Deciding What to Automate: The Human Stakes Model
Four tiers, one rule, and where the human line falls
One question decides the whole portfolio
Most AI programs open with the wrong one, and the portfolio never recovers.
Return to Marcus for a moment. In Chapter 1 he was sitting in a hospital corridor while his father was in surgery, trying to file the time off that would let him stay. The request itself was trivial. What was not trivial was everything around it: the fear, the paperwork he could not focus on, the quiet worry about how it would look to his manager.
Ask "can this be automated" and the answer is obviously yes. A time-off request is structured, rule-bound, and instant. But that question misses the person in the corridor entirely. The better question is not whether the work can be automated. It is how much the work touches a person's dignity, standing, or livelihood, and therefore how visible a human has to be while it happens.
That is the whole of it. The more a decision touches a person, the more visible the human has to be. We call this the Human Stakes Model, and the boundary it draws we call the human line. Everything below the line can move fast and quiet. Everything above it needs a name, a face, and someone accountable.
Built for the engineer, felt by the employee
Sequencing by what is easy to ship produces a portfolio blind to what it feels like to receive.
Here is the design mistake the model corrects. Most AI programs sequence their work by what is easiest to build. Clean APIs, structured data, low legal exposure. That instinct is understandable, and it is wrong. It produces a portfolio optimized for engineering convenience and blind to what the work feels like on the receiving end. You end up automating the layoff notification because the data was tidy, and leaving the password reset to a ticket queue because integration was annoying.
Sort the work a different way. Not by difficulty, but by what is at stake for the person on the other side. Four tiers fall out cleanly, and each one carries its own rule about who leads.
Time off, expenses, pay questions, benefits, policy lookup.
Onboarding, learning, goal setting, career conversations.
Promotion, performance ratings, pay decisions, internal moves.
Investigations, complaints, layoffs, health, bereavement, exits.
Notice what the tiers do not say. They do not say the agent disappears at the top. Even in a layoff, the agent can assemble the record, check the entitlements, and draft the paperwork. What changes as you move up is not how much the agent does. It is who the employee sees, and who answers for the outcome.
Where the human line falls
Plot the real work by what is at stake and who leads, and the boundary draws itself.
The tiers are a scale. The grid is the map. Take the interactions HR actually runs and place each one on two axes: how much is at stake for the person, and who should lead. A pattern appears immediately. Low-stakes work clusters bottom-left, where the agent can lead alone. High-stakes work climbs to the top-right, where a person has to be in front. The diagonal between them is the human line.
The map exposes the mistakes an easiest-first program makes. A layoff and a policy lookup are equally automatable. They sit at opposite corners of what matters. The grid does not tell you what you can build. It tells you what you are allowed to hand over.
Automation should be governed by what is at stake for the person, not by what is convenient for the process.
Print that on the wall of the room where you decide what to build next. It is the single sentence that keeps a well-run program from quietly becoming a cruel one.
How to place a decision you have never seen
You will not memorize a grid. You can carry four questions into any planning meeting.
When someone proposes a new agent, do not argue about feasibility. Ask about the person on the other side. Four questions, asked in order, place almost any interaction on the map without a debate.
The fourth question is the tie-breaker you apply to any of these: would the person want to ask a human why? If the honest answer is yes, an agent alone is not enough, no matter how clean the data is. People forgive a slow answer. They do not forgive being processed.
The hard cases are where the model earns its keep
The model is easy at the corners. It earns its keep in the messy middle.
Clean examples prove nothing. Three hard ones show where the human line actually sits.
- A declined promotion. The agent assembled a flawless case, and the outcome was no. This looks consequential, and the decision is. But the moment of telling is personal. A person delivers a declined promotion, in their own words, and stays to answer the next question. The agent never sends that message.
- A wellbeing check-in that surfaces a safeguarding concern. It began as a developmental, agent-supported conversation. Mid-way, the employee discloses something that signals risk to themselves or others. The stakes jump tiers in a single sentence. The design that matters is not the check-in. It is what happens the instant a concern appears.
- A resignation that should have triggered a retention conversation weeks earlier. By the time the exit form arrives, the human moment has already been missed. The failure was not the agent processing the resignation. It was the absence of a human prompt when the early signals appeared. The line is not only about the final act. It is about when a person should have entered.
Two of those three cases are the same problem: stakes rise mid-interaction, and the handoff to a human has to be instant and clean. That is a design discipline, not a hope.
Not a raw transcript. The person needs the summary, the trigger that raised the stakes, what the agent already did, what it deliberately did not do, and the one thing the employee is waiting to hear. Handoff without context is just a longer wait.
Draw your line this month
You do not need the full portfolio to start. You need the boundary agreed and written down.
- List your top 20 HR interactions and place each on the grid: what is at stake, who should lead. Do it as a group, out loud.
- Mark the human line. Agree which interactions sit above it. Write the list down and get your HR leadership to sign it.
- Pick one routine interaction below the line to hand fully to an agent, and one consequential one where the agent prepares and a person still owns the outcome.
- Design one escalation path end to end: the trigger, the context that transfers, and what the named human sees on arrival.
- Drawing the line once and walking away. Stakes move mid-interaction. A check-in becomes a safeguarding case. If the line is a static list and not a live trigger, the model fails exactly when it matters most.
- Treating human-led as agent-absent. The most consequential work still needs the agent doing the preparation. Pull it out entirely and you have not protected the employee. You have just made the human slower and less prepared.