Source: raw/Masterclass_-_How_FDEs_make_1M_yr_deploying_AI_agents.md

Creator: Greg Isenberg (Startup Ideas), guest “Voss” of Veric Agents (spelling from auto-captions) · URL: https://www.youtube.com/watch?v=1a5HxU52vCQ · Platform: YouTube (auto-captions, fetched 2026-10-21) · Duration: not stated in the raw file

Disclosure. The guest runs an agent-deployment firm and recruits on air: “if you are this person, we really need to talk. We want to hire you,” and later “I need to hire people. So, if people can apply.” Isenberg says he has no stake in the company. A Google voice-AI segment early in the episode is sponsored. Every client number below is the guest’s own, anonymised and unaudited.

A forward-deployed engineer (FDE) embeds inside a business and rebuilds its processes with agents. Voss’s thesis is in the title of his article, “Don’t apply AI”: AI “can’t be applied like a coat of paint,” it requires “process re-engineering.” His method is five steps (map, re-engineer, build, deploy, roll out), and the core of it is sorting every process step into one of four buckets. The method works the same at a $5B public company and a 60-person accounting firm, which makes it usable by a solo consultant.

Key Takeaways

  • Most AI pilots fail because they “slap AI on.” Voss’s examples: “rolling out a license with [Claude Code] to everyone,” or “building an agent that doesn’t really understand the workflow.”
  • The documented process is wrong. At a $5B-revenue public software company, the quote-to-cash process documented by Deloitte had 7 steps. Process-mining agents on the CRM found “a 20 step process with seven different loops”: 61% of requests loop back to step one, legal sends 12% back, and 30% need re-approval.
  • Sort every step into four buckets: delete it; plain code (no judgment, “if X then Y”, or a switch case); agentic (judgment backed by enough historical data); human decision (approval, negotiation, signing, payment).
  • Payment stays human because of phishing invoices: an agent that judges an invoice legitimate and pays it end to end is the risk.
  • Build inside systems of record, not on a new surface. Agents mine and act in Salesforce; human-in-the-loop “is a message in Slack.” Clients who spent “$10 million one time” migrating ERPs will not move again, and nobody needs retraining.
  • You rarely need a frontier model. “Very few cases where we’ve seen the need to deploy Fable or Astra.” They use Opus 4.8 or Sonnet, GPT at lower thinking, and open models. “Benchmark every single workflow against every single model.”
  • Personal agents are sidekicks, not deployments. Dots, Muse, Grok Bot and Instinct are chat-driven “sidekick” agents. Background agents “don’t bother you… they ping you when they need to,” and personal-agent governance for enterprise “isn’t there yet.”
  • Reallocation, not layoffs. Voss says his Fortune 500 and 1000 clients move finance staff to higher-leverage work rather than cut 50% of them. This is the vendor’s framing.

The method

1. Discovery: three inputs

  1. Interviews down one department. In finance: heads of AP, AR, reconciliations, billing, banking and FP&A, then down the chain. Knowledge “lives in their heads.” Ask why things are done this way, who really decides, “which step is theater,” and how often exceptions happen. Isenberg calls this building “a human API.”
  2. Mine the systems of record (Salesforce, NetSuite, Dynamics, Workday) for three to four weeks. What enters, how often it is corrected, and what follows from which record type reconstruct the real SOP.
  3. Existing documentation in SharePoint, Drive, Notion, Slack, Teams, Gmail and spreadsheets.

The split varies. A large company has a decade of CRM history. An SMB mostly has “people’s heads” and no system of record.

2. Show the map, then re-engineer

  • Show leadership the real map before proposing anything. “It breaks their heart, but they need it.” The CRO and CFO at the software company told the team, “you understand our department better than we do.”
  • Bucket each step (above). Voss’s rule for the agentic bucket: there is enough history to be concrete. His example is an invoice line for monitors, which is “very likely going to be office supplies,” with known vendor exceptions.
  • Michael Hammer’s point, cited by Voss: speeding up each of 20 steps may not shorten the process, because “it’s the cycle time between steps that makes all the difference.”

3. Build, deploy, prove

  • Put the estimated monthly agent cost next to the redesigned map (Isenberg’s tip). It is “shockingly low” relative to people.
  • Baseline the KPIs before building, then hold yourself to them.
  • Engagement timeline: a four-week audit, a four-week build, then proof at three and six months.

4. Private-equity roll-outs

  • A PE firm with 26 portfolio companies, 10 departments each and 10 workflows per department cannot be done company by company.
  • Group portfolio companies by system of record (for example five on NetSuite, four on Dynamics) and learn what each system already does, so each FDE solves the similarities once.
  • PE firms then ask for “the playbook” to apply to the next acquisition.

The numbers (guest-reported)

CaseBeforeAfter
Accounts payable (anonymised client)17 steps7 steps
24-day cycle time6 days
6 exception loops1
18% straight-through rate87%
$31 per invoice$6 (“an 80% reduction”)
Month-end close (typical CFO)“about 22 days… 4 weeks, 6 weeks”target “4 days or 8 days”
60-person accounting firm (not Veric’s client), $12M revenue, 400 clientssaid 6 stepsactually 14, with loops at 70% and 35%

Roll-up context. Voss says Thrive Holdings has “35 engineers across 70 firms” and reports “tax returns 30% faster at 98% accuracy.” His valuation logic: buy a firm for 2B, 8B. See the AI roll-up blueprint, which records a different Thrive count.

Selling it

  • Sponsors. CFOs are often the sponsor (“self-proclaimedly notoriously skeptical”), along with CIOs, CEOs and CROs.
  • Value in three buckets: cost savings, revenue uplift, risk mitigation.
  • Sell the outcome each buyer wants. A CFO wants a faster close at the same cost. A CHRO wants to “hire better people faster and train them quicker”; “they don’t want to save money here.”
  • Pay. Isenberg frames FDE pay as a share of the value delivered (“a million dollars”). Voss says some PE firms give FDEs about 0.5% of carry on the exit delta.

The FDE skill stack

  1. How the work actually gets done: the domain process and what each system of record does and does not offer.
  2. Shipping production code that calls those systems with auditability, governance and security. A CS degree is not required.
  3. The AI layer: model choice, what a model can be trusted with, evals, harness tuning, and rollback when an agent takes a wrong action.
  4. Executive communication.

“Usually you just have one or two of these.”

Try It

The guest’s five-day starter plan, for yourself first:

  1. Monday. List every app that holds your stuff (inboxes, file stores, messaging) and which one wins when they disagree.
  2. Tuesday. List 20 things you did last week (paid a bill, cancelled a subscription).
  3. Wednesday. Write one process step by step, with its variants and exceptions.
  4. Thursday. Put every step in a bucket: delete, deterministic code, agentic, or still you.
  5. Friday. Offer one workflow to a few SMBs. Do it free the first time if needed, using skill files, a personal agent and your own agents, with a timeline.

For an SMB, Voss suggests starting in skill files inside whatever tool they already use (Copilot, Claude Code or Codex) before building custom agents.

Open Questions

  • No client is named, and none of the before/after figures are independently verified.
  • Thrive Holdings’ firm count differs between sources: “35 engineers across 70 firms” here (as of Voss’s own research) against “close to 50” accounting firms in the roll-up article. It may be a later count or include non-accounting firms.
  • Guest name. “Voss” and “Veric Agents” come from auto-captions.
  • On-prem demand. Voss has seen “zero” requests for on-premise GPUs, even at banks and pharma. One firm’s experience.