Every organization adopting AI hits the same fork: bolt AI onto existing workflows (fast, safe, incremental) or rebuild the workflow around AI (slow, risky, transformative). At Anthropic's Claude for Science briefing, Bristol Myers Squibb CEO Chris Boerner described a third path: let "a thousand flowers bloom" first — broad experimentation — then rebuild around what actually works. Toggle all three below.
Drag to rotate. Gray blocks = human process steps. Gold = AI. In bolt-on mode AI hangs off the side of an unchanged pipeline; in rebuild mode the pipeline is redesigned with AI at the core and humans at the judgment points.
The case for bolt-on
Attach AI to steps that already exist: draft the email, summarize the document, triage the ticket. No process redesign, no retraining, value in weeks. This is where nearly everyone starts — and where most measurable gains live today (studies of AI coding and support assistants report 14–35% task-time reductions for augmented workers).
The trap: you inherit the old process's shape. If a 9-step approval chain was designed around human bottlenecks, making each step 20% faster still leaves you with a 9-step chain. Efficiency compounds; transformation doesn't happen.
The case for rebuild
Ask what the workflow would look like if AI had existed when it was designed. In drug development — BMS's world — that means AI doesn't just summarize trial documents; the pipeline itself is restructured so models generate candidate analyses continuously and humans move to validation and judgment.
The trap: rebuilding before you know where AI is reliably strong burns credibility and capital. That's the logic of "a thousand flowers bloom": run wide, cheap experiments first, measure which use cases stick, then commit to redesign where evidence is strongest.
Decision guide
| Signal | Favor bolt-on | Favor rebuild |
|---|---|---|
| Task error tolerance | Low — humans must stay in the loop each step | High or verifiable — outputs can be checked cheaply |
| Process age | Recently optimized | Designed around pre-digital constraints |
| Evidence base | Few internal pilots yet | Pilots show consistent wins in this workflow |
| Bottleneck type | Individual task speed | Handoffs, queues, and coordination |
| Competitive stakes | Cost center | Core differentiator (e.g., R&D throughput) |
Rule of thumb: bolt on to learn, rebuild to win. The sequencing — experiment broadly, measure honestly, redesign narrowly — matters more than picking a side.