The ABC Framework is MING Labs’ internal taxonomy for splitting work between humans and AI agents inside a hybrid organisation . Every role decomposes into three layers: A — judgment and relationships, B — structured expert work, C — routine operations.[S1]
The framework was extracted from operating MING Labs since January 2026 — including the failure that forced us to name it. When we fired an AI agent after thirteen days , the diagnosis was that we had given the agent a capability without assigning a layer to own.[S2] ABC is the corrective: every agent gets a domain, not just a verb.
The split looks like this in practice:
| Dimension | Human role | Agent role | Outcome owned |
|---|---|---|---|
| A — Judgment & relationships | Owner | None | Trust, accountability |
| B — Structured expert work | Director | Executor | Quality of analysis |
| C — Routine operations | Reviewer of exceptions | Owner | Throughput, accuracy |
Layer A is non-negotiably human: hiring, pricing, client trust. Layer C is non-negotiably the agent’s: inbox triage, data preparation, pipeline hygiene. Layer B is the contested middle and the place where most “AI initiatives” stall — without a supervision loop, B work either gets dumped on the agent (quality drops) or hoarded by the human (throughput stalls).
ABC is not a job-replacement model. It is a role-design model. The role stays; what changes is which layer each party owns. The framework forces a question most “AI strategy” decks skip: for any given role on the org chart, who owns A, who owns B, and who owns C? When the answer is “the human, with AI assistance,” nothing changes — the AI is a tool, not a colleague, and the productivity story stays anecdotal.
The opposite failure is just as common: assigning C to an agent without naming the outcome it owns. That is how Major Tom shipped 764 messages and zero value. The fix is not a better prompt; it is a better contract. ABC is the organisational counterpart to Agent Experience on the outside: clear ownership inside, clear legibility outside. Both ride on the same insight — agents need a domain, not just a verb.
