Every Verb Makes You Choose

The question that shows up right after you name the verb.

The short version (by humans, for busy humans)

We put up a piece earlier this week arguing that any piece of work breaks into a chain of verbs, and each verb points at a capability: a rule, a model, a language model, or a person. The day it went up, we were in a design review where naming the verb wasn't the hard part anymore.

The hard part was the second question hiding inside every verb: once you know what it points to, how much room do you give it?

Rules are predictable going in and rigid once they're written. Agents are flexible and only traceable after the fact, through their logs. Most teams make that trade-off once, for the whole project, when they should be making it once per verb. That's usually a default deciding for you, not a choice anyone actually made.

If you want where that thought went, it's below. If you haven't read the original piece yet, that's the better use of your next five minutes.

The full version (human on the loop, for depth, yours or your AI's)

Opening scenario

The day after we put the verb piece up, we were in a design review for a piece of our claims intake pipeline, nothing dramatic, the step that decides which queue a new claim lands in before anyone touches it. Someone on the team had specced it as an agent: read the claim, weigh a handful of signals, decide the queue, log the reasoning.

We found ourselves doing the exercise from the piece almost automatically. Read is a language model. Weigh the signals is either a rule or a model, depending on whether those signals are things we can score from loss history or things we already have a threshold for. Decide the queue, if the thresholds are known, is a rule wearing an agent's clothes.

That's when we noticed we'd stopped at the easy question. Naming the verb only told us what it could point to, not what we should actually give it, and that second part turned out to be the harder half of the exercise.

The problem

Naming the verbs in a workflow is the discipline that gets you the map. What the piece didn't spend much time on, because it wasn't trying to, is the choice that shows up right after the map is drawn: for this specific step, how much predictability are you willing to give up for flexibility, and did you make that call on purpose?

Most shops don't make that call step by step. They make it once, at the project level: "yes, let's use AI for this." Then whatever the vendor's default configuration happens to be becomes the answer for every verb inside that project, whether or not each one earned it.

That's an expensive way to end up over-powered on the steps that didn't need it, and under-examined on the one step that did.

Here's what we keep coming back to

The verb piece already names this same trade, in passing, and moves past it. We wanted to sit in it for a minute, because it's the part we think a lot of teams skip.

1. The trade is real, and it isn't a spectrum of buzzwords

Rules are predictable in advance, auditable off the page, and rigid. Agents are flexible, traceable after the fact through their logs, and hard to predict going in. That's the actual shape of the risk you're taking on, one step at a time, not a marketing distinction. Cynthia Rudin's research on high-stakes decisions makes a version of this same point outside insurance entirely: interpretable models beat black boxes precisely when someone has to explain the decision after the fact, not just get it right. [Source: Cynthia Rudin, "Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead," Nature Machine Intelligence, 2019]

The design review we mentioned defaulted to "agent" for a queue-routing decision that had three known thresholds behind it. Nobody chose that deliberately. That's where the tooling landed. A default is not a decision, even when it feels like one because it shipped.

What this means for your Tuesday morning: the next time a step on your roadmap gets labeled "agentic," ask who decided that, and what they were trading away to get there.

2. The industry already knows how to run this trade correctly, in one place

Filed actuarial models are the proof. They sit in the middle of the line on purpose, interpretable because regulators require it, not because anyone thought interpretability was a nice-to-have. That's a discipline the industry has practiced for decades in ratemaking. It hasn't been carried over to every other AI decision on the roadmap. [DL analysis]

What this means for your Tuesday morning: if your actuaries can explain why a rate model sits where it sits on the predictability-to-flexibility line, ask the same question about the underwriting or claims workflow next to it. Chances are nobody's asked yet.

3. The mistake is making the trade once per project instead of once per verb

A single workflow should mix all three: a rule here, an interpretable model there, an agent somewhere it's actually earned. Treating the whole workflow as one trade-off is where the risk creeps in, either you over-buy flexibility on a step that just needed a rule, or you under-buy it on the one step where a rule was quietly standing in for a judgment call nobody wrote down.

What this means for your Tuesday morning: for the workflow you mapped after reading the verb piece, go back and ask, verb by verb, "did we choose this, or did we inherit it?"

4. Most steps stop at the first or second rung, and that's healthy

We're putting up a carousel this week that walks this exact ladder, rule, model, language model, agent, stop at the first yes. Worth a look if you haven't seen it. Most steps in a real workflow stop early on that ladder, and that's not a failure of ambition. Reaching for more power than a step needs buys you risk you didn't have to take. That said, this doesn't apply evenly. Some steps, a certificate-of-insurance request is the usual example, have no real judgment hiding in them at all, and running this exercise on those wastes a meeting.

What this means for your Tuesday morning: if every step on your AI roadmap lands on "agent," that's worth a second look, not a celebration.

So what

Naming the verb answers what a step could point to. It doesn't answer where it should actually sit on the line between predictable and flexible, and that second question is the one we think most teams skip, because it feels like a detail instead of a decision.

It's most of what AI strategy actually is, one step at a time, not a detail. And it's worth knowing now, because the same trade you're making quietly today is the same one you'll eventually have to explain out loud. That's a different post, one for another week. For now, the useful move is smaller: pick one verb you already named, and ask whether you chose where it sits, or whether something else chose it for you.

A note before you go

If you haven't read the piece this is responding to, start there: it's the one that actually teaches the method. This is where that thought led next.

We'd like to hear where yours goes, so reply here or find us on LinkedIn.