Forward deployed engineers are getting paid $1M a year right now.
The good ones walk into a company, rebuild one broken process with AI agents, and hand back $5M to $25M in value. Take 10% of that and the salary stops sounding crazy.
But here's what nobody shows you: the actual method. The thing they do on day zero inside a real company.
Vasuman Moza runs Varick Agents, where his team does exactly this for Fortune 1000 clients. He went on Greg Isenberg's podcast and walked through AI process reengineering step by step, with anonymized numbers from real engagements. A $5B public software company. An accounts payable overhaul that cut the cost per invoice from $31 to $6.
I took notes on the whole 54 minutes. Here's the playbook.
Why "just apply AI" keeps failing
Vas published an article called "Don't Apply AI" a few weeks before this episode. The argument: AI isn't a coat of paint. You can't brush it over a company and expect margin to appear.
Picture a typical mid-size acquirer. Eight acquisitions over the years. Five regions. 23 different systems of record. Chicago runs SAP, Salesforce, and Workday. Toronto runs NetSuite, HubSpot, and ADP. Every region does things its own way.
Slap a chatbot on top of that and you get the same mess, faster.
Vas quotes Michael Hammer, who studied process redesign 30 years ago: speed up every one of your 20 steps and the process barely gets faster, because the time lives between the steps, in the handoffs. That's where a 24-day cycle hides.
So before you write a single agent, you have to find the real process and rebuild it. That's the whole job.
Day zero: interviews, process mining, documents
When Varick lands inside a company, the first move isn't code. It's three parallel digs.
1. Interviews. Start with one department. In finance that means the head of AP, AR, reconciliations, billing, banking, FP&A. Most of the process lives in people's heads, not in any doc. Every company has the person who "has been here 20 years and just handles it". Your job is to turn that person into what Greg calls a human API.
2. Process mining. Put read access on the systems of record (Salesforce, NetSuite, Dynamics, Workday) and watch for 3 to 4 weeks. What data comes in, how often it gets corrected, which record types trigger which actions. The logs tell you the truth the interviews can't.
3. The document sweep. SharePoint, Drive, Notion, Slack, Teams, Gmail, spreadsheets. Sometimes outdated, sometimes gold.
→ A big enterprise has 10 years of Salesforce history to mine. A small business has almost nothing written down, so interviews carry the weight. You pick the mix per company.
The $5B company that didn't know its own sales process
Here's the example that made me sit up.
A public software company, $5B in revenue, 150+ products. The CRO brought Varick in. The company already had a process document (a consulting firm had mapped it for them): build a quote, submit it, deal desk, approve, send, negotiate, sign. Seven clean steps.
Varick's process mining agents found the real version: 20 steps with seven loops.
- 61% of quote requests looped back to step one
- Legal sent deals back 12% of the time
- 30% of the time a new quote had to repeat the whole approval chain
Nobody inside the company had ever seen their own department drawn that honestly. The CRO's reaction, per Vas: "I feel like you understand our department better than we do."
And that map alone is worth money. Before you've built anything, you've handed the executive team a picture of their own company they've never had.
Sort every step into four buckets
Once the 20 steps are on the wall, each one goes into exactly one bucket:
- Delete. The step shouldn't exist in a post-AI world. Kill it.
- Plain code. No judgment needed. If X then Y. A simple API call, not an agent.
- Agentic. Judgment plus enough historical data. Building a quote, categorizing an invoice line item, routing an exception.
- Human in the loop. The risky calls: approvals, negotiation, signing, sending payment.
That last bucket protects you. An agent that pays invoices end to end is a phishing scam waiting to happen. One convincing fake invoice and the money's gone. A human approving payments in Slack costs seconds and saves you the horror story.
Notice how few steps end up agentic. In the $5B engagement it was five out of 20. Most of the win came from deleting steps and writing boring deterministic code. The agents handle the judgment calls, nothing more.
Build inside the systems of record, not next to them
This is the pitch detail that wins enterprise deals.
Varick's agents mine Salesforce and act in Salesforce. Human approvals arrive as Slack messages. Nobody gets a new dashboard or a new login to learn.
Why it matters: Vas has clients who spent several years and $10M migrating from one ERP to another. If your AI pitch starts with "first, move off Salesforce", you've already lost the room. Nobody retrains 400 people for your tool.
