Compression has a different shape in each domain
Chuck Ros published a piece last week on why we measure AI's impact at the wrong altitude. Jobs are portfolios of tasks, and AI compresses each one differently, so occupation-level claims like "engineers are 30% more productive" average away the detail leaders need. He's right, and I want to take it one level further.
A generic compression matrix puts all knowledge work on one map. Compression has a domain shape. Map the product development lifecycle and go-to-market separately and they disagree, which changes how you staff and where the next limit shows up.
So I ran it. The bands below are directional estimates, drawn from published research and calibrated against shipping AI-native product since March 2025. They describe the industry, and your team is a separate question.
| Task | Judgment | Compression | |
|---|---|---|---|
| Product operations / SDLC | |||
| Status and cycle reporting | Standups, cycle summaries, dashboards | Low | 80 to 95% |
| Documentation and release notes | Tech docs, changelogs, SOPs | Low | 70 to 90% |
| Test authoring | Unit and integration coverage, eval cases | Low | 70 to 90% |
| Code generation | Features, refactors, boilerplate via agents | Medium | 60 to 90% |
| Spec and PRD drafting | Requirements, acceptance criteria, context docs | Medium | 50 to 80% |
| Customer signal synthesis | Feedback triage, interview and ticket rollups | Medium | 50 to 80% |
| Code and design review | Judging agent output, catching what matters | High | 20 to 40% |
| Roadmap prioritisation | What is worth building, sequencing, tradeoffs | Very high | 10 to 30% |
| Architecture decisions | System design, build vs buy, model routing | Very high | 0 to 10% |
| Ship / no-ship call | Final quality bar, accountability for landed | Very high | 0% |
| Go-to-market | |||
| Lead research and enrichment | Account lists, firmographics, contact data | Low | 80 to 95% |
| Competitive intelligence | Gathering and summarising market moves | Low | 70 to 90% |
| Content production | Blog, social, email, SEO drafts | Low | 70 to 90% |
| Campaign assets and creative | Landing pages, ads, visuals, variants | Low | 60 to 90% |
| Sales collateral and battlecards | Decks, one-pagers, objection handling | Medium | 50 to 80% |
| Outbound sequencing | Personalised sequences, high volume | Medium | 50 to 80% |
| Pipeline reporting and forecasting | Rollups, board slides, forecast drafts | Medium | 40 to 70% |
| Positioning & messaging | Category framing, narrative, what not to say | Very high | 10 to 30% |
| Pricing & packaging | Tiers, value metric, margin structure | Very high | 0 to 10% |
| Sales conversations and negotiation | Discovery calls, closing, relationships | Very high | 0 to 10% |
High >50%Moderate 10 to 50%Minimal <10%, human judgmentDirectional estimates, not evidence-scored
Read the halves together and notice where the green sits. The lifecycle compresses in the production middle, across code, tests, documentation and first-draft specs. Go-to-market compresses at the top of the funnel, in account research and content volume. In both, the work that compresses is work whose output was already text, code or structured data.
The first finding is that each domain hits its constraint in a different place. In an NBER study of AI coding tools, developers using agents produced several times more code and 65% more pull requests. Releases rose only 20%, as the constraint moved from writing to review, integration and approvals. In go-to-market it moves to positioning, pricing and the conversations that close, which need different owners in the room.
The second finding is that go-to-market compresses more and yields less. More of its volume work automates aggressively, and its human-advantage zone runs on trust and negotiation. Review capacity grows when you redesign the process, while trust grows at the speed of relationships. Lifecycle time saved can become throughput, and go-to-market time saved mostly buys volume that piles up against a relational limit. Microsoft's field experiment across thousands of knowledge workers found individual work compresses faster than collaborative work. Email hours fell by around a third for regular users, and meeting time barely moved.
The third finding is that both maps end at the same cell, the call to ship and the call to close, with zero compression, full accountability and one owner. When one task gets dramatically faster, another task becomes the limit. If you're redesigning roles around AI, start at that cell and work backwards.
Use a map like this directionally, and multiply band by exposure. 80% compression on a task worth 2% of the week is a rounding error. A 40% band on a task that eats a third of the week is the roadmap. Watch the constraint, since freed hours mean nothing if review, pricing or the sales calendar can't absorb them. More code, content and pipeline are production metrics, so score whether more of what shipped landed. The bands can't show where your constraint sits this quarter, so read your own signals, task by task.
The code result is the NBER working paper "Writing Code vs. Shipping Code" (Demirer, Musolff, Yang). The email and meetings result is Microsoft's randomised field experiment with M365 Copilot across dozens of firms. The bands are my estimates, informed by that research. Chuck Ros's "We're Measuring the Wrong Thing" made the altitude point well enough to build on.