TwinLadder Weekly

Issue #45 — A Regulator Just Asked Professionals to Show Their Working

18 July 2026 · Weekly intelligence on judgment, governance, and the boards accountable for both

Editor's Note

From Alex —

On 15 July the Solicitors Regulation Authority closed a consultation that had been open since 22 April. Its title is Strengthening our continuing competence approach. Outside the profession, almost nobody was tracking it.

The proposal asks a solicitor to record how they identified their own learning and development needs, "having considered all aspects of their role and work", and to keep the evidence for at least three years. It takes for the regulator a power to require specified solicitors — or a whole practice area — to complete prescribed learning where a competence concern is identified.

Hold the first of those. Since 2016 the profession's competence regime has run on an annual declaration. The solicitor reflects, writes a statement, and declares. The statement is the artefact. What the SRA has put out for consultation would make the artefact the working.

I want to be exact about the stage this has reached. The consultation is closed. Nothing is in force. The SRA now considers the responses it received, and whatever it adopts will arrive later, on its own timetable and in its own words. What exists today is a published regulatory position, and the position is that competence should be evidenced.

Two things make that worth a board's attention well outside England and Wales. The first is where it came from — a professional regulator, on a timetable nobody had marked, while everyone was reading the AI Act. The second is that in the same fortnight the Institute of Internal Auditors rewrote the Three Lines Model and aimed it at the board.

Liga has the analysis.

— Alex

What the SRA Put Out for Consultation

Strengthening our continuing competence approach opened on 22 April and closed on 15 July. It proposes the most significant change to how solicitors in England and Wales evidence competence since 2016, and it has three moving parts.

The record. A mandatory learning-and-development record, with evidence retained for at least three years. The operative sentence asks solicitors to "Record how they identified their learning and development needs, having considered all aspects of their role and work."

The floor. A minimum of three hours a year of facilitated group discussion on professional ethics, led by a solicitor who holds a current practising certificate. Hours, in a room, with a named person leading.

The power. The SRA proposes that it may "require some or all solicitors to complete specific learning and development where we identify a competence concern." Named individuals, or an entire practice area, directed by the regulator into prescribed learning.

Notice which of the three does the structural work. A declaration answers a question about a state: am I competent. A record of how the need was identified answers a question about a process: what did I examine, what did I find, what did I then do about it. The first can be produced in five minutes at renewal. The second has to have happened.

For readers who run model risk, the shape will be familiar. Validation has never accepted a modeller's assurance that the model works; it asks for the test, the sample, the challenge and the date. The SRA has floated that standard, applied to the human.

A Court Reached for the Same Instrument

The day before the consultation closed, a bankruptcy judge in Texas prescribed the remedy directly.

On 14 July the US Bankruptcy Court for the Southern District of Texas, Houston Division, entered a memorandum opinion in In re Rosslyn2016, LLC, Case No. 25-34507, a jointly administered Chapter 7. The court found petitioning creditors and their counsel in civil contempt of an order of 30 April 2026 and held them jointly and severally liable to the trustee for $29,877.00 in attorneys' fees and costs. The opinion opens on the stakes: "This Court addresses a matter of profound concern regarding the integrity of the bankruptcy [system]."

Among the fabrications was a quotation attributed to In re Dinnubilo — a sentence that appears in that opinion nowhere, and in no other published opinion either. Somebody read a passage, found it apt, and put it in a filing without opening the case it came from.

Then the sanction. Alongside the money, the court required counsel to obtain six hours of State Bar of Texas continuing legal education specifically on the use of generative AI in the courts, as a Rule 11(b) sanction.

Six hours, on a named subject, ordered by the body that caught the failure. That is the SRA's proposed power, exercised by a court, on a single lawyer, in the same week the regulator consulted on holding it in general.

One detail deserves its hedge. The hallucination database that tracks these matters records the tool in this case as Westlaw Precision, a paid professional legal-research product. That attribution sits in a database field; the opinion itself names no tool. If it holds, the brand of the vendor had stood in for the reading of the case.

