{"slug": "your-best-ai-pilots-are-cementing-the-process-you-meant-to-kill", "title": "Your Best AI Pilots Are Cementing the Process You Meant to Kill", "summary": "McKinsey & Company's August 4, 2026 report warns that successful AI agent pilots are inadvertently cementing outdated processes into code, creating a 'pilot trap' where working agents become load-bearing and harden inefficient workflows. The report notes that roughly two-thirds of today's HR activities could be fully automated by 2030, and that in the destination state, about 20 percent of human time would be spent on 'agentic capability management'—configuring, testing, and monitoring agents. The authors argue that the instinct to 'start small and learn' is no longer conservative, as any specialist can build a working agent in an afternoon using low-code tools.", "body_md": "###\n[\nOpinion\n](https://www.unite.ai/series/opinion/)\n\n# Your Best AI Pilots Are Cementing the Process You Meant to Kill\n\n[Add Unite.AI to your preferred sources on Google](https://www.google.com/preferences/source?q=unite.ai)\n\nMcKinsey published a piece on August 4, 2026 about building an HR function around agents, and buried in the setup is the most useful sentence anyone has written about enterprise AI this quarter. Describing what happens when teams build agents without a plan, the authors write that [“every ungoverned agent hardens a fragment of the old operating model into code.”](https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/escaping-the-pilot-trap-building-hr-for-the-agentic-era)\n\nThey call the result the pilot trap: speed at the level of the task, drift at the level of the design.\n\nThat sentence is worth more than the framework it introduces, and it applies far outside HR, to a lot of companies running agent pilots right now who can’t see the outcome, because the pilots are working.\n\n## The trap is the pilot that succeeds\n\nHere’s the mechanism. A payroll specialist or a recruiter, the article notes, can now configure a working agent in an afternoon using low-code tools, the build-by-configuring platforms that put most of the work in clicking and writing instructions rather than writing code. Not a prototype. A working one, wired into a real workflow, producing real output.\n\nWhich means the sequence everyone was taught, start small then learn then expand, is no longer the cautious option. McKinsey puts it flatly: the instinct to “start small and learn” is no longer conservative. When any specialist can stand up an agent in an afternoon, starting small happens whether leadership sanctions it or not.\n\nWorth being precise about where the article stops and the extrapolation starts. McKinsey’s pilot trap is mainly an argument about wasted effort — teams perfecting steps that a proper redesign would delete. The version below is the harder one, and it’s mine: the wasted effort is recoverable, and the lock-in isn’t.\n\nThe failed pilots are fine. A pilot that fails gets switched off and forgotten and costs a few weeks. The dangerous ones are the ones that work, because a working agent immediately becomes load-bearing. Somebody’s Tuesday depends on it. It gets defended in the budget meeting. And whatever process it automated, including all the steps that only exist because of a system you replaced four years ago or a policy nobody has read since, is now encoded, running, and much harder to argue with than it was when it lived in a document.\n\nThat’s the trap. You didn’t automate a workflow. You ratified one. Which is the uncomfortable version of the increasingly common finding that [AI results depend more on the shape of the workflow than on the tools pointed at it](https://www.unite.ai/purpose-built-ai-financial-institutions/). If the workflow is the variable that matters, then hardening a bad one is the most expensive thing a successful pilot can do.\n\n## The number they buried\n\nThe headline figure in the piece is that roughly two-thirds of today’s HR activities could be fully automated or fully automated in delivery by 2030. That number will get quoted everywhere this week and it is the least interesting thing in the article.\n\nThe interesting number is in the time-allocation exhibit. In the destination state McKinsey describes, roughly 20 percent of human time goes to what they call agentic capability management: configuring agents, writing and reviewing the logic those agents follow, testing them before release, monitoring their performance and drift, and retiring them when the workflow changes. The article doesn’t define drift; read it as the slow slide where an agent’s output stays plausible while quietly stopping being right.