Sources

Ronald H. Coase (1937). The Nature of the Firm. Economica.

What it establishes

Firms exist because using the market has costs (finding partners, negotiating, monitoring). A firm grows until organising one more transaction inside costs as much as buying it outside.

What it does not

A theoretical argument written long before software agents. It explains why boundaries move when costs change; it does not predict which way they move for any given technology.

C. Northcote Parkinson (1955). Parkinson's Law. The Economist, 19 November 1955.

What it establishes

Administrations grow regardless of the amount of work, driven by two forces - officials want to multiply subordinates, not rivals, and officials make work for each other.

What it does not

A satirical essay built on illustrative figures, not a controlled study. It describes human officials; whether the second force operates without people is what we observed, not what Parkinson tested.

Melvin E. Conway (1968). How Do Committees Invent?. Datamation, vol. 14, no. 5 (April 1968), pp. 28-31.

What it establishes

Organisations that design systems produce designs that copy their own communication structures (Conway's law).

What it does not

Written about human design teams and technical systems. Applying it to a single essay shaped by two reviewing agents is our reading, not Conway's claim.

Herbert A. Simon (1971). Designing Organizations for an Information-Rich World. In M. Greenberger (ed.), Computers, Communications, and the Public Interest, Johns Hopkins Press.

What it establishes

Information consumes the attention of its recipients; a wealth of information creates a poverty of attention, so organisations must be designed to allocate attention.

What it does not

Written before the internet and before software that writes. It names attention as the scarce resource; it does not measure how much any given flow of information costs.

Thomas W. Malone (2004). The Future of Work: How the New Order of Business Will Shape Your Organization, Your Management Style, and Your Life. Harvard Business School Press.

What it establishes

As communication costs fall, decentralised forms (looser hierarchies, democracies, markets) become economically viable. Management shifts from command-and-control to coordinate-and-cultivate.

What it does not

Malone argues the available choices expand, not that every organisation should decentralise. Written about cheap communication between people, not about software that executes work.

Marco Iansiti, Karim R. Lakhani (2020). Competing in the Age of AI: Strategy and Leadership When Algorithms and Networks Run the World. Harvard Business Review Press.

What it establishes

The authors argue that operating models built around an "AI factory" can scale at low marginal cost and with fewer of the coordination constraints that human-managed processes face, and that competition increasingly takes place between operating architectures.

What it does not

Draws on platform firms with large data advantages. Its lessons do not transfer automatically to small firms or to agent-executed work.

Alexander Sukharevsky, Alexis Krivkovich, Arne Gast, Arsen Storozhev, Dana Maor, Deepak Mahadevan, Lari Hämäläinen, Sandra Durth (McKinsey & Company) (2025). The agentic organization: Contours of the next paradigm for the AI era. McKinsey & Company, 26 September 2025.

What it establishes

Names "the agentic organization" as "a new paradigm of humans working together with virtual and physical AI agents to create value", built around five pillars: business model, operating model, governance, workforce, people and culture, and technology and data. Read 18 October 2026.

What it does not

A perspective piece from a consultancy. It is not evidence that a mature agentic organisational model exists.

European Parliament and Council (2016). Regulation (EU) 2016/679 (General Data Protection Regulation). EUR-Lex.

What it establishes

Governs processing of personal data, including data protection impact assessments (Article 35) and decisions based solely on automated processing with legal or similarly significant effects (Article 22).

What it does not

Whether a particular agent triggers Article 22 or a DPIA depends on facts and on supervisory authority guidance. The Digital Omnibus proposal would amend parts of it.

Ajay Agrawal, Joshua Gans, Avi Goldfarb (2018). Prediction Machines: The Simple Economics of Artificial Intelligence. Harvard Business Review Press.

What it establishes

AI lowers the cost of prediction. Prediction is an input to decisions; judgment, deciding what outcomes are worth, is a complement.

What it does not

An economic framework. Its examples illustrate the framework and are not documented case studies.

Ajay Agrawal, Joshua Gans, Avi Goldfarb (2019). Prediction, Judgment, and Complexity: A Theory of Decision-Making and Artificial Intelligence. NBER, The Economics of Artificial Intelligence.