Why Recording the Working Is a Harder Test Than It Sounds

The SRA's record turns on a self-assessment. The solicitor identifies the need; the record captures how. Which raises the question the week also supplied an answer to.

On 16 July, Frontiers in Psychology published work by Qiuhan Zhu and colleagues at Zhengzhou University on how people offload thinking to AI. The design was a three-wave time-lagged survey at two-week intervals, N = 589 university students and early-career knowledge workers. The authors separate two modes. Dependent offloading treats the machine as a substitute for one's own thinking — outputs accepted with minimal evaluation, the structuring of ideas delegated. Autonomous offloading uses the machine as a scaffold and keeps the structuring.

The two modes diverge downstream. Dependent offloading correlated with reduced intrinsic motivation and with greater transfer of cognitive authority to the machine; autonomous offloading preserved both. One sentence carries the consequence: "Both offloading modes yielded comparable immediate benefits despite divergent downstream correlates, suggesting that potentially maladaptive AI engagement may be difficult for users to detect from immediate experience."

Take the caveat with the finding. The outcome measure is perceived cognitive functioning — what participants reported about their own thinking, collected by survey. No unassisted performance test sits behind it.

The implication for a competence record is direct. Two solicitors work a Tuesday afternoon on the same kind of task. One reads the machine's draft against the file and rebuilds the argument; the other reads it, agrees, and sends it. Both finish on time. Both feel the work went well. At the annual declaration, both write that they kept their drafting skills current, and both mean it.

A record of self-identified needs inherits that blindness. It is only as good as the professional's capacity to feel the debt, and the debt is quiet by construction.

What would make such a record evidential? Three answers already exist, in disciplines that have had to solve it.

Count the re-judgments, not the overrides. Where a decision is visible on its face — an audit opinion, a clinical read, a credit memo argued in writing — the measurable control is the rate at which a competent human forms the judgment again, from the file, blind to the machine's answer. Overrides record only the cases where that independent judgment happened to differ. A deskilled reviewer and a flawless machine produce identical override rates. The re-judgment is the control; the override is its outcome.

Seed the queue. Where the work arrives as volume — a claims queue, a fraud desk, a transaction-monitoring team — the reviewer sees only what the machine surfaces, so failure is detectable only when the machine happens to fail. Testing the eye that rarely means testing it too late. Drop a handful of known-bad cases back into the stream each month, cleared by the machine, with the answer recorded in advance, and at month's end there is a number no dashboard would otherwise produce: of the planted misses, how many did the human catch. A fire drill for judgment, with a pass rate the board sets before the quarter starts.

Account for the volume nobody saw. A model triages five hundred items and surfaces twenty. Every competence measure lives on the twenty. The four hundred and eighty left the human loop entirely — decided, and silent. A record of how professionals maintained their judgment on the cases they were shown says nothing about the cases they never were. The standing question is what evidence exists that any of them were safe to leave unseen.

Each of the three costs money and produces a document. Each converts a claim about competence into a record an outside reader can audit — the standard the SRA has now put to the profession in public.

The Board Was Handed the Same Question, Twice

On 8 July the Institute of Internal Auditors published two updated Statements of Position, replacing the former position papers: the Three Lines Model, and the Role of the Internal Audit Function in Enterprise Risk Management. On the practitioner reading of the changes, the refreshed model targets the board as its primary audience, sets independence out as checkable tests, separates assurance work from advisory work with different safeguards, and asks internal audit to coordinate assurance across the organisation through assurance maps and aligned risk taxonomies. The IIA's own framing is that "collaboration, coordination, and reliance among the three lines...improve risk coverage and the reliability of information communicated to the board and senior management."

Read that against the SRA. One document asks a professional to show the working behind their own competence. The other asks a board to hold a map of who assures what, tested against criteria somebody can check. Both replace an assurance with an artefact.

Two more arrived in the same fortnight, from two more directions.

