Making everyone AI-native

01Map the work before you apply AI.

Ask how a process works and you get the documented version. Watch it, and you find the real one: the extra checks, the side spreadsheet, the exception only one person knows how to handle.

Daniel Kornum of Varick Agents, a forward-deployed engineering firm, writes that the gap between a team’s written procedure and how its work actually runs is usually large, and largest where exceptions pile up.

Automate the documented process and you automate the wrong thing. Every serious AI decision starts with a map of the real work.

02Redesign the work. Don’t just automate it.

In 1990, Michael Hammer argued in the Harvard Business Review that using technology to speed up old processes mostly gets you faster old processes. His title said it in three words: “Don’t automate, obliterate.” AI makes the same mistake easier to make, and at a larger scale.

So the first question isn’t which agent to buy. It’s which steps should exist at all. Every step of a workflow belongs in one of four buckets:

• DeleteWork that shouldn’t exist once the process is redesigned: waiting, re-keying, chasing.
• Plain codeFixed rules that run the same way every time. Cheaper and more predictable than a model.
• AgentJudgment on messy inputs, where you can check the result against past cases.
• HumanDecisions where a mistake would be costly, public or hard to undo.

Most steps don’t need an agent. In Zapier’s data on 20 million live workflows, an AI step shows up in about 4% of automations. Most of the work is moving information reliably from where it lands to where it needs to go.

03Forward deployment works. It doesn’t scale by hand.

The people who do this well are forward-deployed engineers, or FDEs: engineers who sit inside a team, map how its work really runs, decide where AI belongs, then build it and measure it. Palantir created the role. Today OpenAI and Anthropic hire FDEs too, and Andreessen Horowitz has called it the hottest job in startups.

But good FDEs are rare and expensive, and they can sit with only a few teams at a time. When the engagement ends, much of what they learned leaves with them. Most companies will never get one into every team.

04The best FDEs you could hire already work for you.

The accounts-payable lead who knows which vendors always send the wrong purchase order. The operations manager who reroutes the exceptions nobody wrote down. The account executive who knows which approvals actually matter. They already have the hardest part of an FDE’s job: they know how the work really works. What they lack is the time and the tools to map it, measure it and build on it.

Too often, outsiders arrive, ask your busiest people to walk them through their work, and leave with a plan to automate it. We think that’s backwards.

AI transformation done to people stalls. Done with them, it compounds.

05Prove it on your own work.

Benchmarks can’t tell you what will happen in your close, your support queue or your deal desk. Two accounting studies from Mercor show why.

Both results are true. The difference is the work. So test an agent the way you’d test a new hire: in shadow mode, alongside the real work, with nothing going out, scored against what your people actually did.

Keep the cases and the judgments behind them. They become your evals, and your evals become one of the most valuable AI assets you own: a written record of what “correct” means at your company.

06Price AI by the task.

The most capable model isn’t always the right one. For routine steps, a smaller model often does the job for a fraction of the cost. For hard, high-stakes steps, the frontier model is cheaper than its mistakes. Decide task by task, with your own evals, and check again as prices fall.

Make the build-or-buy call the same way: before you spend, on evidence. Our guide to AI model cost per task shows the math.

07Share what works.

In every company, a few people have already figured out how to make AI useful in their own work. Their prompts, skills and workflows sit in personal folders and private chats, where nobody else can use them.

AI-native companies treat them as shared assets. They spread the skills that work, measure which ones help, and retire the ones that don’t. At Workflows.io, the whole team works from one shared repository of skills and agents, as Fivos Aresti describes.

08Keep people in charge.

As models take on more of the legwork, Andrej Karpathy expects more of our work to “rise up the abstractions into oversight and understanding.” That’s the job we want to give people: designing the work, checking it and improving it.

Not every decision should be handed over. A person approves anything risky, every agent leaves a record of what it did, and the people closest to the work decide when an agent has earned more responsibility.