# Why the HR Three-Pillar Model Breaks Canonical URL: https://theunclej.com/blog/human-in-the-loop-03-why-hr-three-pillar-breaks Markdown URL: https://theunclej.com/blog/human-in-the-loop-03-why-hr-three-pillar-breaks.md Description: Human in the Loop, Part 3. The three pillars are not getting kicked over by AI. The old interface stopped carrying current. What HRBP actually sells is judgment translation — and that half is the first to get discounted. Category: AI Organization Tags: AI, Organization Design, Human in the Loop, Enterprise AI Published: 2026-05-22 --- # Why the HR Three-Pillar Model Breaks ![Human in the Loop | Why the HR Three-Pillar Model Breaks](/imgs/posts/human-in-the-loop-03-why-hr-three-pillar-breaks-en/01-cover.webp) ![V03 HR Three-Pillar Interface Fracture Map](/imgs/posts/human-in-the-loop-03-why-hr-three-pillar-breaks-en/02-figure.webp) > This is Part 3 of the _Human in the Loop_ series. ## I sat in that HRBP middle pillar I sat in that HRBP middle pillar. Not as an observer. I sat in that seat and worked through dirty information every day. The boss talks strategy. The frontline hears tasks. The frontline talks blockers. The boss hears excuses. The HRBP sits in the middle. On the slide, the job is called "serving the business." On the ground, the job is harder than that. You translate the boss's intent into role-level moves. You translate the frontline's complaints into organizational problems. You translate human emotion into management signal. Then you compress all of that into one piece of judgment the person in the corner office can read, sign, and stand behind. That is why I never liked the phrase "serving the business." Too thin. What an HRBP actually sells, what pays the bills, is not the smile, not the communication, not the deck. It is judgment translation. Not a pretty phrase. But honest. It asks you to know what the boss is actually afraid of, and where the frontline is actually stuck. To hear the heat inside the business, and to know which heat is emotion and which heat is a sign that role design, workflow design, or incentive design has cracked. The moments I felt this most were rarely the formal reviews. It was after the meeting. The boss just delivered a clean paragraph. Business owner nodded. Employee nodded. Meeting adjourned. Out in the hallway, someone pulls you aside and says, almost whispering: "Are we really doing this?" That sentence is worth more than the meeting minutes. Because it tells you there is a piece of judgment in this organization that has not been spoken out loud. The boss thinks the order has been delivered. The middle thinks they already understand. The frontline is still waiting for a more specific answer. The HRBP's value sits in that crack. Not relaying the message. Knowing whether the crack is an information crack, a responsibility crack, a capability crack, or just the boss never said it clearly. Why this used to be worth money: the material was scattered. A piece in the meeting minutes. A piece in the emails. A piece in the interviews. A piece in the performance records. A piece in the project retrospectives. And a much bigger pile that lives nowhere in any system — in the hallway, at lunch, in those two minutes after the meeting. The HRBP's value was loading those fragments into a head, kneading them, filtering them, and turning them into one sentence the boss could actually hear: This is not a people problem. This is an organization problem. AI walks in, and the trouble starts. Not because AI suddenly replaces the HRBP. It is never that simple. The real trouble is that AI does not need to replace the HRBP outright. It only needs to discount the most common half of the HRBP's work. Collecting. Sorting. Relaying. Summarizing. Drafting. These motions used to compound over years of time, relationships, and writing chops. Now a batch of 1-on-1 notes can be clustered. A pile of emails can be summarized. Dozens of project retrospectives can be pattern-mined. Feedback from multiple departments can be compressed into a conflict map. What an HRBP used to take a month to assemble in judgment material can now hit the table in a few hours. Material on the table. Material is not judgment. That is what this piece is here to take apart. AI is not kicking the HR three-pillar model over. It is more like turning on the lights. The lights come on, and people suddenly see: a lot of what companies call HRBP was never an organizational diagnostician. It was a high-end relay clerk. Not a pleasant sentence. The person in the corner office should still hear it. Because what gets rewritten next is not the org chart of the HR department. It is the judgment chain inside the company. Who picks up the signal? Who runs the diagnosis? Who prices the conclusion? Who