On 7 July Claudia Buch, Chair of the ECB Supervisory Board, wrote to the chief executive of every significant institution under European banking supervision about AI-enabled cyber threats. The letter places the duty squarely: "Responsibility for responding to the evolving cyber-risk environment primarily lies with banks' management bodies." Each bank must submit a comprehensive action plan to its Joint Supervisory Team by 31 October 2026, with roles, responsibilities and timelines named.

The day before, the FCA published the Mills Review, AI and the future of retail financial services, led by Sheldon Mills and drawing on 140 written submissions. On accountability it is direct: "Accountability becomes harder to trace, especially where models change over time and firms depend more heavily on model providers and external AI agents." The FCA's position is that the Senior Managers Regime still applies.

Then the sentence a stewardship practitioner wrote on 15 July, describing how asset managers should build machine assistance into voting. Will Goodwin of Tumelo set the requirement as a property of the record: "Every decision must be reconstructable: which policy applied, which rule fired, which data point was used." He wrote it about proxy votes. It is the test the SRA proposes for a solicitor's development record, the test the IIA's assurance map implies, and the test an examiner will apply to the ECB action plans in November.

The instrument keeps recurring because the alternative has stopped working. When a fluent machine sits inside the work, an assurance that the work was done well carries almost no information. The reconstruction carries all of it.

Russell Reynolds Associates, writing on 7 July and drawing on its own 2025 global board-culture study, puts at 43% the share of boards that have never undergone an externally run evaluation; the underlying sample goes unstated, so treat the direction as firm and the precision as approximate. Their framing: "Effective governance now depends on whether the board has designed itself to make better judgment more likely."

Roughly two boards in five have never had an outside reader of their own working.

Somebody Has Already Priced This

The objection to all of it is cost, and the objection is real. Recording how competence was established, sampling re-judgments, seeding a queue, maintaining a panel of experienced people to score against — all of it consumes hours that used to be billed. So the practice needs a body with a long enough clock to carry it.

One industry has already put an invoice on the formation of a professional and published the number.

In 2006 the ICAO adopted the Multi-crew Pilot Licence, built on the proposition that an airline first officer can be manufactured, competency by competency. It is competency-based instead of hours-based. Of its minimum 240 flying hours the majority may be flown in full-flight simulators; training runs in partnership with a specific airline from the first day; and it delivers a first officer to the right-hand seat in roughly eighteen months, with as little as forty hours in an actual aircraft. The European pilots' union has criticised exactly that dosage. Researchers at Lund University who studied graduates of the licence found no indication of significant shortcomings against conventionally trained pilots. The United States and Canada never adopted it.

The argument is live, and it is an argument about dosage. Lufthansa has run a flight school at Bremen since 1956 and folded it into a European Flight Academy in 2017, taking people with no flying experience to the group's cockpits in about two years. United opened an academy in Arizona in 2022 and drew a 2025 lawsuit from former students alleging the programme underdelivered — a designed pipeline can still be run badly. What the industry argues about is throughput, cost per seat and simulator hours.

Germany has published the same arithmetic for craft. Its federal vocational-training institute runs a recurring cost-benefit survey of firms that train apprentices. In the 2017/18 wave a German training firm spent an average of €20,855 per apprentice per year; the apprentice's productive work returned €14,377 of it; the firm carried the balance of €6,478 per apprentice per year. A visible net cost, carried at national scale, on purpose, and recouped in recruitment never incurred and in knowing exactly whom they had formed.

Gary Becker explained in 1964 why most firms decline that arithmetic. General training raises the trained employee's market wage by the full value of the skill, so the firm that pays for it funds a rival's hire. Judgment — the capacity to tell sound work from fluent work — is close to perfectly general. Last summer the economist Luis Garicano named the current version: the AI Becker problem, in which the machine eats the junior work whose fees quietly financed professional formation.

