For the last 1 year, my company has been pushing hard on GenAI adoption. And rightly so. The productivity gains are tangible, everyday improvements to how we write, review, and ship code.
But recently, I ran into a scenario that made me wonder how this fits with one of our more established engineering processes: the promotion review.
Backing up a little bit, a few weeks ago, I was exploring a design for storing confidential data in one of our systems. I wasn’t going to implement a bespoke encryption system from the ground up but I did need to map out the options available within our infrastructure, evaluate the trade-offs, and settle on a path forward. I started the exploration with Claude and gave it the required context.
Within about 30 mins, the model did most of the exploration work. It laid out the available options, explained the trade-offs, and recommended the path I feel I would have ultimately chosen. I didn’t find anything in its analysis that I strongly disagreed with. I did go and verify its claims against the source-of-truth docs but it compressed what would have normally been days of reading internal wikis and working through the alternatives into a couple of hours.
The solution still needed to be checked against our actual constraints and I did need to adjust a few minor things. But the amount of useful work it had already done was hard to ignore.
From a productivity standpoint, this was a massive win. Later that week, as I was asked to write promo feedback for a colleague, and in that exercise I recounted what I had most recently done in my own design exercise.
The signal is disappearing for mid-level engineers
In many engineering career ladders, the move from a mid-level engineer to a senior engineer, often L4 to L5 or L5 to L6 (sometimes), depends heavily on a few behaviors:
- Navigating ambiguity.
- Evaluating somewhat complex technical trade-offs.
- Driving technical design and architecture.
- Having a strong sense of judgement.
A design document has traditionally been a useful piece of evidence for these behaviors. You could see which options the engineer investigated, the edge cases they considered, and the compromises they chose to make. When doing a promo review, even if you were not a reviewer when the design was being proposed, you could ask the original reviewers what changed during the discussion and how the engineer responded to feedback and questions.
Of course, even before AI, engineers get help from colleagues, reuse existing patterns, and incorporate feedback. A polished document has never been proof of independent reasoning and is not supposed to be. We should be extracting from existing body of work but the challenge is in exercising the right judgement on the best available path forward for the given problem being solved.
AI makes the evaluation of someone’s reasoning that much harder. A model can now produce a substantial part of the exploration and a convincing explanation of the recommendation before the engineer had to explore much of the ambiguity of the problem space.
When an L4 engineer submits an excellent design document that was heavily co-authored by Claude, what exactly can I infer from it for a promotion review? How much of the reasoning can the engineer explain, challenge, and adapt when the circumstances change?
The engineer may have shown excellent judgment in framing the problem, supplying the right context, and recognizing a good recommendation. Or they may have accepted something plausible that they don’t fully understand. The finished document can look very similar in both cases.
That is the part I find difficult as a reviewer - especially when I was not part of the reviews for the design artifacts. We are encouraging engineers to use a tool that improves the output while making one of our familiar signals harder to interpret.
Is L6+ an exception?
My initial reaction was that this matters less for Staff engineers and above, often L6+. Their evaluation usually draws on a wider set of evidence: identifying important problems, shaping technical direction, coordinating work across teams, and getting people with competing priorities to agree on a path forward.
AI can help write the technical brief before a difficult meeting. The engineer still has to understand the disagreements, negotiate the compromises, and take responsibility for the outcome.
That makes the signal less dependent on a few technical design documents. It doesn’t make those levels immune. AI can also help with strategy documents and stakeholder communication. The same attribution question exists; I suspect it is simply easier to answer when an engineer’s impact is visible across teams and over time.
The L4-to-L5 transition feels especially exposed because we often use technical design work as evidence that someone is ready to operate with less guidance.
I don’t know what we could look for instead.
If AI removes much of the legwork, the signal has to come from how an engineer uses and owns the result. But I don’t have a clean answer for what that looks like.
When I tried to think through the alternatives like relying on live design reviews, weighing final execution more heavily, or requiring a trail of decision records, they either feel insufficient or an onerous ask on them to collect all that.
We are asking people to use these tools, and using them effectively may itself demonstrate the judgment we’re looking for. I don’t want to penalize that efficiency.
Maybe we should just update the expectation on how someone operates itself for promo reviews. If a model can do the heavy lifting of exploration, perhaps “navigating ambiguity” simply means something different for an engineer today, and the career ladder itself needs to change.
I’ve reached out to one of our Principal Engineers to start this discussion, and I expect companies might take a while to change their guidelines, if at all. In the meantime, I’m curious how other teams are approaching this.
Are your promo rubrics changing? What evidence helps you evaluate “navigating ambiguity” when an AI model can do much of the initial exploration?