Series: hub | executives | middle managers (this piece) | front line | frontline staff
I | The epicenter: the most pressure, and the most opportunity
Generative AI, intelligent agents, and process automation are pushing Taiwanese companies from “passing things up the chain, layer by layer” toward “deciding on the spot.” This is AI Flattening: once the layers thin out, information and decisions no longer travel up and then down. They are connected directly by the data panels and conversational entrances of frontline POS, customer service, WMS/OMS, and cloud ERP.
The tremor is strongest in the middle. The span of control is being enlarged whether you asked for it or not; the right to approve is being narrowed whether you agreed or not. The number of meetings seems to fall; on-the-spot judgment multiplies.
Many companies have already let AI take over scheduling, reports, checks, and reminders, and the middle manager’s role as “information broker” is leaving the stage quickly. At the same time, the front line needs someone even more to string scattered actions into a path that can actually be walked; the executive layer needs someone even more to translate strategy language into executable boundaries and a beat.
Middle management has not disappeared. It is being forced to upgrade: from the manager who “stays late because they are busy” to the builder who is accountable for the mechanism, the path, and the people.
II | The role remade: from “information broker and sign-off” to “mechanism coach and problem-namer”
The middle manager used to be a switch: aligning, reporting, signing off. Today they are more like a groundskeeper plus a coach. What they maintain is whether the path is clear: at which step new customers get stuck, which two extra questions conversion asks, whether fulfillment swings because the promise was written unclearly or because the resources were not in place.
What they coach is turning “done right once” into “done right every time”: breaking a successful case into a template, playing a failed trajectory back to the team, rewriting “experience that lives in a feeling” as “a rule anyone can use.” After AI has automated the visible processes, the distinctive value of the middle concentrates in the mechanism that is invisible and yet shapes results every day: how a problem is named, how disagreement is tightened, how a promise is written clearly, how an exception is stopped.
III | Value anchors: hold on to three things; the rest can be outsourced
First, the right to name the problem: define the problem in task language, not department language. Replace “which of our KPIs was missed” with “which thing the user needs to finish,” and resources will flow to the right path.
Second, the ability to close the path: string “acquisition → conversion → fulfillment → repurchase → word of mouth” into the shortest executable path, and for each node set entry conditions, a default solution, and an exception fallback.
Third, the talent-growth track: turn “I do this well” into “the team does this steadily,” so that a newcomer’s growth curve is steeper and shorter, and a veteran’s experience can be reused. Hold these three, and the team will have inertia; only then will results be stable.
IV | Daily work rearranged by AI: from “managing the work” to “managing the mechanism, the people, and the exceptions”
Once reports, scheduling, and alerts are automatic, the timetable has to be rewritten. A workable allocation is 7-2-1: seventy percent on optimizing the mechanism and the path, twenty percent on coaching and alignment, ten percent reserved for exceptions on the floor.
Optimizing the mechanism and the path means that every extra question, every rollback, and every failure can be located and repaired. Coaching means transplanting a problem-solving frame from your head into the team’s. Exception-handling means that, in the gray zone, “there is someone to ask, and someone who dares to call it.”
Once the beat is stable, “fighting fires” is replaced by “building.”
V | A shared factual base: pull the argument back from “who is right” to “what the facts are”
Collaboration usually stalls not because of attitude, but because the facts are not the same. Middle managers have to lead in building and maintaining a shared factual base: a unified definition for transactions, conversion, fulfillment, service, word of mouth, and cost; a unified timeline for key nodes; a unified record of experiments and versions.
Any discussion first returns to the base; any conclusion is written back into the base. For every critical path, define a minimum necessary data set (follow-up steps, first-pass rate, promise-fulfillment rate, share actually used), then add average lag and peak stability.
When everyone is looking at the same piece of reality, the wrangling falls away, and decisions can be replayed.
