Nobody Asked For an Agentic Horse

“To save time is one of the most intelligent human drives and I believe it is responsible for the development of the loom. Despite all of the advances in regard to speed and the changes the loom has undergone, the underlying principle has not changed in thousands of years.” — Anni Albers, On Weaving(opens in a new tab)

I was watching baseball this weekend and noticed that mixed in with the usual auto glass and stuffed-crust pizza commercials were AI ads. A lot of AI ads. One presented its offering as a crab claw that could help you grab onto whatever software task you need to do. Another abstractly called itself an “agentic platform,” maybe hoping to sell with FOMO. Another talked about how it could update your spreadsheets for you. 

As someone who works in small business tech, that one especially caught my eye, because the protagonist was a small business owner—bravely soldiering on in a warmly lit pizza shop, surrounded and outnumbered by dark skyscrapers (I did wonder how many of said ominous, anonymous skyscrapers might house AI companies).

One of the things I find fascinating about this moment in technology is that while the vast potential of AI is widely recognized, there’s a tremendous shortage of imagination in how it’s actually being employed. As in those ads, it's easy to view AI as just a tool: one that enables us to do whatever we are already doing, just faster and cheaper—the apocryphal faster horse(opens in a new tab). But what’s the point of a horse if it doesn’t get you where you need to go?

The tools trap

In the standard SaaS model, access to software is the service a company provides. That’s useful when there’s a job or a task the user needs to complete for themself. A document editor can be a tremendously useful and satisfying tool for writing (if you enjoy the human activity of writing).

Unfortunately, many SaaS companies have made the mistake of imagining every problem as something customers can do for themselves, if given the tool. 

The more companies see problems this way, the more tools they build. The more resources they want to invest in building tools, the more they need to deflect support costs to self-service.

As increasingly self-service software tools have been applied to a wider range of more complex use cases, they’ve become harder to learn and use. As increasingly complex tools have proliferated, they’ve become increasingly dis-integrated with one another, resulting in increasingly frustrating experiences for the people who use them. There’s a reason the word “enshittification” was the word of the year in 2023(opens in a new tab).

Truth to material

If you’ve never read Christopher Alexander’s The Timeless Way of Building(opens in a new tab) (or if you haven’t read it in a long time), there’s never been a better time to give it a look. One of Alexander’s core arguments is that “there is a central quality” in everything, that “takes its shape from the particular place in which it occurs.” 

Alexander extends this notion from architecture to systems: “A system has this quality when it is at one with itself; it lacks it when it is divided. It has it when it is true to its own inner forces; lacks it when it is untrue to its own inner forces. It has it when it is at peace with itself; and lacks it when it is at war with itself.” 

He’s getting at the core of what design is—what the Bauhaus designers referred to as “truth to material.” And this is a big part of what’s wrong with most AI tools at the moment: they aren’t true to the material of the underlying technology. 

A chatbot presents a computer as if it were a person, rather than letting the computer behave like a computer. Teaching a computer to sound like a human isn’t going to make it a human, anymore than applying a marbled finish to plastic gives it the structural qualities of marble.

And if the computer in question is providing you with software-as-a-service, layering a conversational veneer on top isn’t going to fix whatever was broken or confusing about that software in the first place.

Service as software

What we're working on at Justworks is very different. We don’t just make software tools, nor are we simply a servicing business. Our customers pay us to own complex problems end-to-end in HR, hiring, compliance, and insurance. 

A medical records system might not add a newborn baby to her parents’ healthcare before issuing the hospital bill. A generalized LLM might give someone the wrong answer about termination law. A country may change its tax regulations without providing notice (let alone an API). The real world is messy.

Handling these kinds of problems for our customers means we have to look at the problems around the problem. We have to design complete service flows that tightly integrate our customer-facing software to internal tooling, to automated workflows, to humans with specialized subject matter expertise.

Justworks has always been “agentic,” because we’ve always been an outcome-driven business. We have to solve the whole, hard problem. We deliver service as software, not the other way around.

AI doesn't change that, but it gives us something new and powerful to work with. AI itself is software, which we can use as a tool. But we can also use it as a material to make software out of.

Our Experience Design team is approaching AI this way, by leaning into its material affordances and qualities with openness and curiosity. Rather than trying to make software behave like a human, we’re asking how we might use AI for what it’s good at.

• Instead of defaulting to conversational chat interfaces for every use case, we're exploring ways to let users make natural language requests where it's appropriate, and enabling the system to respond with contextually-appropriate UI that makes the next action simpler;

• Instead of forcing users to navigate a fixed menu structure that forces them to think in terms of our system architecture, we're exploring ways to use what we know about that customer’s context to focus their attention on the most important things;

• Instead of automating away human agents with software, we're using AI agents to proactively identify problems before they occur, and to tighten integration between our backstage servicing teams.

The goal remains the same: delivering on the whole outcomes our customers need. AI is simply a new material that lets us do that in new ways that enable us to deliver on those outcomes better. 

At Justworks we see AI as an opportunity to show up as a fundamentally different kind of company—a well-designed company that helps our customers handle the messy reality of the world, rather than asking them to muddle through it themselves.