signs? Who holds the bag? Until those questions get redesigned, keeping or killing the three pillars, renaming them, restructuring them — none of it changes anything except the business card. ## Ulrich is not wrong. The interface is just old. To talk about the three pillars, you have to start by laying the tone flat. Ulrich is not wrong. No reason to dunk on him either. In 1997, Dave Ulrich published _Human Resource Champions_. The weight of that book was not in coining HR jargon. It was in pushing HR one step away from transactions and toward value creation. The three-pillar shape everyone is familiar with grew out of that language: Shared Services for standardized service. Center of Expertise for policy and specialist solutions. Business Partner close to the field, carrying organizational judgment back to the business. Standing in 1997, this design was not crude. It was advanced. Because large organizations did have a real problem at the time. If HR stays drowning in attendance, payroll, onboarding, transfers, and ticket queues, you cannot talk about being a business partner. Concentrate the expertise. Standardize the service. Put people next to the field. That division of labor solved an actual problem in its actual era. But it ran on one default assumption. Judgment in an organization moves because humans carry it. Policy gets written by people. Workflows get walked by people. Business signal gets relayed back by people. Organizational judgment gets compiled by people. Accountability sits with people. The three pillars are not really three columns. They are an interface for how humans move judgment around inside an organization. That is where the problem is. In 2026, the interface is still in place. The current has changed. Standardized service no longer has to be carried by a shared services team. AI workflows can handle large volumes of high-frequency requests directly. Specialist policy no longer just sits in a Word file inside a Center of Expertise. It gets written into prompts, rule files, system permissions, agent boundaries, and audit conditions. Business partners are no longer the only path through which judgment moves. AI is already participating in the first-pass clustering of meeting notes, interviews, performance text, role materials, and project retros. So this piece is not a critique of Ulrich. It is not. What this piece is actually saying is this: What Ulrich solved was the collaboration interface for humans. Once AI enters, work gets split into tasks, judgment, accountability, oversight, and rollback. The original columns do not fall over immediately. But the load-bearing logic underneath each column has already been swapped. A lot of companies are still doing three-pillar reflection from inside the HR bubble: HRBP is too soft. COE is too far. SSC is too mechanical. None of those critiques are wrong. They are critiquing "the three pillars are not running well," not "the premise of the three pillars has changed." These are two different problems. Not running well, you can fix with process changes, owner changes, reporting line changes. Premise changed, you have to re-ask: where does judgment originate, who handles it first, who reviews it, who is accountable for the result, how does it get rolled back when it goes wrong. Ulrich's later work did not freeze in 1997. The line kept moving outward — outside-in, value creation, organizational capability. So the public framing has to stay disciplined. It is not "Ulrich is outdated." It is "a lot of companies are still running a simplified three-pillar model, while AI has already entered the flow of judgment itself." That is the point of citing back. Respect the root system. Do not kneel under it. ## SSC did not disappear. It got repriced. Inside the three pillars, the first one AI bites is SSC. But do not hear that the old way. A lot of people hear "SSC gets hit by AI" and immediately ask: Is the Shared Services Center going away? Not necessarily. I would actually argue SSC is the pillar most likely to get repriced — and the one most likely to find its new seat first. Simple reason. SSC was designed to be the concentrated processing layer for standardized, workflow-driven, high-frequency transactions. The reason it exists is not to preserve manual motion. It is to deliver repeatable, stable, rule-shaped HR service at scale. When AI walks in, what gets taken away is not SSC's value. What gets taken away is the pile of low-value motions SSC used to have no choice but to do with humans. IBM AskHR is a useful reference. In IBM's published AskHR material, the system covers 80 countries and multiple business units. In 2024 it handled 2.1 million employee conversations, with roughly 94% of employee inquiries automated. IBM Consulting's digital