VI | Path operations on the ground: from “campaign sprints” to “a path that runs steadily”
Campaigns matter, but they should serve the optimization of the path. Take e-commerce or delivery as an example, and walk “a new customer’s first completed order” from the start: is the entrance search, or a conversational recommendation? Can the default combination solve the problem at a glance? Have the detours of bundling and price-comparison been folded away? Is the promise written up front, or hidden at the bottom?
Is inventory and lead time stable? Is the fallback after a failure clear? Break this path into five measurable nodes, and at each one cut friction, shorten the steps, and raise the first-pass rate.
The repurchase path is the same: take “share actually used” as the intermediate goal, and make “used well → comes back” a design objective, not a slogan. The core is not “do more,” but one step fewer. One step fewer, and it runs more smoothly; more smoothly, and it is more stable.
VII | Exception management: from “putting out the fire after it breaks out” to “slowing down at the boundary”
Sort exceptions into four types: data (inconsistent definitions, missing or drifted values), rules (a conflict between the default solution and a constraint), promises (unclear time-of-arrival / substitution / compensation), and resources (not enough inventory, service capacity, or delivery capacity). For each type, define a trigger threshold — a downgrade action — a rollback condition — and a deadline for the after-action review.
For example, when promise exceptions exceed the threshold, the system automatically switches to a “more stable, slightly slower” option, and at the same time triggers an alert and compensation; within 24 hours, complete the after-action review, and rewrite the unclear promise as a sentence that can be quoted.
In a Taiwan setting, sudden events such as typhoons or earthquakes also need a one-press downgrade and a compliant notification process (disclosure of personal data, notice that operations have stopped), so that the spread of risk is stopped at the boundary.
VIII | Authority and boundaries: write “minimum closed-loop accountability” clearly
Once the layers thin out, authority often cannot be granted because the boundaries are unclear. Write clearly who is accountable for the composite metrics of which path, who has coordination or veto rights at which key nodes, and attach a trajectory that can be replayed before, during, and after the fact.
Authority downward should come with a threshold for “when you may act first, and when you must ask first.” Accountability upward means proactively submitting a short report of “facts — judgment — options — recommendation.” Once the boundary is clear, authority will not be a dab; once authority is in place, the team can grow real judgment.
IX | Talent and the team: turn “an expert’s intuition” into “steadiness for most people”
AI lets some specialties overflow into most people’s hands, and it also devalues “rough experience.” The middle manager’s talent strategy is to structure the tacit knowledge of the people who are good at it: settle useful scripts, stable default combinations, and common objections and follow-up questions into templates and a glossary; turn learnable after-action fragments into material that can be replayed.
Turn “my way of doing it” into “our way of doing it.” Roles also have to be rearranged around a human–AI ensemble: content roles lean toward defining intent and quality control; operations roles lean toward exception and promise management; product roles lean toward path design and data narrative.
Advance platform capability and the remaking of roles at the same time, and the team’s average will be lifted as a whole.
X | Performance and after-action review: from outcome KPIs to “mechanism health”
If you only look at outcomes, it is easy to “add budget” rather than “repair the mechanism.” Put mechanism health on the weekly board: are follow-up steps continuing to fall? Is first-pass rate rising steadily, without swinging by time of day? Can promise-fulfillment rate hold at the peak? Does the share actually used rise as the default combination is improved?
Then add three meta-metrics: strategy half-life (how soon this path’s strategy needs a major rewrite), decision lag (the time from discovery to taking effect), and correction rate (after the review, was the right practice written into the system and reused). If health is poor, repair the mechanism first; only after health improves do you talk about scaling. The order cannot be reversed.
XI | Cross-department collaboration: align the worldview with a shared language
Beyond the factual base, there has to be a shared language. Turn the naming of the outside world and the definition of internal actions into a glossary that can be quoted: what “actually used,” “promise met,” “default combination,” and “downgrade action” mean. In meetings, cut the abstract adjectives; use words that can be verified.
When collaborating, first ask “are we looking at the same set of facts,” and only then “where do our mechanisms differ.” Once the semantics are aligned and the facts are consistent, collaboration moves from persuading each other to combining mechanisms.
