Automate First, Cut Second

The IRS has shed more than 28,000 employees since early 2025 — roughly a quarter of its entire workforce. The stated reason? Technology and AI will make up the difference. The agency is modernizing. It’s automating. It doesn’t need as many people as it used to.

There’s one problem with that story. In March 2026, the GAO — Congress’s own watchdog — released a report on the IRS’s AI program. What they found was striking: the IRS had fired 63 of the people who were actually building the AI. The research and analytics unit, one of the agency’s two main AI hubs, lost most of its staff. And as of mid-2025, 77 of the IRS’s 126 active AI initiatives were still in development.

The GAO’s conclusion was direct: the IRS’s AI program is now at serious risk of failure.

Let that sink in for a second. The agency cut the people it needed to build the automation it was using to justify cutting the people. That’s not a transformation strategy. That’s a sequencing problem — and it’s one that’s playing out well beyond the IRS.

Cut Now, Automate “Soon” — A Pattern Worth Recognizing

When a company announces AI investment and headcount reductions in the same breath, the implied logic is that the automation is replacing the work. But if you look at the actual timeline, the cuts almost always come before the automation is proven — sometimes before it’s even deployed.

This isn’t a new trap. Anyone who worked through the RPA era of the 2010s saw a version of it play out. Companies announced massive automation initiatives, reduced headcount, and then discovered the systems were brittle and narrow — and critically dependent on the institutional knowledge that had just walked out the door. Invoice formats changed. Systems updated. New entity structures appeared. RPA didn’t handle exceptions well, and the people who did were gone.

AI is more capable than RPA — I genuinely believe that. But the sequencing problem is identical. You cannot remove the people who understand a process before the automation of that process has been validated in production.

The Questions You Should Be Asking

Whether you’re advising a client going through an AI-driven transformation or evaluating what’s happening inside your own firm, there are four questions worth asking before anyone starts counting on automation to cover headcount reductions:

• Is the automation in production, or still in pilot? POCs are impressive. Production is a different beast. Getting from pilot to live is hard; staying live is harder.
• Who maintains it once the people who built it are gone? AI systems require ongoing maintenance. Business processes change, data formats shift, edge cases multiply. Someone has to own that work.
• What’s the error rate on real data — not demo data? Systems that perform well in controlled pilots routinely underperform on production data. You need to see the real error rate before you restructure around the assumption that the work is being done.
• Who catches the exceptions if the people who knew them are gone? Every process has edge cases. If the people with the institutional knowledge to handle those exceptions have left, there’s no safety net when the automation misses.

These aren’t anti-AI questions. They’re basic project management. Automation should precede the headcount reduction, not justify it retroactively.

The Right Sequence

The defensible path is slower and less photogenic than a single press release. It looks like this: deploy automation in parallel with your existing team, validate performance on real workflows, close the gaps, and then right-size around what’s actually proven to work. Nobody disagrees that over time, AI will change how many people certain tasks require. The question is whether you’ve actually validated that before you act on it.

The IRS tried to skip that sequence. The GAO documented what happened. It’s the clearest possible case study for what “automate first, cut second” actually means in practice — and why the order matters.

Key Takeaways

• Automate first, cut second. The only defensible AI-driven workforce reduction is one where the automation is proven before the headcount is removed.
• “AI is replacing this work” is easy to say and hard to verify. Ask the sequencing questions. The IRS’s situation is visible because the GAO audits them — your clients’ AI claims won’t come with that scrutiny.
• The people building the AI are often the people being cut. Institutional process knowledge is required to build good automation of that process. Lose the people, lose the knowledge needed to make the automation work.
• RPA taught this lesson once. AI is teaching it again. The direction of travel is real — but sequencing is the variable that determines whether you get efficiency or a gap.
• Maintenance is the conversation nobody is having. Getting to production is hard. Staying in production — as processes evolve, data changes, and edge cases accumulate — is a whole other challenge. Don’t underestimate it.
• This is an advisory opportunity. Clients going through AI transformations need someone asking hard sequencing questions before the cuts are made, not after. That’s a conversation you’re positioned to lead.

Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: [lesson link]