economic
Dark Factories

The J-Curve of AI Adoption

The productivity pattern where adding AI tools to existing workflows causes an initial drop before eventual improvement. Most organizations are stuck in the dip. BCG research quantifies two modes — "Deploy" (tools only, 10-15% gain) vs "Reshape" (workflow redesign, 30-50% gain). The METR 2025 RCT shows the dip is real and measurable.

The J-curve is the shape of the productivity curve when you add AI tools to an existing software development workflow:

Productivity
     │
     │
 ─── ┼ ─────────────────────────────── baseline
     │         ╱──────────────
     │        ╱
     │       ╱
     │      ╱
     │     ╱
     │ ─╲─╱
     │    ╲
     │     ╲
     └──────────────────────────► time
          add    dip    recovery
          tools

Why It Happens

Adding powerful AI tools to an existing workflow creates friction before it creates flow:

  1. Cognitive overhead: Deciding when to use AI vs. write it yourself
  2. Context switching: In/out of AI interfaces breaks flow
  3. Verification overhead: Reviewing AI output takes time (often more than writing it)
  4. Trust calibration: Learning which suggestions to trust and which to fix
  5. Workflow misfit: AI tools designed for clean tasks; real work is messy

The METR 2025 study captured this: experienced developers took 19% longer with AI tools, while believing they were 24% faster. The workflow disruption outweighed the generation speed.

Escaping the Dip

Teams that achieve 25–30% productivity gains don’t escape the J-curve — they redesign around it.

The difference is structural. Adding Copilot to your existing workflow = staying in the dip. Redesigning your entire development process around AI = getting past it.

What workflow redesign looks like:

  • Different ticket structure: Tasks designed for AI execution, not human execution
  • Different review process: Evaluating outcomes, not reading diffs
  • Different CI/CD: Pipeline designed for AI code at volume
  • Different meetings: Eliminating coordination that only existed for human implementation
  • Different hiring: Prioritizing specification ability and judgment over coding speed

BCG: Deploy vs. Reshape

BCG’s “AI at Work 2025” report is the most rigorous quantification of the two adoption modes:

ModeDescriptionProductivity Gain
”Deploy”Add off-the-shelf AI tools to boost existing workforce10–15%
“Reshape”Reimagine functions through full workflow re-engineering30–50%

The 3–5x difference between modes explains why results vary so widely across organizations using ostensibly the same tools.

BCG’s headline findings:

  • 72% of respondents use AI regularly — but only the “Reshape” subset captures the full value
  • 62% of respondents report at least 25% productivity increase after workflow redesign
  • Some organizations achieving 80% gains in specific functions
  • “AI high performers” are 3x more likely than peers to be scaling agent use

The dip is most pronounced for organizations with the longest-standing workflows, deepest legacy systems, and most layered hierarchies — which is most large enterprises.

Developer-Specific Data

McKinsey’s software engineering research (2025):

  • Documenting code: 50% faster with AI assistance
  • Writing new code: ~50% faster
  • Code refactoring: ~1/3 faster
  • Average time saved: 6 hours/week per developer
  • Top teams: 16–30% improvement in productivity, quality, and time to market

GitHub Copilot in controlled trials:

  • Task completion time dropped from 2h41m → 1h11m (55.8% faster) in controlled settings
  • But: developers require ~11 weeks to fully realize gains — the ramp is real
  • Trust problem: 46% of developers don’t trust AI output accuracy (Stack Overflow 2025), up from 31% in 2024 — adoption is rising while trust falls

The Harvard/BCG “Jagged Frontier”

The foundational complexity: AI doesn’t uniformly speed everything up.

  • Junior analysts on well-defined tasks: 30–40% efficiency gains
  • Experienced staff on well-defined tasks: 20–30% efficiency gains
  • Complex tasks where AI output was accepted without sufficient critique: ~23% dip in performance

The “jagged frontier” — where AI dramatically accelerates some tasks while degrading others — explains why workflow redesign matters. The critical organizational capability is knowing which tasks to route to AI vs. humans, not just having AI available.

The Measurement Problem

The METR study finding (19% slower, believes 24% faster) reveals the measurement problem:

  • Developers feel faster because generation is fast
  • But end-to-end task time is what matters — including review, debugging, integration
  • The feeling of speed isn’t correlated with the speed of shipping

This is partly why the J-curve is hard to escape: it feels like AI is working, so teams stop investigating whether it actually is.

What “Redesigning Workflows” Actually Means

The organizations reporting 25–30% gains have done the hard work of asking:

  • What decisions do humans actually need to make here?
  • What coordination exists only because humans are doing the implementation?
  • What review process makes sense for AI-generated code at volume?
  • How do we hire differently to staff this new structure?

This is organizational change management, not technology adoption. Most engineering managers are trained for technology; few are trained for organizational design.

Level-Mapping the J-Curve

  • Level 0–1: Deep in the dip. Added tools, haven’t redesigned. METR territory.
  • Level 2: Partway up the curve. Some workflow redesign. Some gains.
  • Level 3: Past the bottom. Meaningful gains. Humans reviewing features, not code.
  • Level 4–5: Past the curve. 25–30%+ gains. Full workflow redesign. Or: dark factory territory.