XII | The Taiwan context: find a “minimum viable path” inside compliance, security, and ecosystem collaboration
Taiwanese companies have to balance compliance, security, cost, and speed under frameworks such as the Personal Data Protection Act, the Labor Standards Act (employee monitoring), the NCC (communications and data), and the FSC (financial cloud and model use). Data classification, permissions, and audit have to be written into the process, not only into a document. Tool selection should prefer what can be replaced, extracted, and audited, so that business logic is not locked into a single vendor or a proprietary format.
The suggested order of advance is to start from the smallest scheme that can actually run: first get the path working in one or two business units or a first-tier city — the factual base can be stood up, the default combination is actually used, promises can be fulfilled, exceptions can be replayed — and only then copy it across. Multi-cloud plus localization, replaceable models, and migratable rules will earn positive feedback faster than “building the whole stack yourselves.”
XIII | A personal working method: how a middle manager’s day can be “steady, then fast”
Three questions at the start of the day: what is the goal of this path today? What is the most fragile assumption? Which three small experiments will test it? The first meeting of the morning is devoted to “path alignment”: around facts and actions, turn disagreement into tasks plus thresholds, and assign owners and times on the spot.
A quiet period before noon: review yesterday’s log and the replay of key exceptions, and write down two reusable practices. In the afternoon, two fifteen-minute 1:1s: talk through the logic of judgment and the state of energy, not only the numbers.
In the evening, maintain the beat: check whether an inserted order is breaking the rhythm, whether a metric is worsening with no one calling a stop, whether a downgrade or rollback is needed. Close with three lines: which mechanism was repaired today? Which template was settled? Whose capability was seen, and enlarged?
XIV | Energy management: treat yourself as a “critical mechanism” too
The drain on middle managers often comes from being pressed from above and below at high density. Set two guardrails for yourself: first, a quiet period that cannot be interrupted (for thinking and replay); second, an escalation threshold that cannot be skipped (so that every large and small matter does not rush straight to you).
Treat yourself as a critical mechanism in the system: it needs regular maintenance, a clear interface, and a visible boundary. Leadership is a marathon; recovery is scarcer than the ability to sprint.
Closing: middle management does not disappear; it becomes scarce
AI has taken away the work that can be computed, and it has also lit up the real value of the middle more brightly. When you can name a problem in task language, thread collaboration together with a path, use facts and a shared language to cut the wrangling, drive after-action review with mechanism health, and enlarge the team’s average with templates, you have completed the jump from “the manager who stays late because they are busy” to “the builder who lets the organization keep getting better.”
The layers will keep getting thinner, but this kind of middle manager will become more and more scarce — any organization that wants to run steadily, and for a long time, needs this force at the epicenter.

David Yang (Yang Hung) is the founder of Future Insights Lab. He writes AI Future Reflections, with a focus on Agentic Commerce, AIGC, data methods, and organizational renewal.
More deep-dive pieces
- ChatGPT will now show brands how often they are mentioned. That is not proof they were recommended well.
- AI will not only eliminate people: the functions being enhanced, and a manager’s handbook for this week
- AI unemployment ranking: the high-risk functions most likely to be eliminated or largely replaced
- Will AI cause unemployment? Start with the enhance / accelerate / eliminate axes
The judgment, restated
My judgment is that middle managers will not disappear, but “the information broker” will lose its job. The people who remain will orchestrate, align, and be the decision nodes that can be replayed.
Three things this week
- List three information chains that still depend on you to pass things up and down, and turn them into dashboards visible on the spot.
- Write a set of orchestration rules: who may decide on the spot, and when an issue must be escalated.
- Run one cross-layer replay: the basis for the decision, how exceptions were handled, and the next round of corrections.
If you are stuck at the epicenter, you are welcome to subscribe to Future Insights Lab, or reply to discuss the middle-manager orchestration bottlenecks; you can also go back to the hub and the executives piece.
