Automate the robotic work. Augment the rest.

The five-minute email that cost the client.

Picture the email. An account executive has a strong Tuesday call with a fifty-seat prospect. Her follow-up sequence is automated — that was the productivity win of the quarter. Wednesday morning the prospect opens: "Hi {{first_name}}, great speaking on Tuesday." The proposal attached is right. The deal is dead anyway. Saving five minutes cost the client.

The example is fictional; the pattern is anywhere automation has been rolled out with enthusiasm. And the wrong lesson to draw is "AI can't be trusted with email." The right lesson is that this task sat on the wrong side of a line: client-facing, relationship-heavy, unstructured — and it was handed to a machine end to end, no human between draft and send. The same technology, placed one step earlier in the workflow — drafting the follow-up for the rep to check and sign — would have saved four of the five minutes and kept the client.

That line has a name. On one side, automation: the machine owns the task. On the other, augmentation: the machine accelerates the task, a human owns it. Most AI strategy, reduced to its useful core, is deciding which side each piece of work belongs on.

Two questions decide the split.

You do not need a committee for this decision. Two questions place almost any task.

First: how structured and predictable is the work — both input and output? CRM logging, meeting notes, data enrichment, list building: defined inputs, verifiable outputs, the same shape every time. Machines are excellent at work with a shape. Proposals, pricing calls, negotiation prep, the message that revives a stalled deal: open-ended inputs, judgment-shaped outputs. There, the machine's contribution is speed, not the decision.

Second: what does a mistake cost — and who sees it? An error in an internal brief costs minutes. An error in front of a customer costs trust, and trust does not come back at the price it left. The higher the cost of a bad output and the more visible the failure, the stronger the case for a human between the machine and the outcome.

Put the answers together and the rule almost writes itself: structured and low-stakes → automate. Unstructured or high-stakes → augment. The follow-up email from section one fails both questions at once — unstructured relationship work, customer-visible failure — which is why it was the most expensive five minutes of the quarter.

For borderline tasks, one tiebreaker settles most arguments: how cheaply can you verify the output? Automate what you can score at a glance — a booked meeting, a correct field, a logged call. Where checking an output is itself expensive, that is exactly where a human belongs, and where their attention matters most.

The frontier is jagged. Test, don't assume.

The split would be easy if AI's ability were a level line — everything below some difficulty automated, everything above it human. It is not. AI capability is a jagged frontier: two tasks that look equally hard to you can sit on opposite sides of it, and which side is not obvious until you test.

The cleanest evidence comes from a field experiment run at Boston Consulting Group by researchers from Harvard, Wharton, MIT and Warwick, covering 758 BCG consultants. On tasks inside the frontier, consultants using AI finished faster and produced significantly higher-quality work. On a task deliberately designed to sit just outside the frontier — one built to look similar in difficulty — consultants using AI were 19 percentage points less likely to reach the correct answer than colleagues working without it: roughly 60–70% correct, versus 84.5% for the control group. Same calibre of people, same tool, opposite result. The tool didn't change; the task did.

Three working rules follow. Test at the task level, with your own real work — not the vendor's demo tasks. Never extrapolate from one success to the neighbouring task, however similar it looks. And retest on a rhythm, because the frontier moves as models improve: a task that failed the test in January may pass it by summer, and the teams that retest are the ones that notice first.

Augmented does not mean unattended.

Augmentation has a failure mode of its own, and it is sneaky: the better the machine performs, the less attention the human pays. Fabrizio Dell'Acqua — one of the jagged-frontier researchers — showed this in an earlier experiment with professional recruiters and gave it a name: "falling asleep at the wheel." Recruiters working with a higher-quality AI made worse assessments than colleagues given a weaker one, because the better the AI seemed, the less genuinely they checked. The oversight still appears in the process diagram; it just stops happening. The most dangerous output is the one that has been right twenty times in a row.

So design the judgment in, rather than assuming it. And one more distinction earns its keep here — augmentation runs in two directions, and choosing the direction is part of the design.

The goal is humanization.

It is worth saying plainly what the split is for, because "automation" still triggers the wrong reflex — that the point is fewer people. In a sales team the point is the opposite. The robotic share of a rep's week — the logging, the formatting, the assembling, the retyping of things already typed — is work a machine should do precisely so the human hours move to work only humans can do: sitting with customers, reading a room, making the judgment call on price, earning trust.

That is also the honest test of whether your split is right. Don't audit the tech stack; audit the calendar. If the machine hours went up and the customer hours didn't, you automated the wrong things — or automated the right things and spent the freed hours on new admin. The split from section two is not an IT decision. It is a decision about what your people's time is for.

Automate the robotic work. Augment human performance.

1. Dell'Acqua, McFowland, Mollick et al. · "Navigating the Jagged Technological Frontier" — Harvard Business School Working Paper 24-013, with Boston Consulting Group (2023; later in Organization Science, 2025). 758 BCG consultants; quality gains inside the frontier, a 19-percentage-point correctness drop just outside it.
2. Dell'Acqua · "Falling Asleep at the Wheel: Human/AI Collaboration in a Field Experiment on HR Recruiters" (working paper). Higher-quality AI led to lower human accuracy as reviewers disengaged.
3. Harvard Business School Online · AI Essentials for Leaders. The automate/augment/avoid decision, frequency × value, and the AI-in-the-loop vs human-in-the-loop scenarios.