The Great Inversion

For two hundred years, professional-services firms have largely been built in the same order: hire the professional, then buy them tools. The accountant, lawyer, administrator, underwriter or compliance professional is the unit of production. Software, paralegals and offshore teams exist chiefly to make that person faster.

That order is beginning to flip.

The argument is not that professional judgment disappears, nor that regulation will permit unsupervised automation. It is that the labour economics of regulated knowledge work are becoming incompatible with a model in which every routine unit of production must be performed personally by a scarce, licensed professional.

Demographic ageing and long qualification pipelines constrain the supply of professionals. AI and software systems make it possible to specify, execute, test and evidence an increasing share of repeatable production. The firms that win will not simply give professionals better copilots. They will reorganise themselves so that humans and agents work together as one production organisation.

Agents will perform permitted recurring work. Engineers and technical operators will encode, maintain and improve the domain logic. Licensed professionals will concentrate on ambiguity, materiality, exception handling, professional judgment, accountability and sign-off.

This is the Great Inversion.

Fund administration is the wedge. Humans and agents are the firm.

---

1. The labour constraint is real

The strongest version of the argument does not claim that every profession is uniformly “running out” of people. It makes a more useful claim: several regulated professions face an interlocking pipeline and succession problem, especially in ageing, high-income jurisdictions.

The issue is structural. Qualification typically requires education, supervised practice, examinations and regulator-controlled admission. A recruitment cycle cannot replace a cohort whose pipeline began five to ten years earlier.

In the United States, NASBA data reported 653,408 actively licensed CPAs as of August 2025, down from 671,855 a year earlier; 27,994 new candidates entered the CPA-exam pipeline in 2024, the lowest count since NASBA began tracking the series in 2008. The widely repeated claim that “75% of CPAs are retiring” needs care: the underlying 2015 AICPA estimate concerned members eligible to retire by 2020, not a forecast that three quarters would necessarily leave practice. The direction is nevertheless clear: the profession faces a material pipeline challenge. ¹

Germany illustrates the wider demographic setting. Destatis’s 16th coordinated population projection puts the working-age population aged 20–66 at 51.2 million in 2024. By 2035, it is projected to decline by 3.2 million to 4.9 million across high-to-low long-run migration scenarios; without net migration, the decline is about 6.2 million. The precise scenario is less important than the implication: large baby-boom cohorts are retiring into smaller successor cohorts. ²

Demand is becoming more complex at the same time. AIFMD II strengthens reporting and disclosure expectations around matters including delegation, fees and expenses, and supervisory reporting. The revised ELTIF framework has applied since January 2024, with supplementary regulatory technical standards taking effect later that year. This does not mechanically prove that administration demand will outstrip labour supply. It does increase the premium on reliable, traceable and scalable production systems. ³

2. The alternative workforce is technical

The inversion depends on a different division of labour.

The relevant contrast is not an engineer replacing an accountant or lawyer. It is a firm composed differently:

• Technical operators translate domain rules into workflows, structured data models, validations, test cases and escalation paths.
• Deterministic services and bounded agents ingest, classify, reconcile, calculate, draft and route permitted work.
• Qualified professionals investigate material exceptions, exercise professional judgment, supervise the system and accept responsibility for final outputs.

AI is useful because it can package precedent, rules and workflow assistance for less experienced workers. In a field experiment involving 5,179 customer-support agents, Brynjolfsson, Li and Raymond found that a generative-AI assistant increased issues resolved per hour by approximately 14% on average, with gains of roughly 34–35% among novice and lower-skilled workers. This evidence concerns customer support, not audit, legal advice or fund administration. It supports an augmentation and knowledge-transfer proposition, not a claim that professional judgment is autonomous. ⁴

The implication is narrower and more important: more repeatable application of professional knowledge can be captured in software, test suites, structured data and bounded agent workflows. Judgment remains scarce at the boundary—where facts are novel, rules ambiguous, stakes material or responsibility cannot be delegated.

3. The economic inversion

A headcount-led professional-services firm adds capacity mainly by hiring, outsourcing or acquiring teams. Its marginal cost remains linked to skilled labour because people perform much of the production.

An inverted firm changes the marginal unit. Each completed workflow can improve the next through accumulated rules, test cases, exception patterns, reviewer decisions, data mappings and evidence. A confirmed reviewer finding becomes a regression test, validation, workflow revision or escalation rule. The result is a firm that becomes more capable with every completed human-and-agent handoff.