The tasks in question were inventoried in public on 13 July, by two of the firms doing the removing. The Institute of Student Employers interviewed Jenny Taylor at IBM and Jake Bustos at Coca-Cola on how entry-level roles are being redesigned. Taylor: "Our entry level hires have always performed real roles... The real skill now is knowing how to use AI." Bustos names what the machine has taken on: "AI is increasingly supporting junior employees with first drafts, research summaries, data analysis, presentation building." Both firms are keeping their programmes and hiring for adaptability, AI literacy, critical thinking and learning agility.

That list is a precise inventory of the work through which a junior used to acquire judgment, published by the firms that removed it. Set it beside the labour-market read from five days earlier. Indeed's Hiring Lab found on 8 July that "71% of the increase in software development job postings between May 2025 and May 2026 is from senior roles, and 37% is due to jobs that mention AI in their title." The market is hiring people who already have judgment, from a pool being refilled more slowly than it is drawn down.

The cost of formation falls now. The return falls ten to fifteen years out, on nobody's current number, in careers that outlast every executive currently measured on the spend. Management runs on a one-to-three-year clock. This erosion runs on a ten-year one. Inside a company one body has a mandate built to outlast everyone who reports to it, and that body is reading this.

What This Means for Boards Right Now

One. Treat the SRA consultation as a published regulatory position, and act on what it signals. Nothing is in force; the responses are still being considered. What the document establishes is that a professional regulator has written down an evidential test — record how the need was identified, keep it three years, submit to a direction to fix. Regulators read each other. The question for a board outside England and Wales is which of its own regulated professions sits closest to the same move, and whether the firm would have anything to hand over if asked this year.

Two. Fix the evidence before the training. A learning budget produces attendance records. What an outside reader can test is a re-judgment rate sampled blind against a held benchmark, a seeded-case pass rate for the desks that see only what the machine surfaces, and a sampling regime over the volume no human sees at all. Any AI competence programme funded without one of those three attached is buying attendance. On the Frontiers evidence, the professionals inside it will report that it worked.

Three. Put the reconstruction requirement on one page and check who already owes it. The ECB action plans are due to Joint Supervisory Teams on 31 October, with roles and responsibilities named. The IIA's refreshed Three Lines Model asks for an assurance map the board can read. The Mills Review says the Senior Managers Regime still reaches an autonomous system, and that tracing accountability through one is harder. Three separate bodies, one requirement: a decision that can be rebuilt afterwards from a record, by somebody who was not in the room.

Which leaves the question the week actually put on the table, and it belongs at the top of the next meeting: could any of our professionals show how they established their competence — and could we?

Reading List

What happens to a profession when the first-draft work that funded formation is absorbed by a machine, and what has to be paid for explicitly once it is gone: The Apprenticeship Is Breaking — and Almost Nobody Is Saying So
The task-level version: which entry-level work built which judgment, and what the removal costs a firm ten years out: The Broken Learning Ladder: AI Is Removing the Work That Built Expertise
The SRA's other significant AI decision, read for the conditions the regulator attached: SRA Approval of Garfield.Law: Regulatory Implications for AI Legal Services

What We Are Watching Next

Whether the SRA publishes a consultation response, and whether the power to direct prescribed learning survives into whatever it adopts
Whether another professional regulator in Europe — in accountancy, medicine or engineering — opens a comparable consultation on evidencing competence
What significant institutions put in the action plans due to their Joint Supervisory Teams by 31 October, and whether any board sets the risk tolerance those plans are built on before they go
Whether any internal audit function publicly maps AI onto its assurance map under the refreshed Three Lines Model, and which line is named as testing the models
Whether the FCA follows the Mills Review with anything that states how the Senior Managers Regime operates over a system that changes after it is approved

The next issue goes deeper into one of these. If you want a specific function or sector covered, reply to this email.

— Liga

TwinLadder Weekly is a weekly intelligence report on judgment, governance, and the boards accountable for both. Subscribe at twinladder.ai/newsletter. Forward this issue freely.

A Regulator Just Asked Professionals to Show Their Working