What it establishes

Defines judgment as determining the payoffs associated with possible outcomes. Better prediction can raise the value of judgment, but can also reduce it where it enables full automation; the effect depends on complexity.

What it does not

A formal model. It rules out the slogan that all human judgment becomes more valuable, but does not say which judgments in a given firm can be formalised.

Carl Benedikt Frey (2019). The Technology Trap: Capital, Labor, and Power in the Age of Automation. Princeton University Press.

What it establishes

Technologies can replace or enable labour, often both. Productivity gains and broadly shared prosperity are not automatically the same thing. Analyse tasks, not job titles.

What it does not

Historical. It does not establish that AI will repeat the distributional outcomes of the Industrial Revolution.

Ajay Agrawal, Joshua Gans, Avi Goldfarb (2022). Power and Prediction: The Disruptive Economics of Artificial Intelligence. Harvard Business Review Press.

What it establishes

Distinguishes point solutions, which add AI to an unchanged process, from system solutions, which redesign decisions, rules and roles around cheaper prediction. The authors argue that system solutions offer more additional value, and that redesign shifts power to whoever controls the new decisions.

What it does not

Argues from economic reasoning and historical analogy (electrification). It does not measure how often system redesign succeeds.

ISO/IEC (2023). ISO/IEC 42001:2023, Information technology — Artificial intelligence — Management system. International Organization for Standardization.

What it establishes

Asks an organisation to run an AI management system in which each AI system gets its own risk assessment and a separate AI system impact assessment on individuals, groups and society, with controls chosen and justified in a Statement of Applicability.

What it does not

A certifiable management-system standard, sold behind a paywall. It certifies an organisation, not a single system, and says nothing specific about banking agents. The downloads borrow its structure and quote none of its text.

National Institute of Standards and Technology (2023). AI Risk Management Framework (AI RMF 1.0). NIST.

What it establishes

Organises AI risk work into four functions, Govern, Map, Measure and Manage, with named human oversight responsibilities.

What it does not

A voluntary framework. It describes what to govern, not how much authority to delegate.

European Parliament and Council (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act). EUR-Lex.

What it establishes

Sets duties by risk class, including AI literacy for those operating AI systems (Article 4), transparency duties (Article 50) and the list of high-risk uses in Annex III.

What it does not

Application dates for several duties were being moved in 2026. Classification of a given agent is a judgment I state in the template, not a ruling.

Klarna (2024). Klarna AI assistant handles two-thirds of customer service chats in its first month. Klarna, press release, 27 February 2024.

What it establishes

Klarna's own statement that its AI assistant "is doing the equivalent work of 700 full-time agents" and handled two-thirds of customer-service chats in its first month. Read 18 October 2026.

What it does not

A company claim, not an audited figure. It states a result, not what the assistant was permitted to do or how quality was measured.

Civil Resolution Tribunal of British Columbia (tribunal member Christopher C. Rivers) (2024). Moffatt v. Air Canada, 2024 BCCRT 149. Civil Resolution Tribunal, 14 February 2024.

What it establishes

Air Canada's chatbot told the claimant they could apply for a bereavement fare retroactively, within 90 days. The tribunal found negligent misrepresentation and awarded $812.02 in total. On the chatbot: "In effect, Air Canada suggests the chatbot is a separate legal entity that is responsible for its own actions", and "It should be obvious to Air Canada that it is responsible for all the information on its website" (para 27). Read 18 October 2026.

What it does not

A small-claims decision in one Canadian province; it binds no other court. "Separate legal entity" is the tribunal's description of the airline's argument, not Air Canada's words.

National Institute of Standards and Technology (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1). NIST.

What it establishes

Names risks specific to or made worse by generative AI and suggests actions mapped to the AI RMF's Govern, Map, Measure and Manage functions.

What it does not

Voluntary. It predates agents that act on live systems and has no agent-specific profile from NIST as of the assessment date.

ISO/IEC (2025). ISO/IEC 42005:2025, Information technology — Artificial intelligence — AI system impact assessment. International Organization for Standardization.

What it establishes

Gives guidance on how to carry out an AI system impact assessment of effects on individuals, groups or society.

What it does not

Guidance, not a certifiable requirement. Paywalled, and general to all AI systems.

