The Relay, Not the Replacement: Human–AI Workflows
By the end you'll be able to
- Decompose a real task into steps and classify each by AI-fit: delegate, retain, or draft-then-review.
- Design handover points — what the human briefs downhill, what the human checks uphill.
- Evaluate a workflow by its weakest checkpoint, not its most impressive step.
- Predict
- Allocate
- Decide
- Field note
⚡ 30-second warm-up — From Lesson 17: what's the three-question test?
Is this checkable? Would the diet plausibly cover it? What happens if it's wrong? Today that test becomes one half of a bigger design — the "uphill baton" in a workflow you build yourself.
Two colleagues use the same AI for the same report. One saves three hours and ships their best work; the other ships a confident error to a client. Same tool, same model. Predict what differed.
Start with the wrong question
"Can AI do my job?" is almost unanswerable, because almost no real job is one task — it's a bundle of very different steps. The better question is which parts of the bundle each of you should run.
Commit to a prediction
The colleague who shipped a confident error most likely…
This lesson changes one thing: you'll stop asking whether AI can do "the job", and start unbundling any task into legs — deciding who runs each one, and engineering clean handovers between them.
Discover the mechanism
Almost every real task is really several: understand the client (context only you hold), gather precedents (AI with retrieval, then you spot-check sources — Lesson 13), structure options (AI drafts, you judge fit), price it (your call — stakes and accountability), write it (AI's first pass from your brief — Lesson 11), check every claim (you — Lesson 17), send it (you — it's your name on it).
AI legs are bounded and checkable; human legs involve context, judgement and consequences. The handovers are where quality is made or lost: a rich brief downhill, proportionate verification uphill. Teams that thrive with AI aren't the ones with the best prompts — they're the ones with the best-designed baton passes, and a named human at the anchor leg.
choosing which model "grade" runs an AI leg (fast and cheap vs the most capable) is a related skill this lesson's Go deeper opens — match the grade to the leg, not to the whole project. Lesson 21 completes the picture: what the final signature actually means.Work is a sequence of legs; the craft is choosing which runner runs which leg and engineering clean baton passes. AI legs: volume, drafts, transformation, first-pass search. Human legs: framing, judgement calls, stakes decisions, accountability.
- Downhill baton = the brief — objective, context, constraints, examples (Lesson 11). Uphill baton = the check — verification proportionate to stakes (Lesson 17), the tally test where people are affected (Lesson 18).
- Where it breaks: relay legs are fixed and sequential; real workflows loop and branch (agents, Lesson 16, run several legs internally — with the gates you already know to place). Runners also tire; the AI doesn't — which tempts over-delegating judgement legs precisely because the volume legs went so well. Fluent success at leg three is zero evidence about fitness for leg four.
Build a workflow yourself
Conceptual illustration — a deterministic consequences engine, not a live model. Pick a scenario, allocate each step, choose the check for anything AI touches, then run the quarter.
Where this bites at work
Luka's redesigned acquittal workflow: AI drafts sections from his structured brief with the improvement loop capped at two rounds; retrieval grounds every requirement in the funder's current guidelines; he personally verifies every reported figure against programme records; he signs the acquittal, his name to the funder. Six hours becomes two — with quality up, not down.
Spotting a leg worth automating takes four yeses: it recurs frequently, success is describable explicitly, errors are cheap to detect and reverse, and the blast radius of a bad run is bounded. Fewer than four — keep it manual, revisit quarterly.
"Using AI well means writing better prompts" is the misconception to retire. Prompting is one baton on one leg. Decomposition, allocation, gates and verification are the actual skill — a work-design skill, learnable by anyone who runs a process, no code required.
Now make the call
Maya's firm is rolling out a client-facing AI assistant. It's drafting excellent client responses. Someone suggests: "since it's doing so well, let's have it also decide loan pre-approval eligibility." What does Maya recommend?
Try asking the assistant whether it's ready too — worth seeing that self-assessed readiness is still just generation.
Apply it to your work
Pick one recurring task. Write its steps down, mark each AI, human or hybrid, define both batons for every hybrid step, and name the human anchor. One page — worth more than any prompt collection.
Prove it to yourself
Two quick questions — untimed, retryable, and they count toward ★ Mastered (completed + decision + self-check all correct). This is what defeats the feeling of "that made sense" evaporating by next week.
Explain it in your own words
Two or three sentences for a colleague. Then compare against the rubric — it checks ideas, not wording, and only you see your note.
✓ Lesson 20 stamped: Workflow Architect
You allocated a real workflow, ran it, and watched the same model produce wildly different outcomes depending on your design. One lesson remains — the one that makes all the others safe to use at full speed.
Evidence & review — how we know what this lesson claims
| Claim | Type | Basis | Review risk |
|---|---|---|---|
| Task decomposition and allocation, not prompt wording, is the primary driver of AI-assisted work quality | How it works | Human–AI complementarity research | Medium — reviewed quarterly |
| AI legs suit pattern-heavy, checkable work; human legs suit judgement, stakes and accountability | How it works | Series synthesis of Acts 3–4 | Low |
| Model "grade" should match the leg (cost × quality × risk), not the whole project | Practice-based guidance | Provider pricing and capability documentation | High — reviewed quarterly |
| Four-part automation-suitability test (frequency, explicit criteria, cheap-to-detect/reverse errors, bounded blast radius) | Practice-based guidance | Series synthesis | Low–Medium |
| The workflow relay builder | Teaching device | Deterministic consequences engine, labelled illustrative | Medium |
| The relay-race analogy | Useful mental model | Limits stated in the lesson | Low |
Concept review due: January 2027. Product-behaviour claims: quarterly.
Where you are in the machine
Human judgement now includes designing the whole relay, not just one leg. Select any layer for its role, lessons and your live progress.
The machine, bottom to top: training data (2) → token pieces (3) → prediction engine (1, 4–9) → the desk (10–11) → senses (12) → live knowledge (13) → tools (14) → memory (15) → agent loops (16) → human judgement (17–21).
Next: the anchor leg
One lesson remains — the one that makes all the others safe to use at full speed: judgement, verification, and what responsibility means when a machine did the drafting.