跳到正文
原文
Accounting Today(美国会计行业媒体)· L. Gary Boomer·· 11 天前站内评分47

L. Gary Boomer认为智能代理工作队伍首先是组织设计问题

Boomer's Blueprint: The agentic workforce is an org design problem

导读

L. Gary Boomer提出,智能代理工作队伍不是可采购的功能,而是需要配置权限、责任与控制的角色。每名接触实际工作的代理都应持有包含八项属性的书面规格并由具名负责人承担责任,自主程度分为建议、附理由建议、经批准执行和在边界内执行四级。作者建议先用决策边界图区分预测与判断,选择高业务量且监管暴露度低的流程并行运行,以人工推翻率检验设计,并同步调整初级员工培养机制。

正文
AI agent

Last month, I argued that most firms have a drawer full of AI tools rather than an "intelligence stack." This month, I want to talk about the harder half of the same problem. Once you have the stack, who works inside it?

Processing Content

Every vendor with a demo has picked up the phrase "agentic workforce," which makes it sound like something you buy. But an agent isn't a feature you switch on. It is a role you staff, and our profession already has the discipline required to staff it well.

Start with the premise that your org chart isn't a picture of your firm. It is a latency map. It records how long information takes to travel and how many people it must pass through before someone can act. Every box exists because coordination was expensive. When coordination gets cheap, the map no longer describes the territory.

The end of headcount as the unit of capacity

"The Organizational Singularity" by Salim Ismail names the new operating logic "elastic agency." Instead of a fixed pool of people arranged in a hierarchy, you run a single pool of capability that blends human and synthetic, internal and external, and scales up or down without reorganizing the firm. Two artifacts replace the org chart at the working level:

  1. A capability registry that says what the firm can do and who or what does it; and,
  2. A graduated authority model that says how much latitude each has.

For accounting and advisory firms, this changes the busy season conversation more than anything else on the horizon. Capacity has always been a headcount question, which made it a recruiting question, which made it a market you don't control. Elastic agency turns part of that into an architecture question that you do control. Not all of it. The judgment work doesn't become elastic, and you should be suspicious of anyone who says it will. But the coordination scaffolding around the judgment work absolutely does.

No spec, no agent

Our profession has an advantage that I don't think most firm leaders have fully registered.

The framework requires every agent touching real work to carry a written specification with eight properties:

  1. A stated purpose;
  2. An autonomy tier;
  3. A permission envelope;
  4. A memory boundary;
  5. Escalation rules;
  6. An evaluation suite;
  7. A telemetry and audit trail; and,
  8. A reusability scope.

One named human owner, accountable for what the agent does, sits above all of that.

Read that list again in our vocabulary: scope, authority limits, access rights, retention policy, escalation, testing, audit trail, and ownership. We've been writing this document for a century. We call it a control. The rule that Ismail recommends is, "No spec, no agent."

The autonomy tier is the property partners will want to argue about. There are four rungs, and you climb them one at a time. The agent advises, meaning a human asks and the agent answers. The agent recommends, meaning it proposes options with reasoning attached. The agent executes with approval, meaning it does the work and waits for a signature. Or the agent executes within bounds, meaning it acts on its own inside a defined envelope and reports afterward. Most firms should have almost nothing on the fourth rung in year one. The ones rushing there usually don't have an evaluation suite to tell them when it goes wrong. Resist the temptation to invent a fifth category. Four is enough, and the discipline is in the promotion, not the taxonomy.

Two of the eight properties deserve extra attention because firms tend to skip them.

The permission envelope must have scope isolation, approval thresholds for destructive actions and soft-delete windows for anything irreversible. The cautionary case in "The Organizational Singularity" involves a misconfigured token that let an agent go from normal operation to a deleted production database and three months of erased backups in nine seconds. Nobody was malicious, but the envelope had no walls. Now imagine the equivalent inside a firm holding client financial data under a professional standard of care.

The reusability scope determines whether your agents become capital or clutter. An agent built for a single engagement is an artifact. An agent built to be forked, adapted and reused across the practice compounds. Ask the question before you build, because retrofitting it is expensive.