delivery team also restructured more than 10,000 HR processes into 70+ workflows, improving service delivery efficiency by 40%. Do not read that as "the chatbot is impressive." That misses it. What it actually says is this: the underlying task shape of SSC has changed. A Shared Services Center used to mean: Someone picks up. Someone looks it up. Someone processes it. Someone replies. That same capability is now compressed into knowledge bases, intent recognition, workflow orchestration, permission calls, exception escalation, and service monitoring. SSC is no longer just shared service. It is starting to become the AI workflow orchestration layer. That shift is far bigger than "a few HR ops headcount got cut." Media reports have circulated numbers about IBM HR roles being replaced. Public writing should not promote those to the level of an IBM official statement. Even without that figure, IBM's own published language is enough to say one thing: High-frequency HR service is moving from human ticket center to automated workflow center. This changes what the person in the corner office should be looking at. The old SSC lens was response time, satisfaction, process compliance, cost control. The AI-era SSC lens is a different set of questions: Which requests can complete on their own? Which requests must escalate to a human? When the knowledge base is wrong, who fixes it? At what quality drop do we trigger rollback? None of those are the craft of a traditional SSC lead. They look more like product operations, workflow governance, and responsibility-chain design. So the danger for SSC is not that AI eats it. The real danger is companies buying AskHR-style products as "service bots" through a procurement process. That is when it dies. Because what you bought is still a tool. Only when the person in the corner office treats it as an AI workflow orchestration layer — and redefines automated handling, human escalation, accountability, and quality thresholds — does SSC actually finish its AI-era repricing. Klarna's customer-service rebound is a reminder for this layer. High-frequency service automating does not mean complex problems can run without a human exit. AI can carry volume. The organization still has to design the escalation path. The new work of SSC is not guarding every ticket. It is designing which tickets no longer need a human guard, and which tickets must put a human back into the loop. Once that is clear, SSC is not the first pillar to die. It is the first one to get a new price tag. ## COE is not writing faster. It is writing executable. The change in the COE pillar is more hidden than SSC. The first reaction is usually: LLMs can draft policy now. They can benchmark systems. They can give me a new performance scheme. So part of COE's expert work gets replaced. Sure. That will happen. But that is the skin. What actually gets rewritten in COE is not "who drafts the policy." It is "what form policy takes." COE output used to be policies, process documents, templates, handbooks, training material. Stored on a shared drive. Stored in the OA. Stored in SharePoint. Stored in a knowledge base. By default, policy was a document. Once AI enters, the document stops being the final form of policy. Policy starts entering prompts, permissions, validation rules, system actions, and agent boundary conditions. Employees no longer read the policy and then act. Employees get constrained by rules inside the workflow, hinted inside the tool, escalated inside exceptions, traced inside the audit. That is the deeper COE shift: Policy moves from Word document to executable rule. Take travel policy. The old version: COE writes a policy. Whether the employee reads it, understands it, or whether the approver is loose — most of it gets caught at the back end, if at all. With AI in the loop, travel policy can become form-fill hints, anomaly amount blocks, auto-explained rules, and human-escalation conditions. Policy is no longer just a "norm." It is a judgment node inside the workflow. Take performance policy. The old version: COE defines the cycle, scoring rules, calibration. With AI in the loop, meeting notes, project retrospectives, goal changes, and cross-team feedback can all be folded into performance material. The question for COE is no longer: How do we write the performance policy? It is: Which material can AI summarize? Which judgments must not enter the evaluation automatically? Does the employee have the right to appeal? Who reviews an AI-generated performance summary? At this point COE's expertise matters more, not less. It just no longer expresses itself as a pretty handbook. It expresses itself as splitting policy into boundaries, rules, permissions, exceptions, and rollbacks. I feel this firsthand inside