# Ungoverned AI Pilots Are Welding the Old Organization in Place: The Pilot Trap, Named by McKinsey Canonical URL: https://theunclej.com/blog/hr-pilot-trap Markdown URL: https://theunclej.com/blog/hr-pilot-trap.md Description: Our read: the pilot trap is not "too many pilots." It is pilots substituting for design instead of feeding it — every ungoverned agent welds a fragment of the old operating model into code. Speed at t Category: ai-organization Tags: ai-organization, piece-00000179 Published: 2026-10-21T05:06:47.823Z Updated: 2026-10-21T05:06:48.964Z --- ![og-cover](https://assets.theunclej.com/uploads/9807ef7f67d5330e2b580246d41be13d.png) # Ungoverned AI Pilots Are Welding the Old Organization in Place: The Pilot Trap, Named by McKinsey > **Our read: the pilot trap is not "too many pilots." It is pilots substituting for design instead of feeding it — every ungoverned agent welds a fragment of the old operating model into code. Speed at the task level masks a vacuum at the design level, and by the time you see it clearly, the old organization has come back wearing agents as a disguise. That judgment is McKinsey's (August 2026); what this piece adds is the unpacking, the verbatim quotes, and the audit.** > > Reader's question: are the AI pilots running in your company feeding the future blueprint, or giving the old organization a new lease on life? > > Scope note: every claim traced here comes from a McKinsey article published in August 2026 (Komm/Hancock/Durth); it is the firm's consulting experience generalized, not survey statistics, and "an organization" and "a global auto manufacturer" are McKinsey's anonymized cases. Our own additions — the three-question audit and the "double noose" framing — are signed as ours and kept separate from McKinsey's claims. In August 2026, McKinsey published an article titled "Escaping the pilot trap: Building HR for the agentic era." Its opening line is the question every HR leader now asks: we understand the general destination, but how do we get there? McKinsey's answer starts not with a roadmap but with cold water: the way you are running things right now may be welding shut the very thing you intend to change. ## What the pilot trap actually is Start with the default reading. Twenty years of transformation methodology teach the same sentence: start small, validate with pilots, scale what works. Lean startup, agile transformations, digital pilots — all the same logic. Applied to AI: pick a few use cases, pilot them, roll out the winners. That logic breaks in the agent era. Not because the methodology was wrong, but because its premise is gone. The first shift McKinsey names: a frontline employee can now configure a working agent in an afternoon — a payroll specialist will do, so will a recruiter; low-code tools, no waiting in IT's queue (txt L26-27). McKinsey calls this "the greatest accelerant of the transformation and, at the same time, its greatest governance challenge": local experimentation without guardrails hardens into a fragmented landscape of one-off agents faster than any central team can catalog it (L27-29). The second shift is harsher: **"starting small" no longer requires your approval.** Leadership doesn't sanction it; the front line starts anyway. The question is not whether to start small — it is whether those efforts unfold within a deliberate design or instead of one (L40-43). Hence the definition. In prior transformations, moving without a blueprint cost time; this time, it costs the blueprint itself — **every ungoverned agent hardens a fragment of the old operating model into code** (L43-45). McKinsey's name for it is the pilot trap: **speed at the level of the task, drift at the level of the design** (L45-46). In plain language — and this translation is ours, not McKinsey's wording: a company that governs nothing accumulates well-running AI pilots, some of them optimizing processes that should not exist in five years (McKinsey's qualifier is "in some cases"). We take the mechanism one step further: with no blueprint to check against, the fastest pilots start making the blueprint's decisions — that extension is ours. ## Three judgment calls nobody wants to make McKinsey says the North Star must take a position on three core tensions, and none of the three is a technical question. All are judgment calls (L60-95): **First, who owns the orchestration platform.** The agents belong to HR, but the platform they run on is enterprise infrastructure — identity, security, integration, orchestration. Give it wholly to IT and HR waits in line, losing the pace advantage that makes agents valuable; let HR build its own and the enterprise inherits a shadow stack. The same question repeats inside HR: which agents are built centrally as products, and which are configured locally by the people who own the workflow (L70-74). **Second, investment or surveillance.** Agentic HR can finally tailor its work to the individual rather than the segment: anticipating a career move, curating the next role, tailoring development to how someone actually works. But the same capability, pushed one notch too far or introduced without transparency, stops feeling like an investment and starts feeling like surveillance. McKinsey's test is precise: **the line is not set by regulation; it is set by whether employees believe the personalization serves them** (L82-83). **Third, automating work or accelerating people.