Teamday
← Back

Vasuman Moza: How Varick's Forward Deployed Engineers Work

Greg Isenberg
agentsenterpriseautomationbusinessfuture-of-work

How Vasuman Moza and Varick Agents Deploy AI Agents as Forward Deployed Engineers

Vasuman Moza is the founder and CEO of Varick Agents, a company that sends forward deployed engineers (FDEs) into enterprises and private equity portfolio companies to rebuild whole departments with AI agents. On Greg Isenberg's channel he walks through anonymized client engagements: a $5 billion public software company, a PE firm with 26 portfolio companies and an accounts payable rebuild, plus a peer's engagement with a 60-person accounting firm whose six-step process turned out to have 14 steps. Isenberg frames the episode around why a forward deployed engineer can earn $1 million a year: if you drive $10 million of value for a company, paying you 10% of it is an easy trade.

Don't apply AI: Moza's article "Don't apply AI" sets the thesis. "AI isn't something that can be applied like a coat of paint. It's something that really involves process re-engineering." He points to AI roll-ups such as Thrive Holdings, where engineers placed inside acquired firms report tax returns done 30% faster at 98% accuracy, and notes that a company bought for $1 billion whose margin doubles can sell for several times that. Spreading AI over a company with 23 systems of record across five regions, by contrast, only makes a broken process run faster.

Step one is process mapping: On day zero, an FDE interviews people department by department (in finance: AP, AR, reconciliations, billing, FP&A), because the real process lives in the heads of people who have been at the firm for 20 years and "just handle it". Then process-mining agents sit on the systems of record (Salesforce, NetSuite, Dynamics, Workday) for three to four weeks to see what actually happens, and finally the FDE reads whatever documentation exists in SharePoint, Drive, Notion, Slack and spreadsheets. At the $5 billion software company, brought in by the CRO, the quote-to-signature process documented in an earlier Deloitte engagement had seven steps. Process mining over the CRM found 20 steps with seven loops: 61% of requests looped back to the start, legal sent deals back 12% of the time, and 30% needed a new quote. The executives' reaction, as Moza tells it: "I feel like you understand our department better than we do."

Four buckets for every step: Each mapped step gets sorted into one of four buckets: delete it, replace it with plain code (if X then Y, no judgment), give it to an agent (judgment backed by enough historical data, like classifying an invoice line item), or keep a human in the loop for risky actions such as approval, negotiation, signing and sending payment. In the 20-step sales process, five steps became agentic. Payment stays with a human because an agent that decides a phishing invoice "looks legit" will pay it.

Build inside the systems of record: Moza's strongest advice to anyone becoming an FDE is to "pitch yourself as building these agents inside their systems of record." Varick's agents mine Salesforce and act in Salesforce, and human approvals arrive as Slack messages. Clients have spent years and, in one case, about $10 million migrating ERPs, so a pitch to move them onto an AI-native CRM loses them. For the PE firm, Varick grouped portfolio companies by system, five on NetSuite and four on Dynamics, so FDEs learn what each ERP already does, fill the gaps, and sell to the same kind of buyer, the CFO, again and again.

The accounts payable numbers: One anonymized AP engagement, mapped in two to three weeks, went from 17 steps to seven, cycle time from 24 days to six, exception loops from six to one, and cost per invoice from $31 to $6. The biggest change was the share of invoices that went through without an exception: "We drove that from 18% to 87% for this client." Moza calls that a process re-engineering result more than an agent result. Citing Michael Hammer, he argues the time hides between steps, in handoffs, not in the steps themselves. He also says the companies he works with mostly move finance staff to higher-leverage work rather than cutting 50% of jobs.

Models and infrastructure: Most workflows don't need a frontier model. Varick uses mid-tier models from the big labs plus open models, and Moza says you should "benchmark every single workflow against every single model." Enterprises often refuse Chinese models, and Varick routes model calls through Azure Foundry, AWS Bedrock and Google Vertex under zero-data-retention terms. Asked about on-premise GPUs for regulated clients such as banks, healthcare and pharma, he says: "Truthfully, we've seen zero of that." He separates sidekick agents you chat with from background agents that do the work and ping you when needed, and says the second kind delivers far bigger ROI, roughly 70-80% faster output versus 10-20%.

What a Forward Deployed Engineer Needs, According to Vasuman Moza

  • Three people in one - Know how the work gets done (systems of record, how a function like accounting should run), ship production code with auditability and security, and own the AI layer: model choice, evals, harness tuning and rollbacks when agents take wrong actions. Then explain it all to senior leadership.
  • Sell the outcome to the buyer - Varick sells cost savings, revenue uplift and risk mitigation. CFOs want lower cost, but also a month-end close cut from about 22 days to four or eight. A chief people officer wants better hires trained faster, not savings.
  • Baseline before you build - Varick runs a four-week audit with proofs of concept, builds agents in the next four weeks, and reports against the original KPIs at three and six months.
  • Get paid on the value - Isenberg's math is 10% of $10 million in value. Moza adds that some PE firms give FDEs a share of the carry, such as 0.5% of the gain between purchase and sale price.
  • Start with one SMB workflow - His weekly plan: list every app that holds your stuff, list 20 tasks from last week, write one process down step by step, sort its steps into the four buckets, then offer to rebuild one workflow for a small business, for free if it's your first.

What Varick's Playbook Means for Companies Deploying AI Agents

Most AI pilots fail, Moza argues, because companies slap AI onto processes nobody has mapped. Varick's playbook is the reverse: find the real process, measure it, delete or hard-code what doesn't need judgment, give agents the steps that do, keep humans on the risky ones, and build it all inside the tools people already use. The same playbook works for companies building FDE teams of their own, and it explains the pay. The scarce skill is mapping and rebuilding the process, not writing the agent.