Meta_J. Organizational governance used to be writing a manual. Now it is closer to writing rule files, prompts, hooks, skills, and review gates. This is not making management more esoteric. The opposite. It pulls judgment that used to live in experience and turns it into something that can be triggered, checked, and reviewed. Which assertions must be traced back to source. Which sensitive information must be redacted. Which content cannot pass through a format gate alone. Which facts must carry their media-source boundary. If those only sit in a "things to watch out for" doc, they are useless. They have to enter the workflow. Enter the checks. Enter version control. Enter the responsibility line. That is COE's new seat. Not writing faster. Writing executable. So COE did not get replaced by an LLM in any simple way. What got replaced is the "policy draft producer" layer. What stays is the AI governance rule designer: the person who decides how policy enters the system, how it enters prompts, how it enters workflows, and how it gets held accountable and rolled back when it goes wrong. If COE still understands its deliverable as a Word doc, it will keep drifting to the edge. If COE can rewrite its deliverable as rules, boundaries, and executable protocols, it is not getting pushed out by AI. It is being forced into the seat where it should have been sitting all along. ## HRBP loses half. The remaining half is more expensive. SSC got a new price tag. COE got a new medium. That leaves HRBP. This middle pillar takes the hardest hit. Because HRBP's reason for existing is exactly that "judgment has to flow between humans, through a human." The information at the business site is too scattered. The boss's intent is too implicit. Employee emotion is too tangled. The real problems in the organization rarely sit in a system field. So HRBP, historically, did judgment translation in the middle. Once AI walks in, half of that reason gets pulled out. Meeting notes get cleaned up. Interviews get classified. Performance material gets summarized. Role profiles get generated. Employee feedback gets clustered. The big chunk of work HRBPs used to accumulate through time, relationships, and writing chops gets steeply discounted. This is not the end of HRBP. But HRBP really is reduced to half. The discounted half is the judgment translator. Collecting material. Tidying expression. Relaying information. Writing reports. First-pass synthesis. The half that stays is the organizational diagnostician. Reading what organizational problem actually sits underneath those materials. These two sound close. They are not. The judgment translator answers: What is the business saying? What should the boss hear? The organizational diagnostician answers: Why does this keep happening? Is the root role design, workflow, incentive, capability, culture — or is it the boss never made the intent clear? A judgment translator can do well on diligence and relationships. An organizational diagnostician cannot. It demands an organizational model, a real read of the business, a feel for people, and evidence discipline. AI can put material on the table. Material is not a conclusion. AI can tell you five departments are complaining the process is slow. It cannot decide for you: is it the wrong authority allocation, the wrong approval mechanism, or the boss never set a clear priority. That is what the new half of HRBP is worth. I would rather call it: Organizational diagnostician. Judgment-premium pricer. The organizational diagnostician decides which layer the problem lives in. The judgment-premium pricer decides which human judgments are still worth paying for, which judgments can be handed to AI for preprocessing, and which judgments must reserve the final human ruling. Wulf et al., in 2025 (arXiv:2507.14034), describe a spectrum of human–machine collaboration with HITL and HOTL. The terminology is not the point here. The point is the HRBP needs to know which loop they are standing in. When AI is uncertain, who catches it? When AI is more autonomous, who supervises? When AI generates a judgment, who decides whether it can enter the formal workflow? If HRBP still understands itself as "business partner," that word will keep getting thinner. Closeness to the business is a position. It is not value. After AI, closeness is no longer the core question. The core becomes: can you design the boundary between humans and AI in the judgment chain. Can you tell which judgments need human review, which judgments machine can run first, and where accountability sits when it goes wrong. This is also why I keep rewriting my own résumé. I no longer want to write myself as traditional HR or traditional OD. Because the seat after the three pillars is not "I know HR management." It