How McKinsey Plans to Survive AI (and Reinvent Consulting) (AI column)
Interesting half-hour episode HBR IdeaCast podcast featuring Adi Ignatius (HBR) Talking to Bob Sternfels (Global Managing Partner: CEO of McKinsey). The main idea revolves around the fact that AI is 30–50Percent "technology" and the rest is organizational design and processes. If you “pin the LLM on top of the old process,” there will be local automation, but not systemic growth throughput.
I looked with great interest and took the following insights that apply to software development.
1The barrier is not in the model, but in the organization Wells, unnecessary handoffs, layers of approvals eat the effect of AI more than the quality of the prompt. 2Human + Agent is a new unit of productivity McKinsey has a narrative inside of tens of thousands of “agents” and a move toward “their own.”1 human agent (However, in the interview it is not clear what is an "agent", so it is not clear what to compare the figures with.). That is, agents are included in the labor force and noted an increase in their number. 3Analytics commoditization → value goes into outcomes If the analysis becomes cheaper, the one who delivers the result and knows how to "subscribe" to the effect wins. Within companies it is similar: less "how many tickets closed", more "what outcome improved". 4Hiring/growth: resilience + learning> perfect score And This section is about changing hiring criteria within McKinsey. Previously, they were looking for excellent students, and then they realized that success is better predicted not by excellent grades, but by successful overcoming crises - it is better to take someone who falls and gets up, knows how to study and work with people. 5Speed wins, even if there are more errors - with a good recovery There is a direct parallel with good engineering practices: small batches, feature flags, observability, fast rollback → reduce the cost of mistakes and can be faster.
If we somehow transfer insights to the advice of engineers, we get about the following: 1)Stop using AI as a magic button. Agent → PR → quality gates → human review (Tests/linters/security scans are not canceled). 2) Look for where the agent removes the handoff rather than “speeds up writing.” Start with end-to-end flow (value stream). 3) Skills from the second wave: judgment + alternatives + counterexamples. Ask the model not "answer" but "risks/checks/what could go wrong"
For the techlids, that means the following.
- Think of AI as an operational transformation, not a set of pilots.
- Choose. 2-3 process and make the entire process (For example, triage → fix → deploy).
- Measure the result with outcomes-metrics + speed/stability balance: DORA (lead time, deploy freq, MTTR, change fail rate) + DevEx/SPACE (So that "acceleration" does not burn the team).
- Manage agents as products: inventory + ownership + audit (Who is the owner, what data sees what tools call, what gates must pass).
#AI #Engineering #Leadership #Devops #Productivity #Management