** Deciding what agents absorb is the easy call. The hard one is naming which human capabilities become more valuable — judgment, relationships, coaching, persuasion — and where the organization will deliberately accelerate them. Embedded in this is a succession problem: **if agents take the entry-level work, the next generation's judgment has to be developed on purpose rather than absorbed on the job** (L90-91). McKinsey adds a blunt coda: none of these tensions is solved once; they are managed continuously — and **a North Star that does not take an explicit position on all three is not yet a North Star** (L93-95). ## The 2030 accounts: a wholesale migration of time McKinsey sketches 2030 in exhibits one and two. Three accounts deserve to be pinned to the wall: Account one: by 2030, roughly two-thirds of today's HR activities could be fully automated or fully automated in delivery (L184-186). Note that McKinsey insists on the distinction — fully automated means the agent completes the work end to end with no routine human involvement; fully automated in delivery means agents execute the workflow while humans retain ownership of policy, outcomes, and exceptions. Payroll is the canonical case: agents run the cycle, humans design the policy and adjudicate the cross-border complication (L188-194). Account two: the structure of an HR professional's day gets rebuilt. Today a typical function allocates roughly 40 percent of its time to operational delivery, 30 percent to program and project execution, 15 percent to ad hoc problem solving, and 15 percent to strategic engagement and innovation. In the 2030 model, operational delivery shrinks to roughly 5 percent, program execution to roughly 10 percent, ad hoc problem solving stays at 15 percent — **strategic engagement and innovation rise to 50 percent, and a new 20 percent category appears: agentic capability management** (L223-229). That 20 percent is not a rounding category: configuring agents in low-code environments, writing and reviewing the logic they follow, testing before release, monitoring performance and drift, retiring them when the workflow changes. McKinsey's words: work that sits closer to coding and product management than to case handling will become a normal HR career rather than a specialist sideline (L237-251). Account three stings the most: **the three-pillar model is sentenced to reassembly**. The Ulrich model that split HR into business partners, shared-services centers, and centers of excellence — the framework every HR professional in China has memorized — merges into a single layer in McKinsey's destination: shared-services centers become largely digitized and agentified, COEs shrink to a few deep experts organized into tribes, and the business-partner population is revamped into workforce pods, one per business unit, each a senior strategist with one or two analysts (L262-266, L282-284). HR as a whole becomes a product-and-platform organization: two product groups (talent products, employee experience products) plus one platform group (people platform and intelligence), topped by an enterprise workforce architecture group that runs workforce simulations and the skills graph — the architecture, McKinsey says, that makes the CHRO a genuine peer to the CIO on agent strategy rather than a customer of it (L274-291). ## HR's double noose All of the above lands on one role: HR itself. McKinsey's phrasing: HR has a dual mandate, as both the architect of the enterprise's human-agent redesign and a subject of that same redesign (L34-35). In the ERP and agile eras, HR was largely a downstream adopter of changes designed elsewhere; in the agent era, it has to transform itself first. The sentence that follows is merciless: **a function that moves slower than the organization it advises, or that automates tasks while preaching workflow redesign, forfeits the credibility to lead** (L36-37). That is the double noose: demand that the whole company transform while stalling your own, and you lose credibility first and your reason to exist second. McKinsey lists "how do we rebuild HR's own capabilities" as one of the five CHRO decisions (L430-433), alongside "what is our North Star," "where do we enter," "who governs," and "what is the value case" — one of the five decisions is a mirror. ## McKinsey's prescription, and our three questions The prescription fits in one sentence: **North Star back** — define the 2030 human-agent operating model for the function first, then work backward to the implementation sequence, capability investments, governance structures, and quick wins (L57-60). Three entry points (bottom-up use cases, end-to-end process redesign, domain-by-domain reinvention) are all legitimate, but they are entry points, not strategies; proof points must be selected, sequenced, and judged against a clear view of the future operating model (L120, L176-180). McKinsey also supplies a reality check: one organization on the bottom-up path implemented 50 use cases and found significant annual savings and better service levels — but McKinsey immediately notes that without ties to a broader redesign, you may be perfecting steps the future operating model eliminates (L126-141). That is where McKinsey's mechanism account ends — the sections that follow cover roles, transition layers, and decision lists. We add three audit questions — ours, not McKinsey's — to tell whether your pilots are feeding the blueprint or welding the fragments: **One: when did you last kill a pilot?