Same logic scales to private equity. A PE firm with 26 portfolio companies can't run 26 separate AI transformations. So Varick groups the portfolio by system of record: the five companies on NetSuite become one engagement, the four on Dynamics become another. One playbook per stack, repeated. (This is the engine behind the whole AI rollup wave: buy the firm at one margin, sell it at a much better one.)
The numbers: $31 per invoice down to $6
The accounts payable engagement Vas walked through is the cleanest before-and-after I've seen on this topic:
- Process steps: 17 → 7
- Cycle time: 24 days → 6
- Exception loops: 6 → 1
- Straight-through rate (invoices needing zero human touch): 18% → 87%
- Cost per invoice: $31 → $6
That straight-through number is the one to stare at. Before: 82% of invoices fell into exception handling. After: 13%. A chunk of that fix wasn't even agents, it was deleting broken loops.
And it's not an isolated case. Thrive Holdings has 35 engineers deployed across 70 firms, turning out tax returns 30% faster at 98% accuracy. Vas has seen month-end close drop from 22 days to 4 to 8 days. Buy a services firm at a 60% gross margin for $1B, push the margin with this method, and the exit multiple does the rest.
One more thing worth stealing: sell the outcome to whoever signs. A CFO wants cost per invoice and close time. A CRO wants quote velocity. A chief people officer wants better hires trained faster. Same agents, different scoreboard.
The five-day plan if you're starting from zero
Greg pushed Vas for the "I have no clients, no track record" version. Here it is:
- Monday: list every app that holds your own stuff. Inboxes, file stores, messaging apps. Map which one wins when they disagree.
- Tuesday: write down 20 things you actually did last week. Paid a bill, cancelled a subscription, chased an invoice.
- Wednesday: pick one process and document it end to end, every exception included.
- Thursday: sort every step into the four buckets. Delete, plain code, agent, human.
- Friday: cold outreach to small businesses and offer to do this for ONE workflow. Free if you have to. The next one is five figures.
A 60-person accounting firm Vas described ($12M revenue, 400 clients) claimed a six-step workflow. The real one had 14 steps, clients re-asked for documents 70% of the time, and partners bounced work back 35% of the time. Some clients sent invoices as photos of notebook scribbles. That's the level of mess waiting for you, and it's winnable with exactly the method above.
You don't need a CS degree for this. You need three skills stacked: understand how the work gets done, ship production code with audit trails, and know which model to trust with what. Vas says people who have all three are NBA-player rare. That's the gap. I broke down the role itself, the Palantir history and the career math, in my earlier piece on the forward deployed engineer job.
On models: most of this work doesn't need a frontier model. Varick runs plenty of workflows on smaller and open source models, and benchmarks every workflow against several before picking. The hard-won asset here is the process map. The model is swappable.
FAQ
What is AI process reengineering?
It's rebuilding a business process from the ground up before adding AI: interview the people who run it, mine the systems of record to find the real steps and loops, then sort every step into delete, plain code, agent, or human in the loop. The term is a deliberate contrast to "applying AI", which layers agents over a broken process and automates the mess.
How do forward deployed engineers make $1M a year?
Value-based math. If one FDE engagement cuts a company's cost per invoice from $31 to $6 across millions of invoices, that's multi-million dollar savings every year. Companies happily pay a slice of that. Some PE firms now hire FDEs with a share of the carry, so the engineer gets a cut of the exit uplift they created.
Do I need to be a software engineer to do this?
You need to ship production code, but not a decade of credentials. Vas's bar is three skills in one person: knowing how work actually flows through a company, writing reliable integrations with governance and rollbacks, and picking plus testing the right model per task. AI coding tools lower the second bar every month. The first one, you earn by doing engagements.
Why not just replace the company's old software with an AI-native tool?
Because migrations are where enterprise deals go to die. Companies have spent $10M and multiple years moving between ERPs. Agents that live inside the existing Salesforce or NetSuite, with approvals in Slack, need zero retraining and zero migration budget. Path of least resistance wins the contract.
Where do I find my first client?
Small businesses with messy back offices. Accounting firms, IT shops, agencies, anyone drowning in invoices and handoffs. Offer to map and fix one workflow. Even free, the first engagement buys you the case study that prices the second one at five figures.
Steal the method, then hear the founders running it
Every week on the Profitable Founder Podcast I talk to bootstrapped founders building businesses exactly like this, with the real numbers attached. If the FDE path has your attention, the interviews will show you what the first $100K actually looks like.
