Back to Blog
leadershipstrategyai-agentslocal-aibuild-in-public

Your Team Is Building Agents. Is It Automation or Augmentation?

May 24, 20266 min readBy Bruce Canedy

The build that prompted this

Over the past two months I've built three local agents on an M4 Max Mac Studio. route-ops plans tomorrow's mobile-espresso-bar stops. bar-prep reads route-ops' output and produces a prep checklist tailored to each stop. supplies-order reads the same route file and decides what we need to restock. They run in OrbStack containers against a local Ollama model and coordinate through a shared volume. No cloud, no API bill, no message broker, no framework. The seven canonical multi-agent patterns — Sequential, Parallel, Aggregator, Router, Loop, Hierarchical, Network — are all reachable from that primitive.

The series so far has been a builder's notebook. This post is the leadership companion: the question I keep landing on once the docker compose file is checked in.

The HBR frame

A new piece in HBR, Why Companies That Choose AI Augmentation Over Automation May Win in the Long Run, argues that leaders are facing a strategic fork. One path treats AI as a way to do the same work with fewer people. The other treats AI as a way to do work the organization couldn't do before. Both can be defended on a spreadsheet. They produce very different organizations five years out.

The authors' core insight is that the choice is not a private executive deliberation. It is a signal that travels fast. Employees decode AI strategy the moment it touches their workflow. When they perceive that AI is being deployed to them rather than with them, behavior shifts: shallow adoption, declining well-being, more "workslop" (low-effort, low-quality AI-generated output), faster attrition of top talent, and a junior bench that quietly erodes.

The piece grounds this in a survey of 1,294 desk workers across the US, Canada, and UK. Two numbers are worth carrying around:

  1. 65% — the higher rate of self-reported workslop among employees who feel forced to adopt AI versus those who feel encouraged.

  2. 32% — the lower intent-to-leave among employees who perceive their company's AI intent as augmentation versus automation.

Same tools. Same dashboards. Different perceived intent. Different outcomes.

Two J-curves

Erik Brynjolfsson's "Productivity J-curve" describes the lag between adopting a new general-purpose technology and seeing productivity gains: a dip first, then a sharp rise. The HBR authors apply it to AI at the firm level, and argue there are two J-curves to choose between.

The automation curve dips shallow. Substitute labor for AI on well-specified tasks, run the same throughput with smaller teams, harvest cost savings within a few quarters. The peak is real but bounded. You have made the same machine cheaper.

The augmentation curve dips deeper. You're investing in re-skilling, in job redesign, in the human-AI coordination routines that have to be discovered by the people doing the work. The dip is longer because the learning is harder. But the peak is higher, because what comes out the other side is not a cheaper version of yesterday's work. It is work the organization couldn't previously produce.

The choice between curves is not a one-time decision at a board meeting. It is a thousand small decisions about which agent to build, who's in the design conversation, where the saved time is allowed to go, and which roles get protected versus quietly eliminated. Each one of those decisions is a signal.

The six-phase progression

The HBR piece lays out the behavioral progression each path generates after the initial AI rollout. It's worth holding both side by side.

The automation path runs toward decline: early resistance, then layoffs that erode well-being and productivity, leaner teams overburdened as workslop rises, attrition climbing as employees seek stability, employer brand suffering, and leadership pipelines hollowing out.

The augmentation path runs toward growth: trust accelerates adoption, sustained well-being lifts productivity, teams develop new capabilities while workslop stays low, retention strengthens and institutional knowledge broadens, employer brand becomes a magnet, and leadership pipelines deepen.

The phases are the same sequence in both paths. Adoption, well-being, capability, retention, brand, pipeline. The difference is the direction of the slope at each phase, and that direction is set by what employees believe the AI is for.

Three levers a leader actually controls

Most of this is too abstract to act on at the level of next Tuesday's sprint planning. Three concrete levers translate the HBR frame into decisions a leader of a small agent-building team can actually pull.

  1. Pilots or passengers in the design conversation. When your team scopes the next agent, the person whose work it touches needs to be in the room, not in a "stakeholder review" afterward. Co-design is the difference between an agent that absorbs context the human had to learn the hard way, and an agent that produces plausible-looking output the human then has to clean up. The survey's 65% workslop gap between forced and encouraged adoption is the price of skipping this step.

  2. Where the saved time goes. The 2025 Indeed Workforce Insights Report found that the time AI saves is mostly redirected into "more of the same tasks." That is automation by accident. You've installed augmentation tooling and gotten automation outcomes because no one explicitly re-routed the freed hours. Augmentation requires a deliberate answer to what does this person do now that they didn't have time for before? — and that answer needs to land somewhere the person recognizes as their work growing, not their workload growing.

  3. The shape of the junior bench. The cheapest agent to build is the one that replaces an entry-level role. The most expensive thing to lose is the pipeline of people who would have grown into the senior roles ten years out. Research from Harvard and Anthropic both flag the same pattern: generative AI is currently protecting top-of-organization roles and compressing or eliminating junior ones. Short-term efficient. Long-term fragile. The future leaders learn judgment by handling the work that an agent can now do faster than they can. Remove the work and you remove the school.

Credible commitment is the thing

The HBR piece is honest that the augmentation path is harder. The dip is longer. The investment is real. The board may not see the curve turn for years. What makes it work, when it does work, is what the authors call credible commitment: a leader who can point at past decisions and show that the words and the actions line up.

Aon's CEO Greg Case is their go-to example. When he tells 60,000 employees that AI will expand opportunity at the firm, they have direct evidence the pledge is real. During COVID he publicly promised no redundancies and funded it with temporary executive pay cuts. Satya Nadella's reinvention of Microsoft from "know-it-all" to "learn-it-all" is the other case study. Both leaders chose the longer J-curve and made the choice legible.

The legibility is the work. A small team running local agents doesn't have a 60,000-person broadcast channel, but it has something more direct: every weekly demo is a credible-commitment moment. Whose work got better this week? Who's in the design conversation for the next agent? What new thing is the human doing with the time the last agent gave back? Those answers, repeated, are the strategy.

What this means for our build

We are a small team experimenting with private local models on a Mac Studio. That smallness makes the augmentation/automation choice more visible, not less. There is no anonymizing layer of 4,000 employees between a leadership decision and the person whose Tuesday afternoon it changes. Each agent we build is a vote.

So far the agents we've built are augmentation votes. route-ops doesn't fire a dispatcher, because there isn't one. It gives back the planning hours that were never going to scale. bar-prep doesn't replace the barista. It lets the barista walk in and start serving instead of spending the first thirty minutes deciphering the day. supplies-order doesn't eliminate procurement. It keeps procurement from being a 7am scramble.

The architecture decision was Sequential vs Parallel. The leadership decision is pilot vs passenger. The first decision is in the docker compose file. The second is in every conversation around it.

References

This is the leadership companion to my M4 Max series on building a local AI development environment. I'm building in public at Technology Playground. If you're thinking about AI for your team — or just want to argue with my take — I'm at bruce.canedy@technologyplayground.com.

Back to all posts