When the small number of large parts leaves me with parts still too big to estimate, I often take the first (in development order) of those parts and break it down again.
Can you estimate the first item? If not, you can decompose that first item in a similar fashion. Now you’ve got smaller, more easily estimated items. If it’s still too big, you can repeat this strategy.
In extreme cases you might need to go to a third level of detail, but realize that each time you’re increasing your uncertainty. That’s to be expected. The longer the time horizon, the more uncertain you are. You don’t want to forget about that uncertainty, though, lest you fool yourself.
I don’t think I’ve ever gone beyond three levels of doing this—but I could if it helps.
Breaking It down Another Level
"I know what we’re missing," Blaise said. "A simple end-to-end Walking Skeleton. Before we route a call based on the number dialed, let’s support a single customer service rep and route any call to them. Before we build a CRM customer search screen, let’s support only a single customer and pull them up automatically. This will connect the major pieces and make sure they’re integrated from the start. Then we can add all the other features to this basic framework."
"What size would that be?" Marion asked.
"I’d say another ’medium.’ It’s got some minor unknowns about the integration points, but that’s the point of it. If it weren’t for those, it would be tiny."
"Let’s break it down."
"That seems pretty simple. We need to…
I think that’s sufficient for now."
Once you get something small enough to estimate, then you can use relative comparisons and affinity grouping to estimate the other large parts that are not broken down in detail. This leaves you with a fairly high level of uncertainty, but does provide ballpark figures. And as you complete the early work, you can recalibrate your estimate with your actuals so far.
You can then extrapolate from that first item, both from its estimate and, later, from its actual. You will have built a really long lever when you do this. A small variation in the initial measurement makes large variations in the expected total. Hold this expectation loosely and check it along the way.
