There are basically two vectors I use to understand the types of AI I might want to deploy inside an org and what is most likely to succeed. I must have spoken to over 50 orgs about AI adoption now and I still see no (or at best, a muddled) distinction between agentic and deterministic ideas. To me that is the critical fork point and it makes sense once you engage with the ideas.
Deterministic vs Agentic
Deterministic means we want strict control over how the AI makes decisions, the process it follows and consistency in the outcome. We have a pre-defined process and you can picture a flowchart, i.e. tagging a customer service ticket and routing it based on the tag. At implementation we break each step down into a machine and chain them together into a factory line. The flow runs to spec every time, never deviates and is highly consistent. Naturally this only works for pre-defined work or process.
Agentic is when we hand the machine a goal and let it figure out a pathway, possibly with soft constraints. The thinking is delegated to runtime. Agents can deliver more with less definition and can surprise you with their over-delivery on a task. Agents are also prone to hallucination, lack consistency on goal, can make mistakes that are hard to capture and require a more sophisticated organization to manage them.
Agents are cooler right now but there is a clear trade off around the type of task and consistency expectations. Free-flowing creative tasks make a lot of sense in agent land where a power user can direct them, but a process that needs to run reliably n times per month without thinking about it ever again fits a more deterministic approach.
Both have differences in how we build them. I'll release a blog next week on Agents, which typically do have a higher upfront investment and maintenance cost over the medium term.
Organization Levels 1, 2, 3
Now let us shift focus to the organizations we are talking to and where they are on the curve.
Level 1 is the baseline of where most orgs are today. Users are early on the curve and there are no full process replacements yet. Here we stay focused on replacing deterministic processes where we can build momentum, define the risk upfront and provide a measurable financial outcome. Agents seem to fit here but let's be honest, we are not ready and that situation is like giving a dog a gun. L1's goal is measurable wins that prove the thing works.
Level 2 is an org starting to shift. Power users are emerging, there are some deterministic process replacements and an AI council supporting varied implementations. We might experiment with Agents here but should focus on getting as much deterministic process replaced as possible. Focus on building components that stack to a better whole and prioritize ideas we might recycle later.
Level 3 is the nirvana; a competent team and lots of existing replacements. The funny thing is once you get here Agents are not that cool anymore. They are a tool just like deterministic flows and you see how it all snaps together.
Pulling it together for ideas
A lot of orgs we speak to have lists of well meaning ideas and a flat matrix to evaluate effort and payoff. They do not have maintenance and reliability in that matrix - and what those needs are per implementation.
When I put the two vectors together it makes a lot of sense where an org needs to focus and they are too early on the curve for lofty ideas. I have never hid my baseline assumption: an org should spam as much dumb, stupid, easy automable shit in level 1 as possible. Printing value, building momentum, getting battle scars as a team and learning what you are doing. I think back to early internet: it paid off more to invest in an online process replacement than ecommerce upfront.
The Deterministic vs Agentic distinction is missing in almost every org we speak to. To me that is one of the groundwork things in any matrix: it impacts how, cost, reliability and future maintenance lift.