is: I can split an organizational problem into roles, workflows, knowledge, accountability, and governance — five layers — and I know which layer has to be rewritten when AI walks in. For HRBPs, this is not pure bad news. The bad news is, the half of yourself that pays the bills through relay work is getting cheaper fast. The good news is, the half that can actually diagnose the organization, price judgment, and design the human–AI loop is getting more expensive. So the future of HRBP is not staying closer to the business. The new HRBP has to stand in the judgment seat: Who makes the call. Who reviews the call. Who carries the call. Who pays for the high-value call. By here, three-pillar collapse is no longer an HR-model question. It is a job-value revaluation question for the person in the corner office. ## Do not let HR rewrite HR by itself By here, all three pillars have changed. SSC did not disappear. It got repriced as an AI workflow orchestration layer. COE did not get replaced wholesale by an LLM. Its policy medium migrated from documents to rules, prompts, permissions, and system boundaries. HRBP did not disappear. Its discounted half is the judgment translator. The half that stays is organizational diagnosis and judgment pricing. This is not a disaster for the HR department. This is a protocol-layer rewrite for the HR function. So the person in the corner office should not hand this back to HR to fix on its own. When HR rewrites HR by itself, it usually turns into three motions: Swap the org chart. Update the job descriptions. Run another round on the capability model. These are not useless. But if you stop there, you are still patching the old interface. Fang Shengtao, Zuo Qian, and Fan Li, in _Chief Organization Officer: From Team to Organization (2nd ed.)_, frame organizational capability as a bundle of elements: formal organization, talent, culture, tools/equipment/AI, processes/mechanisms/systems, and other special elements. I am not unpacking that framework here. Just touching it once and moving on. This piece does not lean on a name to back it. But the framing is useful: AI should not be sitting in the tool procurement list. It is already inside the organizational-capability formula. That pushes the problem out of the HR department and back onto the desk in the corner office. Of the high-frequency service SSC used to carry, which becomes AI workflow, and which must keep a human exit? Of the policies and rules COE used to write, which migrate from documents into prompts, rules, permissions, and audits? Of the business-partner work HRBP used to do, which is just judgment translation, and which is the human organizational diagnosis worth keeping? If those questions are not answered, redrawing the three pillars one more time is still a business-card swap. And that kind of swap is the easiest way to manufacture fake progress. The decks look active. Org chart updated. Role titles updated. Capability dictionary updated. Workshops held. Group photos taken. An "AI-era HR transformation roadmap" written. Then back on the real ground, the problem is the same problem. Employee asks about a policy, the system answers wrong, nobody owns the fix. Business wants to use AI for performance summaries, nobody dares say which material is off-limits for evaluation. Boss wants to collapse a process layer, nobody can say cleanly who picks up the accountability afterward. This is not HR being lazy. This is the problem crossing the HR department boundary. AI split service, policy, and judgment apart. The signing seat is still painted onto the old org chart. That is the awkward part. If you can answer those questions, the next cut goes deeper. Because once roles get rewritten, the employment contract gets rewritten too. What you used to buy from an HRBP was their time — meetings, communication, materials, running process. After AI discounts the notes, the materials, the first-pass analysis, what is left is the organizational diagnostic judgment. How do you price that? What part of a person's judgment are you actually buying? At what price? Who is on the hook when it goes wrong? That is what the next piece, _The New Labor Contract_, takes apart. This piece stops here. The HR three-pillar model is not a keep-or-kill question. The service delivery, policy carrying, and judgment flow it used to carry have already been pulled apart and reassigned by AI. If the person in the corner office leaves HR patching the old blueprint, that is not transformation. That is renovation. --- ## Read on - Previous: [Why the Org OS Cannot Run AI](/blog/human-in-the-loop-02-why-org-os-cannot-run-ai) - Series hub: [Human in the Loop](/blog/human-in-the-loop) - Next: [The New Labor Contract](/blog/human-in-the-loop-04-new-labor-contract)