\n\nA fifth of a function’s human hours spent tending the machines. The article is explicit that this is “not a rounding category,” and it’s right. It’s the same [supervision overhead that keeps turning up as the biggest hidden cost in enterprise AI](https://www.unite.ai/hidden-cost-of-enterprise-ai-botsitting/), now given a percentage. But the piece doesn’t say whether that 20 percent is a transition cost that falls as the practice matures, or the actual steady-state price of running agents at scale. That’s the number that decides whether an agent program is a cost reduction or a cost swap, and it’s missing. If you are modeling savings from automation and you have not put a line in the model for the people who maintain the automation, your model is wrong by roughly a fifth of the affected team.\n\nThe same gap shows up in the case detail. One organization in the piece implemented 50 use cases and “identified significant annual cost savings along with meaningful improvements in HR service levels.” Significant is doing a lot of work there. More to the point: the article warns, in the paragraph right after, that in some cases organizations improve the efficiency of processes that end-to-end redesign later eliminates or absorbs into entirely different workflows. So how many of those 50 survived? That single number would settle whether the bottom-up route is a shortcut or an expensive detour, and it isn’t there.\n\n## The half of the advice a normal company can actually use\n\nMcKinsey’s prescription is to define your 2030 human–agent operating model first and work backward from it. Define the destination, then sequence the implementation, the capability investments, the governance.\n\nThe article is honest about who can do that. Its own destination org chart, it says, requires “product management muscle that most HR functions do not have today, a data foundation strong enough to run a skills graph, and a CHRO with a real enterprise workforce design mandate.” Then: “fewer organizations can start here.”\n\nFor everyone else, including plenty of companies that already have agents running, the North Star is not the actionable half. The diagnosis is. If ungoverned agents encode the old process, then the exposure you have today is proportional to how many agents already exist that nobody approved, and you almost certainly do not know that number.\n\nSo the first move isn’t a 2030 blueprint. It’s a count.\n\nFind out what’s actually running. Not what was approved. What exists. Every automation someone built in a low-code tool, every scheduled job wired into a shared inbox, every assistant configured against a real system of record. For each one, three facts: who built it, what it can touch, and what breaks if it stops. That inventory takes an afternoon per department and it is the only version of “governance” a company without a transformation office can execute on Monday.\n\nThe article’s own fifth CHRO question is the one to steal, and it’s the sharpest of the five: which early choices are foundational, which compound over time, and which can be safely undone. Sort your inventory that way. The agents that can be safely undone are your experiments and you should run more of them. The ones that can’t are your architecture, whether or not anyone decided that. It’s also a useful correction to the instinct that drives most first pilots, which is to point agents at [the most visible process rather than a quieter internal one](https://www.unite.ai/the-worst-first-job-you-can-give-an-agent-is-the-visible-one/).\n\n## Who this actually costs\n\nThe party that loses here isn’t the company that never started. It’s the one two years in with a healthy pilot count and a slide showing adoption climbing.\n\nPilot count is the metric that feels like progress and measures the opposite. Every additional pilot built against the current process raises the cost of changing that process later, because now there’s tooling in the way and a person who owns it and a number in a deck that says it’s working. The company with dozens of working agents is not ahead of the company with three. It is more committed.\n\nThat’s the reframe worth taking out of this piece. The question to ask about your agent program isn’t how many are running or how accurate they are. It’s how much of your current operating model you would have to unwind to change your mind — and whether anyone could tell you that today.", "url": "https://wpnews.pro/news/your-best-ai-pilots-are-cementing-the-process-you-meant-to-kill", "canonical_source": "https://www.unite.ai/your-best-ai-pilots-are-cementing-the-process-you-meant-to-kill/", "published_at": "2026-08-04 14:50:03+00:00", "updated_at": "2026-08-04 15:51:31.182310+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-policy", "ai-ethics"], "entities": ["McKinsey & Company", "Unite.AI"], "alternates": {"html": "https://wpnews.pro/news/your-best-ai-pilots-are-cementing-the-process-you-meant-to-kill", "markdown": "https://wpnews.pro/news/your-best-ai-pilots-are-cementing-the-process-you-meant-to-kill.md", "text": "https://wpnews.pro/news/your-best-ai-pilots-are-cementing-the-process-you-meant-to-kill.txt", "jsonld": "https://wpnews.pro/news/your-best-ai-pilots-are-cementing-the-process-you-meant-to-kill.jsonld"}}