delete it, automate it with a plain rule, give it to an agent, or keep a human on it. A surprising share lands in the first bucket.
What survives gets built inside the tools the company already runs. Salesforce, NetSuite, Slack. Nobody has to learn a new app. When a human has to sign off, it is a Slack message. Most of the work does not need the most expensive model. They measure before and after, then come back months later and prove it.
The episode stays on real engagements, with details changed. A $5B public software company. An accounts payable overhaul that cut cost per invoice from $31 to $6, and collapsed 17 steps to 7 and six loops to one. A private equity portfolio of NetSuite companies served by one engagement pattern. A 60-person accounting firm run through the same four buckets as the $5B case. Vas closes with a five-day starter plan anyone can run on their own work before they hold the title.
Why This Matters
Frontier models are now a commodity buy. The edge Isenberg and Vas keep returning to is deployment: who can find the real workflow, decide where intelligence belongs, and leave a measurable before-and-after. Most pilots fail because they slap a license or a generic agent on a process nobody mapped.
That is also the AI roll-up thesis in the episode. Private equity already buys accounting firms, IT shops, and law practices that run on people and old software, then puts engineers inside to rebuild the work. The same loop is how a company makes itself AI-native without a rip-and-replace, and how an individual gets the reps that justify FDE rates.
The page is for two readers at once. Someone who wants the FDE job, and an operator who wants agents in the systems they already pay for. The five-day plan is the on-ramp. The four buckets are the filter.
Key Points
- The $1M FDE paycheck is a slice of value created. Isenberg's frame: if agents drive $5M, $10M, or $25M, the buyer will share the pie.
- Start by listening, not building. Interview the people doing the work and let agents read company data for weeks. The real process is almost always 3x the one on paper.
- Sort every step into four buckets: delete it, automate it with a simple rule, give it to an agent, or keep a human on it. A large share belongs in delete.
- Build inside the systems of record the company already uses. Salesforce, NetSuite, Slack. Human sign-off is a Slack message, not a new app.
- Use the cheapest model that gets the job done. Most of this work does not need the frontier model on every step.
- Baseline before you build, then come back months later and prove it. The episode's accounts payable case cut cost per invoice from $31 to $6.
- On that same payable map, 17 logged steps became 7 and six loops became one. Compression is the result of the sort, not of a smarter prompt.
- The same four buckets ran a $5B public software company and a 60-person accounting firm. Size changes the systems, not the method.
- Private equity is already running the play: buy a people-and-old-software firm, put engineers inside, lift margin, and the firm is worth more.
- Sell the outcome the buyer cares about. CFO, operator, and sponsor each want a different number. Prove it with before-and-after KPIs.
- A top FDE holds three skills at once: how the work actually runs, production software in the systems of record, and the AI layer (model choice, evals, rollback).
- Vas said they have seen zero demand so far for on-prem GPU clusters, even at banks and pharma. Sensitive-data buyers are still routing through Azure, Bedrock, and Vertex with no-train terms.
Quotable
Quotes are lightly cleaned from auto-generated captions and the source post. Wording may not be verbatim.
"If you're deploying agents and you're able to drive 5, 10, 25 million of value for companies, would they be willing to give you a slice of that pie? It turns out yes."
The pay claim is a share of measured value, not a salary band.
"They start by listening, not building. The real process is almost always 3x longer than the one on paper."
The map is the product of the first weeks. Building before that encodes the wrong workflow.
"Don't apply AI right. This is the reason why most AI pilots fail. They try to slap AI on the business."
A license rollout or a generic agent fails when it never learned the workflow.
"We find the real process. We measure the time, baseline all of the KPIs. We pick processes with owners. We sort every single step. We baseline before we build."
The closing playbook. Order is the point: map, measure, sort, then build.
"We build it once and we deploy it everywhere. And we measure it constantly."
The PE pattern in one line. One engagement shape, repeated across portfolio companies.
Concepts
The role
Forward deployed engineer
Someone whose job is getting a company to use AI inside the systems it already runs. Not a strategy deck. The work is embedded: learn the workflow, decide where intelligence belongs, ship it, and prove the delta.
Value slice, not salary theater
Isenberg's arithmetic: drive $10M of value and a cut of that is the $1M check. Some PE firms, he notes, give FDEs a share of the carry on the companies they transform. The rate follows proximity to the money.
Three skills in one person
Domain: how the work actually runs, including the exceptions. Systems: production software in Salesforce, NetSuite, Dynamics, or whatever the record lives in. AI layer: which model, what you can trust it with, evals, and a rollback when it takes the wrong action. Vas treats the person who holds all three as the rare end of the field.
The method
Human API before the agent
Interviews give the why. System-of-record mining gives the what. Existing docs give the official version. Together they produce every step, its volume, its wait, and its owner. The paper process is the short version.
