Capacity Planning with AI-Assisted Workflows

When we introduced AI-assisted development workflows into our programs, one of the first questions leadership asked was, "So we can do more with fewer people now, right?" The honest answer is nuanced — and getting capacity planning right in this new environment is one of the harder problems I have tackled recently.

What Actually Changed

AI coding assistants like Claude 3.5 and GPT-4o genuinely accelerate certain tasks. Boilerplate code generation, test scaffolding, documentation drafts, and code review assistance all got faster. For some tasks, the productivity improvement is significant — I have seen thirty to forty percent time savings on routine implementation work.

But not all work is routine implementation. Architecture decisions, complex debugging, stakeholder negotiations, and requirement discovery are largely unaffected by current AI tools. If your capacity plan assumes uniform productivity gains, you will over-commit and under-deliver.

My Framework

I break work into three categories for capacity planning purposes.

AI-accelerated work: Routine coding, testing, and documentation where AI tools provide measurable speed improvements. I apply a conservative productivity multiplier of 1.2x — meaning I assume twenty percent more throughput, even though some individuals see higher gains.

Human-intensive work: Architecture, design, complex debugging, and stakeholder-facing activities. No productivity adjustment. These tasks are bottlenecked by human judgment, not typing speed.

Hybrid work: Tasks where AI assists but human judgment is still the critical path — like code reviews where AI flags issues but a human must decide on action. I apply a 1.1x multiplier here.

The Trap to Avoid

The biggest mistake I see is treating AI productivity gains as a reason to reduce headcount preemptively rather than increase throughput. In our case, we did achieve a thirty-percent team size reduction while maintaining velocity — but that happened organically over months as we understood which roles could be consolidated, not as a Day One mandate.

Capacity planning in the AI era requires more granularity, not less. You need to understand your work mix at a level of detail that many program managers are not used to. But getting it right is the difference between realistic commitments and chronic overcommitment.