The layer that changes the most

Partners will ask what this does to the people. The honest answer is that it hits hardest in the middle, where firms are least prepared.

The framework describes the shift as coordinator to exception architect. Something like 80% to 90% of a middle manager's coordination work is within current agent capability. What remains is judgment, relationships, mentoring and resolving genuine ambiguity. That's a better job. It's also a different job, and calling it an opportunity without funding the transition is a category error dressed up as good news.

Take the objection seriously before you act on the opportunity. Managers do invisible translation, integration and tacit-knowledge work, and firms that cut the layer without building the capability underneath watch innovation collapse within two to three years. That objection is correct, and the answer isn't to dismiss it. The answer is order of operations. Translation migrates into the "Interpret" layer of the intelligence stack. Integration migrates into "Orchestrate." You must deliberately elicit tacit knowledge before the person carrying it leaves. Build the stack first, then redeploy the people.

Which brings me to the problem I've been unable to solve and will not pretend otherwise. Call it the missing junior loop. We've always trained future managers and partners by having them do the entry-level coordination work. Tie out the schedules, chase the confirmations, prepare the return, get corrected and repeat. That's the work agents will absorb first. Automate it without replacing the apprenticeship, and you're borrowing from a talent pipeline that takes 10 years to refill.

There are partial answers. Use the intelligence stack itself as an apprenticeship platform, where the junior reviews and overrides the agent, and every override becomes training data. Compress the apprenticeship deliberately rather than letting it evaporate. Build external pipelines through universities and programs. None of this is solved, but firms that treat it as a design problem now, rather than discovering it as a staffing crisis in 2031, will own a talent advantage that compounds for a decade.

Retention changes shape

There's one more consequence firm leaders should think through carefully: When AI collapses the cost of coordination, it collapses the cost of leaving in the same motion. A high-judgment professional who once needed a firm's infrastructure, staff and systems to serve clients well needs materially less of it every year. Retention by friction is finished, and every firm still relying on it, whether through deferred compensation, process dependency or simple inconvenience, holds a wasting asset.

"Retention by resonance" replaces it, according to Ismail. That's built on three things:

  1. Consequence, meaning the person's judgment governs real scope rather than one file at a time;
  2. Legibility, meaning who decided what is visible by design rather than absorbed into a partner's name; and,
  3. Identity, meaning a purpose specific enough to exclude people (a purpose that welcomes everyone binds no one).

Compensation in this model is hygiene. Pay to parity, then stop competing there, because you won't win a bidding war against someone whose cost structure no longer includes your overhead.

Where to start on Monday

Start with a decision boundary map rather than an agent. Take one service line and decompose its recurring decisions into prediction and judgment. Route the prediction work to agents with a written spec and a named human owner. Keep the judgment with the professional whose name goes on the work. Then put the review on the calendar quarterly, because the boundary moves as the tools improve, and a boundary you set once will be wrong within a year.

Pick one workflow for the first wave. The best candidate is high volume, rule-clear, measurable, reversible and low in regulatory exposure. Run it in parallel with the existing process, with success criteria agreed before you begin. Then watch the human override rate. If it falls quarter over quarter, the agent is learning, and the design is sound. If it stays flat, you bought software.

Run one check on the whole program. If your agents map one-to-one onto your current roles, you've automated the org chart rather than redesigning the firm. The point isn't to give every person a digital assistant. It's to organize the work around how decisions get made.

The machine carries the coordination. The professional carries the judgment, the relationship, and the responsibility. Design the roles that way on purpose, and the agentic workforce stops being a software category and becomes your firm's operating model. Think, plan, grow!

GaryBoomer

Visionary and strategist

L. Gary Boomer, CPA, CITP, CGMA, is the visionary and strategist at Boomer Consulting Inc., in Manhattan, Kansas. Read full bio

来源:Accounting Today(美国会计行业媒体) · accountingtoday.com

审计热点 是聚合摘要和阅读索引,原文版权归各来源所有,引用请以官方原文为准。如果你是来源方,希望更正、下架或调整展示方式,可以通过反馈页联系我们。

来自微信公众号的内容,本站仅呈现摘要与原文链接,不转载文章全文;文章权利归属原作者及原公众号

关于使用条款隐私信源