Four buckets
Delete. Plain code for deterministic rules. Agent for the steps that need judgment inside a bounded task. Human for approvals, exceptions, and signatures. Delete first, or you automate a faster mess.
Systems of record, not a new app
Agents write back into the tools people already open. When a person has to approve, the handoff is Slack. Adoption dies when the workflow moves to a tool nobody asked for.
Cheapest model that works
Model choice is a cost and reliability decision, not a brand decision. Most steps in these maps do not need the most expensive model. Evals decide what you can trust. Roll back when the agent acts wrong.
Sidekick versus background
Some agents sit next to a person and draft the next action. Others run in the background on a loop. The sort decides which. A human still owns the signature, the negotiation, and the exception.
The economics
AI roll-up
Buy a firm that runs on people and old software. Put engineers inside. Re-engineer the work. Margin up means the firm is worth more. The episode frames this as already happening in accounting, IT, and law practices.
One pattern, many companies
A sponsor opened the door to a NetSuite portfolio. One ERP, one problem (payables, payments, reconciliations), one set of politics. Map it once, deploy it across the set.
Proof at six months
The sale is the before-and-after. Audit in weeks, build in the next block, then return at three and six months with the old number and the new number. The $31 to $6 invoice cost is that proof in one line.
On-prem is not the current ask
Asked about GPU clusters for sensitive data, Vas said they have seen none of it, including at banks, healthcare, and pharma. Buyers are using cloud model hosts with no-train terms. OpenAI's private intelligence launch is treated as a response to the same worry, not as a reason to stand up hardware yet.
Implementation
Steps are synthesized from the episode method and the five-day starter plan. Not a verbatim checklist from the speakers.
Map your own systems first
List every app that holds your stuff. Mark which one wins when two disagree, and how work routes between them. Vas uses a personal inbox-and-files analogy so the client mess is recognizable before you touch theirs.
Pull a real week of work
Twenty things you did last week, from paying a bill to booking a hold. Count how many were the same step twice. The paper version of your job is shorter than the one you actually ran.
Write one process down, step by step
Pick one loop, such as paying a bill. From the email to the bank, every click, every tab, every wait. Name the owner. This is the human API. Do not open a model yet.
Sort every step into the four buckets
Delete, plain code, agent, human. Delete first. Rules go to code. Agents only where a bounded judgment call remains. A person still signs, negotiates, and handles the exception. Most people start at the model. That is the failure mode.
Baseline the number you will be judged on
Time, cost, error rate, volume. Write it down before you build. The payable case only works as a story because $31 and 17 steps existed as a before.
Build inside the tool they already open
Write back to the system of record. Put the human checkpoint in Slack. Use the cheapest model that passes a small eval. Keep a rollback for a wrong action.
Run the five-day version on a real business
Monday, list the stack. Midweek, map one process and sort it. End of week, ship the smallest thing that removes one step, on a business you know: a dentist, a law firm, the shop on the corner. Do it free if you have no reps. Charge the next one.
Come back and prove it
Return at three months and six months with the old KPI and the new one. Build the pattern once, then deploy it on the next company with the same ERP and the same problem. That is the roll-up motion in miniature.
Tools and Resources
These resources are curated in two groups. Mentioned Resources are pulled directly from the source material, and Suggested Resources are added to help you expand and apply the ideas beyond the original.
The following resources may contain affiliate links. As an Amazon Associate I earn from qualifying purchases at no extra cost to you. This does not influence the placement of links on this page.
Mentioned Resources
| Resource | Description |
|---|---|
| Episode on YouTube | Full 53:59 masterclass. Greg Isenberg with Vas of Varick Agents. |
| Source post | Isenberg's five-point summary of the method, posted with the episode. |
| FDE slides | Presentation linked from the episode. Process maps and the four buckets. |
| Vas article | Written companion linked from the episode description. |
| Varick Agents | Vas's firm. Enterprise AI implementation, forward deployed. |
| FDE in 30 days | Varick's longer on-ramp, linked from the episode. |
| NetSuite, Salesforce, Slack | Named systems of record. Agents write back here. Human sign-off is Slack. |
Suggested Resources
| Resource | Description |
|---|---|
| The Goal (Eliyahu Goldratt) | Constraint and process flow. Useful before you automate a step that should be deleted. |
| The Phoenix Project (Kim, Behr, Spafford) | Work as a system of handoffs, waits, and unplanned work. The shape of the maps in this episode. |
| Thinking in Systems (Donella Meadows) | Stocks, flows, and feedback. Helps separate a loop you should collapse from a loop a human must keep. |
| High Output Management (Andy Grove) | How to instrument a process and manage by the output. Pairs with the before-and-after KPI habit. |
AI Prompt