** A pilot list that only grows means governance is really just approval. An organization with a genuine North Star retires agents on a schedule — McKinsey explicitly writes "retiring them when the workflow changes" into the duties of that 20 percent (L249). The right to retire matters as much as the right to launch. **Two: does your pilot list match your 2030 picture?** Take McKinsey's destination org chart as the template and tag every pilot you are running: does this agent exist in the destination model? If yes, which product group owns it? If no, is it optimizing a fragment that should be eliminated? **Three: which phase is HR's own agentification scheduled in?** If HR's transformation plan puts HR's own overhaul in "phase two" or "when conditions mature," that is not a plan; it is avoidance. McKinsey's credibility sentence works as a yardstick: the business will watch how fast HR transforms itself before deciding whether to believe it. You don't need to fully understand the agentic transformation, and you don't need to first answer "which jobs will exist in 2030." Next time your company greenlights a new AI pilot, run the three questions: does this pilot have a position in the destination picture? Which phase is HR's own equivalent overhaul in? When was the last pilot retired, and why? If any answer is missing, your pilots are still making decisions on the blueprint's behalf. McKinsey has laid the mechanism on the table, including the uncomfortable note that fifty use cases can be optimizing processes that deserve to die. For your own pilot list, at least the books don't have to stay murky. ## Sources and scope - Source: "Escaping the pilot trap: Building HR for the agentic era" by Asmus Komm, Bryan Hancock, and Sandra Durth, McKinsey People & Organizational Performance Practice, August 2026. All L line references point to the physical lines of the local extraction escaping-the-pilot-trap-extracted.txt; the official PDF (sha256 `6fe2f2c1…`) and extracted text (sha256 `63a4535d…`) are archived. - Evidence boundary: all claims are McKinsey's consulting generalizations, not survey statistics; "an organization" and "a global auto manufacturer" are McKinsey's anonymized cases and are not named or embellished here; the two-thirds figure, the 40/30/15/15-to-5/10/15/50/20 shift, and the three-pillar merger are McKinsey analysis, attributed as such. - The "three-question pilot audit" and the "double noose" framing are this piece's own reading, signed as ours and kept separate from McKinsey's claims. - Division of labor on this content line: piece 178 covers the hiring end (the broken junior ladder — buying execution versus buying judgment); this piece covers the organization end — whether pilots feed the blueprint or weld the fragments. Neither re-argues the other's frame. — I'm Uncle J. I turned the judgment of "how to read people and deploy people" into an AI system. No parameter talk; just the books. ## Appendix: Eight load-bearing sentences, verbatim The eight sentences below carry the article's weight, quoted verbatim with our Chinese translation. Any translation error is ours; the English original governs. **1. The definition of the pilot trap (txt L45-46)** > We call this the pilot trap: speed at the level of the task, drift at the level of the design. **2. The one-afternoon threshold (txt L26-27)** > A payroll specialist or a recruiter can now configure a working agent in an afternoon using low-code tools. This is the greatest accelerant of the transformation and, at the same time, its greatest governance challenge. **3. This time the blueprint itself is the cost (txt L43-45)** > In prior transformations, moving without a blueprint cost time. In this one, it costs the blueprint itself—every ungoverned agent hardens a fragment of the old operating model into code. **4. HR's dual mandate and the credibility sentence (txt L34-37)** > In the agentic era, HR has a dual mandate as both the architect of the enterprise's human–agent redesign and a subject of that same redesign. A function that moves slower than the organization it advises, or that automates tasks while preaching workflow redesign, forfeits the credibility to lead. **5. The line between investment and surveillance (txt L82-83)** > The line is not set by regulation. It is set by whether employees believe the personalization serves them. **6. The apprenticeship succession problem (txt L90-91)** > If agents take the entry-level work, the next generation's judgment has to be developed on purpose rather than absorbed on the job. **7. No position, no North Star (txt L93-95)** > None of these tensions is solved once; they are managed continuously. A North Star that does not take an explicit position on all three is not yet a North Star. **8. People remain; where does judgment sit (txt L202-204)** > The question across HR is not whether a human is involved, but where human judgment sits in a workflow that agents largely execute.