Your IT Problem Was Never an IT Problem. Is AI Different?

Ryan Hunter
2 days ago
5 min read

Three years ago I wrote a piece for Municipal Magazine called Your IT Problem Is Not an IT Problem. The argument was simple: there's no such thing as an IT problem.

IT is a potential solution. When a government buys software to fix something it hasn't bothered to understand first, it usually gets the same problem back, faster and more expensively.

I used two examples. One was a clinic that spent months and real money moving to digital x-rays, then discovered the new system couldn't handle re-reads, the workflow where a clinician sends an image back to a radiologist for clarification. Instead of asking why those conversations kept breaking down, they paid the vendor to hard-code an on-screen imitation of sticky notes.

The other was a county thinking about buying agenda-management software. They did the process work first, and found two wins that required no technology at all: board minutes published in one day instead of six to eight weeks, and 85 percent of the meeting paper eliminated, because supervisors already had digital access and nobody had gotten around to discontinuing the old process.

Back then, the honest answer was frequently that you didn't need the software.

Three years on, I've had to reexamine whether that still holds.

Is AI different?

Yes. And no.

What's actually new

Every IT tool I had worked with before was, in the end, incapable of judgment. It moved data from one place to another and applied rules somebody had written down. AI does something else. It can think – or at least, it can do something that looks a lot like thinking. It can take a fuzzy, badly specified task and hand you back something usable.

The clearest example is reading long material. There has never been a cheap mechanical way to read a sixty-page staff report or two hundred public comments before a hearing and say what matters for this decision. Ordinary automation handles almost none of it, because the entire job is a contestable judgment about relevance.

Now you can hand that to a machine and get something useful back. That isn't a faster version of an old thing. It's a new thing.

I'm not writing this as a skeptic. We have been building AI into PPI's own work — drafting RFP responses, managing our knowledge base, absorbing work we used to do by hand — and it has made us faster and more responsive to our clients than we have ever been.

I want the public agencies we work with to have that too.

And you can't guess where it will work

Here's what I didn't expect. Go back to my two examples from 2023, and both of my intuitions have inverted.

Meeting minutes looked like the most boring administrative task in local government, and therefore the obvious candidate. Except I keep hearing from city clerks who are being sold AI transcription constantly, and they hate it.

One important function of meeting minutes is to count and record votes, and AI often just… can’t. A vote taken by a nod or a show of hands isn't on the audio at all. AI cannot transcribe what nobody said. If you want reliable minutes, somebody has to say the vote out loud or record it electronically.

Which means the technology requires a change to how the meeting is run (i.e., a process change).

Meanwhile the transcript itself did get nearly free. Long View, North Carolina, a town of two thousand people, took minutes production from six hours a meeting down to under one. The mechanical half got cheap. But the actual job, deciding what mattered, and accurately recording what was decided, is still hard.

Then there’s radiology. In 2023 I used it as an example of skilled human work that a software rollout had failed to support. AI now reads medical imaging, in many cases more accurately than a human. The time is coming when AI will do meaningful parts of what radiologists do.

So the task I assumed was trivially automatable has resisted automation, and the one I treated as unassailably expert is being encroached on.

AI inherits whatever process you put it on

The reason IT never solved your process problem is that IT had no judgment. It couldn't paper over a contradiction in your policy. It just failed, visibly, and stopped.

AI can paper over contradictions.

And the judgment it brings is frequently the same judgment your staff has been using for years to make incoherent guidance work at the counter.

Give AI poor information, and it will work – it will just be confidently wrong, quickly, and at scale.

This is true of our own work at PPI. I’ve spent the last several months building an AI tool that helps us respond to public RFPs. We call it Athena. Athena is impressive. But getting it to be good has taken a lot of human judgement.

For a long time I was frustrated that Athena did a poor job identifying which of our 1099 associates would be strong fits on a given proposal. Then I realized I had never actually compiled a list of the 1099 associates in our network — what they've done, what they're good at, and who's actually available.

That is not an AI problem. That's our own failure to document standard work.

There’s a silver lining: AI turns out to be an unusually fast and unusually cheap diagnostic for the places your own processes are half-baked. I’d been missing out on opportunities to propose my strongest associates on contracts for years, because I was largely relying on my own memory about who had which skills. Claude forced me to confront and solve that process problem.

What about public agencies?

Imagine the same problem in permitting: You want AI to answer residents' questions. Everyone agrees it's an obvious use case, right up until you go looking for the guidance and find five documents revised at five different times, two contradicting each other, one five years old, and a great deal of the real answer living in the subjective judgment of whoever is at the counter that day.

Drop AI on top of that and you haven't fixed the service. You've built a faster route to your own confusing, contradictory answers, delivered with total confidence and at much greater volume.

Two recent government AI chatbot experiences illustrate the difference:

Pennsylvania rebuilt its unemployment chatbots to recognize the many ways people phrase a question and match them against curated answers rather than generating new answers; it now fully resolves about 65 percent of conversations.
Meanwhile, New York City's business-services chatbot composed its own answers. It made headlines for advising business owners to confiscate their workers' tips. It was decommissioned in February of this year.

Same technology, same problem, same year – very different results.

New York City outsourced its critical thinking, while Pennsylvania used AI to provide better, faster access to validated human expertise.

So, is AI different?

Yes.

It can do work that previous generations of technology simply couldn’t do, including work that involves interpretation, ambiguity, and judgment. That creates opportunities in government that were barely imaginable a few years ago.

But it doesn’t repeal the old rule.

In fact, AI may make understanding the underlying process more important, because a traditional system often breaks when the rules don’t make sense. AI may keep going.

Citizens want responsive government, not more apps. AI can help deliver the first thing, but it can also become “one more app.” The difference isn’t the technology. It’s whether you understood the work before you automated it.

Interested in how AI can fit into your organization?