AI agents pay off when you fix the process first

How forward deployed engineers map, sort and automate real business work

Almost every company wants AI agents right now, but far fewer can name the step they should take over or what that step costs today. A long conversation between Greg Isenberg and Vas from Varick tackles exactly that gap, and its starting point is blunt: AI pays off when you re-engineer the process first, then build agents into it.

The episode centers on the forward deployed engineer, or FDE. An FDE works inside a customer’s business rather than only in a product team’s office, and the job starts with understanding how the work really gets done. If you want the full discussion, the episode is on YouTube and runs about 54 minutes.

Map the real process before you touch the software

So where do you start? With the least glamorous step, which Vas spends a lot of time on from around the 09:55 mark. The method combines interviews with the people who do the work, process mining of the company’s systems of record, and whatever documents already exist. Those sources rarely agree, and the disagreements are often the most useful part.

Imagine a step that everyone describes as a one-day task. The system logs show the same item sitting in a queue for a week. The interviews explain why, and an old document reveals a rule nobody remembers writing down. Only when all three sources line up can you see the process as it actually runs.

Sort every step into one of four buckets

Once the map exists, each step goes into one of four buckets: delete it, handle it with plain code, give it to an agent, or keep it as a human decision. The sorting is where most of the value shows up, which is why the episode returns to it around the 18:11 mark. Plenty of steps should simply disappear. Reaching for AI agents first is the expensive habit to avoid when a simple rule would do.

Human decisions stay human. The aim is not to remove people from judgment calls. It is to stop them spending their hours on work that needs no judgment at all.

Build inside the systems people already use

The second rule is about where the agent lives. Vas argues for building AI agents inside the tools a company already runs on, such as Salesforce, NetSuite and Slack, a point made around 19:04. A new platform asks people to change their habits. An agent that sits inside the CRM or the finance system asks them to change very little, and that makes adoption a much smaller problem.

What an accounts payable overhaul shows

The most concrete example in the conversation is an accounts payable overhaul, the work of receiving, checking and paying a company’s bills. The cost per invoice went from $31 to $6. The discussion of how the accounts payable process gets mapped starts around 28:39.

The number is striking, but the method is what you can copy. Map the steps, sort them into the four buckets, and rebuild the work inside the systems your finance team already uses. Then measure cost per invoice before and after. A result nobody can compare is hard to defend in the next budget meeting.

Three engagements, one method

The episode walks through several other engagements. One involves a $5B public software company, around the 14:11 mark. Another is a private equity portfolio, covered around 21:36. A third is a 60-person accounting firm, discussed around 32:54. Read together, they suggest the same method holds at very different scales.

Why AI roll-ups need a buyer at the top

The episode also looks at the business opportunity behind AI roll-ups. A roll-up buys or partners with a group of small companies in one market and applies the same improvements across all of them. The episode is also direct about who you need to sell to: C-suite executives, a topic it covers around 23:43.

The advice is simple. Sell the outcome each buyer cares about, and prove it with before-and-after KPIs. A CFO and a COO may care about the same process for different reasons, so the pitch should change with the person in the room.

Code, agents or people? Choosing the right tool

The episode covers when to use code, when to use an agent and when to bring in a person, and how to choose AI models for each job (around 34:51 and 36:00). My takeaway is that the model should follow the job, not the other way round. Model choice also depends on how easily you can switch later, which our piece on open models and startup freedom to switch explores.

Sidekicks and background agents

The episode separates sidekick agents, which work alongside a person, from background agents that run without anyone watching in real time (around 38:16). The difference matters because it changes who checks the output and how you measure success. A sidekick needs a good interface, and a background agent needs solid monitoring. Getting a background agent to finish a task reliably is the hard part, and we have written about the hidden cost of finished agent work.

The conversation also touches on on-premise hardware demand around 47:39. Whether you run models on your own machines or rent them, the question is the same: what does switching actually cost you? We explored that in what switching compute really costs.

Why the FDE role pays so well

The episode also asks why FDEs earn so much, around 42:25, and the skill mix explains a good part of it. A top FDE combines domain knowledge, production engineering, AI judgment and strong communication. Finding all four in one person is rare, and that scarcity is a sensible place to start.

What to take away

Start with the process, not the model. Map it with interviews, system records and existing documents, sort every step into delete, code, agent or human, and build inside the tools your team already uses. Measure cost before and after, and sell the outcome each buyer cares about.

Where could your own workflow improve? Ask which step you would delete first, which steps need a person’s judgment, and which number would prove the change. The episode ends with a five-day starter plan around 45:26, and it is the most practical part of the conversation, so listen to the specifics rather than relying on a summary like this one.

The biggest lesson is that AI does not fix a broken process. It rewards the teams that understood their process well enough to change it. Fix the work first, then let the agent do what is left.

AI pays off when you re-engineer the process first, then build agents into it.

The sorting is where most of the value shows up

A result nobody can compare is hard to defend in the next budget meeting.

▶

Sources

• Masterclass: How FDEs make $1M/yr deploying AI agents (Greg Isenberg) — Source episode; conversation between Greg Isenberg and Vas from Varick.