Four metrics make the claim real or expose it as rhetoric:

The fourth measure is the most revealing. The ambition is not zero human intervention. In regulated work, a non-zero review and judgment floor is often necessary and desirable. The ambition is to identify every recurring intervention above that floor and decide whether it is a required control, a genuine exception, a source-data problem or an engineering opportunity.

4. Regulation shapes the firm

Regulation is neither a blanket ban on automation nor a licence to remove accountable people. It defines which activities may be automated or delegated, who retains responsibility, which controls must operate and what evidence must be retained.

The correct strategic objective is not “regulators will accept AI.” It is to build a firm able to demonstrate:

• Which exact artifact was produced and which data, rules and workflow version contributed to it.
• What validations ran and what they found.
• Which exceptions were raised, resolved or consciously dismissed.
• Who made every material decision and under what authority.
• Whether an output changed after review and whether approval was renewed as required.
• That the relevant records are retained, attributable, retrievable and intelligible to an examiner.

AIFMD II reinforces the importance of information, oversight, monitoring and control in relation to delegation. It does not prescribe a universal AI workflow and does not relieve an AIFM or other regulated entity of responsibility. It does, however, favour an evidence-rich architecture over fragmented, retrospective recordkeeping. ³

There is an early cross-vertical precedent in legal services. On 6 May 2025, the Solicitors Regulation Authority authorised Garfield.Law Ltd to provide regulated legal services through an AI-led service model. The authorisation is specific to that firm and regulatory perimeter; it is not a general authorisation for autonomous legal practice or regulated financial activity. Its significance is structural: automated delivery can operate within a regulated entity when governance, client protection and accountable professionals remain real. ⁵

5. Maker-checker becomes evidence

A maker-checker process is often evidenced by a signature, an approval field or an email trail. These may show that an approval occurred. They frequently do not establish the precise version reviewed, what checks ran, what the reviewer considered, what changed afterwards or why an exception was accepted.

The inverted firm makes the control a data structure.

Every material artifact receives a version identity. Findings, validations, reviewer questions, decisions and approvals bind to that version. A material change creates a new version and triggers explicit invalidation or re-routing rules. A confirmed finding sends work back and becomes a candidate for a permanent system improvement. A dismissal remains an attributed professional judgment with its rationale; it does not disappear.

Conceptually:

$$\text{Finality} = \text{valid artifact version} \land \text{required validations passed} \land \text{material exceptions resolved or authorized} \land \text{authorized human approval}$$

The ledger does not replace accountability. It makes accountability visible, inspectable and more scalable. It can also measure the human failure mode most likely to matter at high automated volume: rubber-stamping. Review duration, findings, dismissals, overrides, reviewer throughput and post-approval defects are control indicators, not merely productivity metrics.

6. What the firm becomes

The professional-services firm has historically been organised around professional labour. The partner’s name is on the door; junior staff are trained to do the same work; and tools serve the people who execute it.

The inverted firm is organised around the production organisation itself. Humans and agents work from the same queue, hand work to one another and leave one attributable record. They differ in authority, not in whether they are participants in the firm. Agents perform permitted repeatable work. People make the decisions that require judgment, legal authority and accountability.

The durable advantage is not a model or a chatbot. Models will improve and become widely available. The advantage is the accumulated operating record: workflow graph, rules, test suites, data lineage, exception history, reviewer decisions, authority model and measured control performance.

Footnotes

1. NASBA figures are summarised from NASBA Candidate Performance data and active-licence counts reported in the accounting-industry compilation, “The State of Accounting 2026.” For an important qualification of the commonly cited “75%” retirement figure, see “CPA Retirement Cliff: Separating Fact from Fiction.” Before external publication, replace these secondary references with an archived primary NASBA report and the original AICPA material. ↩

2. German Federal Statistical Office (Destatis), “16th coordinated population projection” (6 July 2026). ↩

3. Directive (EU) 2024/927 of the European Parliament and of the Council of 13 March 2024, “AIFMD II,” Official Journal of the European Union (26 March 2024); European Parliament, “ELTIFs: Delegated Act supplementing Regulation (EU) 2015/760” (2024). ↩ ↩²

4. Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond, “Generative AI at Work,” NBER Working Paper 31161. ↩

5. Solicitors Regulation Authority, “SRA approves first AI-driven law firm” (6 May 2025). ↩