Large language models have become ubiquitous and affordable, and their positive impact on software development is well documented. A great outcome of the new tools is that they have lowered the barrier for coders and non-technical experts alike to apply their knowledge in ways that were previously impeded by the toolsets and incompatible vocabularies. Consequently, the value of human judgement, from a problem solving and business perspective, has gone up dramatically.
Contrary to click-bait headlines and political free throws, we think the new AI economy will make expertise more precious than design, development, and implementation. Just because AI can perform some tasks that previously required specialization, doesn’t mean, as Rogé Karma wrote recently, that AI is yet capable of “dealing with unpredictable situations, meeting subjective measures of success, acting on tacit knowledge, and navigating complex webs of human relationships.” This has been our experience too.
Indeed, integrating knowledge across many sectors, and creating new solutions to new problems we have not seen before, has become the dominant theme. Expertise rules.
In this brief article we’re not going to address the societal implications, from employment opportunities to moral impact, of the new AI paradigm, though we have plenty to say about it, as do others. Of course that doesn’t mean that anyone, from junior tech writers to senior architects, can maintain the status quo ante. Far from it.
Instead, we want to make a narrower argument, as the engineers we are, about our experience in the creation of a suite of new products in data governance, management, analysis, and learning platforms. For background, Pradeep is a product design leader and software expert. And Peter is a grey-haired technology executive with, in the immortal words of Howard Cosell, “diminishing skills.”
The first LLM-assisted product we and the Amida team brought to market is a new software application, TorQE, which fundamentally changes how states can report Medicaid’s and Children’s Health Insurance Program (CHIP) data. It was created by people who understand Medicaid, T-MSIS, data quality, and the real world challenges faced by state agencies who have to report their results to the Center for Medicare and Medicaid Services (CMS). AI helped us move a lot faster, it made the previously infeasible, feasible, but we provided the domain knowledge, judgement, and problem definition.
If you know about T-MSIS, you’re going to fall off your chair.
And if you don’t, this post is still for you: the how’s are way more interesting than the what’s.
The problems with T-MSIS are well known and have been “admired” for years. If AI alone were sufficient to solve them, there would already be successful T-MSIS validation products in the market. There aren’t. That’s because the problem isn’t the code itself, it’s knowing which data elements are wrong, why they’re wrong, and what a state agency actually needs to do differently with the staff and systems they already have.
The gap between that depth of knowledge and the pace at which we could turn it into a product, with excellent documentation, and rock-solid evidence of compliance (this is not quite the elixir of immortality, but, from a time savings perspective, it is close!) was a constant friction, in time, money, and opportunity.
For us, and for many in our industry, the capacity gap between product definition and complex software development abruptly closed over the last few months. We built a comprehensive (and therefore commercially disruptive) platform, with our partners and for our customers at a pace, and a quality, we could not have sustained, never mind afforded, as recently as twelve months ago. Our LLM of choice is Claude.
Of course, no AI can do this work without human supervision. Drop the model into the Medicaid world cold, without context, and it flails and fails. (Believe us, we’ve tried.) The regulatory history, the institutional scar tissue, the “we tried that in 2016 and here’s exactly why it failed,” that knowledge is ours, and it is everything.
As we discovered with TorQE, and what we observe in Medicaid Enterprise Systems (MES) modernization, is that the arrival of LLMs changed what new skills were needed, and where conventional expertise is best leveraged and focused. Of course, it also subtly shifts the power away from the incumbent MES vendors, and back to State Medicaid agencies, where it belongs. When writing code becomes democratized, the competitive advantage is process acumen, business experience, intellectual curiosity, and, yes, human judgment.
And this is why we thought it would be timely to talk about our experience, because it doesn’t match the blaring and dark headlines. As a counter example, our CEO has learned more about System and Organization Controls 2 (SOC 2) audit requirements in two months than two decades in this industry ever taught him. One of our junior data engineers, who knew nothing about front-end development, now ships React applications and wires up APIs as if he has done it forever. And one of your authors, who entered industry writing assembly code for x86 processors, now writes shell scripts for the first time in a quarter of a century. And they work!
With proper guidance and clear rules of the road, our people have not outsourced their thinking, they have expanded their range.
Ted Chiang wrote that “[until] recently, we might have thought that writing computer code at a professional level could be done only by a mind that had subjective experience. Now it appears that LLMs might be able to do this, but we don’t need to attribute subjective experience to them; we can simply acknowledge that we hadn’t anticipated that writing computer code could be treated as a pattern-matching task solvable by huge amounts of computational horsepower and a vast data set of code repositories.”
This is exactly the counter-narrative we wanted to tell. The question is not whether AI is replacing business expertise, because it absolutely isn’t, and we believe it never will. It is fundamentally a prediction machine that is incredibly good at pattern matching. It follows, and can even derive, rules for games and for physics.
We love it not because it is magical, but because it makes us more productive.
And it has changed who gets to scale their expertise, and at what speed. Historically, turning deep domain knowledge into a product required capital, large teams, and long development cycles. That naturally favored organizations with deeper resources.
Now the bottleneck is finding the people who understood the problem deeply enough to define it, and who know what to build. GPS doesn’t tell you where to go, as the expression goes, “if you don’t know where you’re going, any road will take you there.” But it is great at figuring out your route. True, it wasn’t wonderful for TripTiks. But the overall benefit to society is measured in the trillions of dollars.
So, too, will it be for LLMs in software development. The winners will be those with the skills to describe attractive destinations, and business goals, so the software can take us there.