So many people (and some teams) work under mandates to “use AI.” While I am a fan of these two functions for the current LLMs:
- The ability to create fast prototypes for human assessment and judgment.
- And the ability to gain new insights into you, your team's, or your organization's data. (This often requires a lot more data than you might think, but that's a different problem.)
However, I am not a fan of an individual working alone with AI. Not if they are supposed to be part of a team producing a product other people will use. That's because typing speed never matters. The gating factor in product development is the team's ability to understand what the customers need, how fast the team learns from their work, and how fast the team can do all of that. (Those are all the feedback loops I described in Project Lifecycles: How to Reduce Risks, Release Successful Products, and Increase Agility.)
That's where value stream maps can help us see the team's reality. (If you have not yet read my previous series that starts with How to Use Value Stream Maps to Reinforce Agility & Effectiveness, Part 1 (Expert Teams), please consider reading that. I will not repeat what I already wrote.
Let me summarize the two big ideas in that series:
- When people work alone, the overall cycle time increases. That's because the team does not learn as fast as a single person does. But the single person rarely has all the necessary context.
- When people collaborate, the overall cycle time decreases. Collaboration allows the team to learn as fast as possible, retaining context and using judgment.
The same principles apply when people and teams use LLMs. If people and teams are not careful, they can create even larger bottlenecks.
Where LLMs Do Not Address Bottlenecks in Expert-Focused Teams:
Imagine this expert-focused team using AI the way many managers want people to use AI: as individuals.
Notice that the overall cycle time is still about 9 days. And that's with each person spending half the time they had previously. (I have not included any learning time here—I assume the people have now learned how to use their LLM of choice. Learning anything new takes time.)
Without AI, the overall cycle time is 9.35 days. With AI, 9 days. Can anyone tell the difference? That's because the wait times dwarf the work times. People are not available when the work is available for them. That's the bottleneck.
Why are people not available? Often, because managers believe in the fallacy of resource efficiency instead of the empirical data of flow efficiency. (See A Little Scree About AI and the Hard Parts of Product Development for more information.)
But even cooperative teams with working agreements might not reduce their bottlenecks.
Where LLMs Do Not Address Bottlenecks in Cooperative Teams:
What happens with a Cooperative team? It all depends on where the delays are. This team had delays for people to be available:
Let's assume again that AI can help people work twice as fast. That does not change the overall cycle time enough. Without AI, this team has a cycle time of 7.8 days. (I would call that 8 days). With AI, the cycle time is 7.5 days. Again, can anyone tell the difference?
Even cooperation does not manage the delays between everyone's availability. Only collaborative teams can manage those delays.
Collaborative Teams Might Use AI to Reduce Bottlenecks
To keep things “even,” I'm assuming that the collaborative team uses AI to reduce their work time:
Again, the team can find some reduced work time when they work with AI. But the overall cycle time does not change because the bottlenecks are not in the team.
The bottlenecks are between the work times, in the delays.
Value stream maps can help us see those delays and bottlenecks. Next up is how to diagnose the bottlenecks.
The Series:
- How to Use Value Stream Maps to See Where AI Creates Bottlenecks Part 1
- How to Diagnose the Bottlenecks In Your Team to Reduce Cycle Time, Part 2