Map the Work Before You Add AI

The quickest way to get real value from AI in finance is to map how the work actually happens before you add an agent to it.

Why it matters. Forward-deployed engineers (FDEs), a role Palantir pioneered, sit inside a client's business and build AI agents into its real workflows. They're in demand, and the best are paid very well, because their work shows up in the numbers. In a new masterclass on Greg Isenberg's channel, Vas Moza of Varick Agents explains how they do it. His message is simple: don't "apply" AI like a coat of paint. Redesign the process.

What FDEs do that finance teams can borrow

1. Find the real process, not the documented one. FDEs use three sources: interviews with the people doing the work, data from systems like NetSuite or Salesforce, and existing documents. In one engagement, a sales quote approval documented as 7 steps turned out to be 20, with 7 loops. In an SME, most of the process lives in people's heads, so those conversations matter most.
2. Sort every step into four buckets. Delete it, hand it to simple rules, give it to an agent, or keep it as a human decision. Some steps shouldn't be automated at all. They should go.
3. Keep people on the risky calls. Agents suit work that needs judgement and has enough history behind it, like coding an invoice line. Approvals, negotiations and payments stay with a person. His warning: let an agent pay end to end and a convincing fake invoice can go straight through.
4. Build inside the tools you already use. Put agents into the ERP, CRM and chat tools your team already knows, and route approvals as a simple message. Nobody has to migrate or retrain.
5. Pick the workflow before the model. In his experience, most jobs don't need the newest, biggest model. Test each workflow against a few models and use the one that does that job well.
6. Baseline before you build. In an anonymised accounts payable example, the rebuild took the process from 17 steps to 7, cycle time from 24 days to 6, straight-through processing from 18% to 87%, and cost per invoice from $31 to $6. He stresses that this was a process redesign, not just agents. It's one engagement, not a benchmark.
7. Prove it months later. The playbook ends by measuring again at three and six months against the numbers written down at the start. If the saving isn't there, it didn't happen.

What this means for your team

You don't need an FDE to start. Pick one workflow you own, like supplier invoices or month-end. Write down every step, every wait and every exception, then time it. As Vas notes, citing Michael Hammer's classic reengineering work, the days usually disappear in the waiting between steps.

That's the FinanceSignal idea in practice: a cost per invoice or a close timeline is just noise until it tells you what to change. One caution. Gartner and Forrester, quoted in The Register, warn that bespoke builds can leave you dependent on a vendor, so keep the process map and the know-how in-house.

Soft close

The $1M headline grabs attention, but the lesson underneath is practical: start with the business problem, tidy the process and the data, check every figure, and measure the real saving. If you're starting out, the AI Starter Kit for SMEs helps you pick a safe first job and counts time saved after checking. For more tools that turn numbers into decisions, visit Signal Lab.

Summary of a third-party video; figures are the speaker's own and anonymised. General information, not financial, tax or technology advice.

Sources: Greg Isenberg (The Startup Ideas Podcast), Masterclass: How FDEs make $1M/yr deploying AI agents, with Vas Moza, CEO, Varick Agents (YouTube, published 1 October 2026), watched via transcript 4 Oct 2026 — https://www.youtube.com/watch?v=1a5HxU52vCQ · Varick Agents, Don't Apply AI (slide deck linked from the episode), checked 4 Oct 2026 — https://learn.varickagents.com/dont-apply-ai · Simon Sharwood, Palantir's fondness for French food cooked up tech's latest fad – forward-deployed engineers (The Register, published 3 October 2026), checked 4 Oct 2026 — https://www.theregister.com/channel/2026/10/03/palantirs-fondness-for-french-food-cooked-up-techs-latest-fad-forward-deployed-engineers/5300360 · Gergely Orosz, What are Forward Deployed Engineers, and why are they so in demand? (The Pragmatic Engineer, 12 August 2025), checked 4 Oct 2026 — https://newsletter.pragmaticengineer.com/p/forward-deployed-engineers · Michael Hammer, Reengineering Work: Don't Automate, Obliterate (Harvard Business Review, July–August 1990), checked 4 Oct 2026 — https://hbr.org/1990/07/reengineering-work-dont-automate-obliterate