An agent dropped into an unchanged process automates that process’s failure modes at machine speed. The redesign that would fix it is exactly the thing individual tool adoption does not produce: the strongest field experiment we have found no change in the composition of anyone’s work at all.
Why this matters
The cleanest evidence in this dimension is a null result. A randomized six-month deployment across 66 firms and 7,137 knowledge workers found that treated workers spent about two fewer hours per week on email and worked fewer out-of-hours, and that the quantity and composition of their tasks did not shift. The authors’ explanation is the mechanism this hub exists for: email is solitary, so one worker can change her own process; meetings and document ownership require renegotiating norms with other people, and nobody renegotiates norms because a tool arrived. (Microsoft researchers, Microsoft telemetry, 2023–24 Copilot, pre-agentic; it measures time use, not output quality.)
The correlational finding layers on top: organizations reporting significant AI value are 2.8× more likely to have redesigned individual workflows rather than bolted AI onto the existing ones (55% against 20%). That evidence is cross-sectional and self-reported and leaves reverse causality open, since firms already capturing value are the ones with budget and mandate to redesign; take the direction from it and not the arrow.
So process redesign is the complement that makes agents pay, and the ordinary mechanism of adoption fails to supply it. That is a co-invention story economists have told about every general-purpose technology since the dynamo, with one twist that matters here: generative AI needs no hardware, so individual adoption runs ahead of the firm rather than behind it. Staff are already integrating agents into processes the firm has not redesigned, one desk at a time.
Two rules that make redesign tractable at fund scale:
- Triage before redesigning. Standardized, low-variability workflows (expense approvals, password resets, reconciliations with clear rules) are fine to automate as-is. Redesign is worth it where coordination overhead is high, sequences are rigid, and humans make decisions that could be data-driven. Ignore the accompanying modelled uplift ladder: 5–10% / 20–40% / 60–90% is a hypothetical call centre labelled “estimated impact,” with no fund analogue.
- Split owner from verifier. The person who built the agent is structurally the worst person to check it, and a verification step that shares the agent’s failure modes verifies nothing. Make hand-off and escalation points explicit: where the agent stops, who picks up, and what evidence travels with the hand-off. The canonical illustration is an accountant who stops doing the matching and starts setting variance thresholds, exception rules and review triggers; it is an illustration, not a case study.
The counterweight, and it is a real disagreement. The redesign-first position comes mostly from consultancies and from one argument-by-analogy essay. The citizen-development literature says the opposite for this population: automating an existing messy process still delivers value and is the most likely early citizen scenario. Our read is that both are right at different tiers — automate the mess at tier 1, redesign before anything reaches tier 3 — and that a firm which insists on redesign before any automation will simply watch the automation happen without it.
The evidence base for this dimension contains nothing from a fund. The examples are a university HR repository, a modelled call centre, a 2023 RPA-era cohort of large enterprises, and generic knowledge work on early Copilot. The one fund-adjacent example in our sources is hypothetical, so the mechanisms are the part worth borrowing and the numbers are not.
Where you stand
| Level | Looks like | Cheapest next move |
|---|---|---|
| Crawl | Agents are bolted onto existing steps. Nobody has written down where the agent stops and a human starts. | Document the hand-off for one live agent: what it does, where it stops, who checks, and what evidence they see. |
| Walk | Hand-offs documented for important agents. Verification is done by the builder. | Move verification to someone with a different failure mode, and record what they caught. |
| Run | Processes triaged into automate-as-is versus redesign; owner and verifier are different named people; escalation paths tested. | Pick one high-coordination process and redesign it around the agent rather than inserting the agent into it. |
| Fly | Sources of truth are machine-readable, interfaces are API-first, and process changes are reviewed when the agent’s scope changes rather than annually. | Re-verify that the process still has a human decision point that would catch a confident wrong answer. |
Concerns this dimension covers
- Unreliable output acting on your systems — fabrication meeting a process with no verification step.
- Quality debt and orphaned apps — busyware and key-person dependency, the process residue of unmanaged builds.
- Oversight decay — what happens to a verification step nobody budgeted time for.
- Recordkeeping and compliance gaps — the process step that used to generate the record and no longer does.
Controls that answer them
- Human approval gates — the designed boundary a redesigned process places deliberately, where judgment is possible and damage is uncontainable.
- Logging and audit — process integration leaves a decision trail or it did not happen.
Who does the work
- Citizen developer roles — where the owner/verifier split is defined and assigned.
- The reviewer pool — who staffs the verifier half when the builder cannot.
Sequencing and where this is checked
- Day 3 sequencing — process redesign is a Day 2 move that Day 1 visibility makes possible.
- 90 — where a redesign gets tested on one process before it becomes a program.
- Promotion gates — an agent enters a business process through the gate, and the hand-off design is what the gate inspects.
Open questions
- Does agent-readability work (plain-text sources of truth over PDFs) have measurable returns? It is asserted as the immediate change with the highest return, by an author describing his own lab, with no before-and-after numbers.
- Nobody has tested whether an owner/verifier split actually catches more than a diligent builder does. The argument is structural and the evidence is absent.