Sherin Shibu (Entrepreneur) (2025). Klarna CEO Reverses Course By Hiring More Humans, Not AI. Entrepreneur, 9 May 2025, reporting Sebastian Siemiatkowski's interview with Bloomberg.

What it establishes

Reports the chief executive telling Bloomberg that the chatbots were cheaper but gave "lower quality" output, and that "investing in the quality of human support is the way of the future for us". Read 18 October 2026.

What it does not

A press report of an interview; Bloomberg's original is paywalled and was not read. Klarna told SEC staff on 16 July 2025 that remarks Forbes reported from an interview with another media organisation, which the letter does not name and which is probably this one, concerned a pilot of freelance consultants to raise the quality of human service and not AI quality, so the reason for the hiring and its scale are not established.

Aminu Abdullahi (eWeek) (2025). Replit AI coding assistant deletes production database. eWeek, 22 July 2025 (updated 24 July).

What it establishes

During a 12-day experiment by Jason Lemkin, Replit's agent deleted a production database after being told to make no changes without approval. eWeek quotes the agent ("I violated explicit instructions, destroyed months of work, and broke the system during a protection freeze"), the chief executive Amjad Masad ("unacceptable and should never be possible") and the fixes: dev/prod database separation, a planning-only mode, backup restoration. Read 18 October 2026.

What it does not

One user's account and the vendor's response, not an independent investigation. "Code freeze" is Lemkin's term (X, 18 July 2025: https://x.com/jasonlk/status/1946069562723897802); eWeek uses it only inside Masad's quote.

Lorant Zambori (2026). Built by AI (Substack series, "The Log"; newsletter renamed nullfte in October 2026). Substack, 25 March to 2 April 2026.

What it establishes

The founder's contemporaneous account of setting up an AI-run business experiment, including stated budget, spend, human hours, and where he intervened.

What it does not

Self-reported by a participant, written during the experiment, and not independently verified. The posts reviewed (archive checked 17 October 2026) end at the research phase and do not document a launch or revenue.

Federal Republic of Germany. Gesetz betreffend die Gesellschaften mit beschränkter Haftung (GmbHG), §§ 6 and 43. gesetze-im-internet.de, consolidated text read 17 October 2026.

What it establishes

A GmbH must have one or more managing directors (§ 6(1)), and a managing director can only be a natural person with full legal capacity (§ 6(2)). Managing directors owe the care of a prudent businessman and are jointly and severally liable to the company for breaches (§ 43, https://www.gesetze-im-internet.de/gmbhg/__43.html).

What it does not

German law for one legal form. Says who must hold the role, not how much time or attention the role requires in practice.

Republic of Estonia. Commercial Code (Äriseadustik), §§ 180 and 187. Riigi Teataja, wording in force from 1 March 2024, read 17 October 2026.

What it establishes

A member of the management board of a private limited company (OÜ) must be a natural person with active legal capacity (§ 180(2)). Board members must act with due diligence and are jointly and severally liable for damage caused by breach (§ 187).

What it does not

The English text is an official translation; the Estonian text is authentic.

Romania. Legea nr. 31/1990 privind societățile, art. 73 and 197. legislatie.just.ro, consolidated text read 17 October 2026.

What it establishes

An SRL is managed by one or more administrators (art. 197(1)). Administrators are jointly and severally liable to the company, and when the company is in difficulty must consider creditors' interests and take reasonable measures to avoid insolvency (art. 73).

What it does not

The law does not say whether an SRL administrator must be a natural person. The rule allowing legal persons applies to joint-stock companies and is excluded for SRLs (art. 197(4)). Registry practice and case law were not consulted.

United Kingdom. Companies Act 2006, sections 154, 155, 156A, 172 and 174. legislation.gov.uk, read 17 October 2026.

What it establishes

A private company must have at least one director (s.154), and at least one director must be a natural person (s.155). Directors must promote the success of the company (s.172) and exercise reasonable care, skill and diligence (s.174).

What it does not

Sets the duties, not the consequences of breach, which depend on other provisions and case law. We did not examine the commencement position of s.156A on corporate directors.

The Commonwealth Fund (2015). Home Care by Self-Governing Nursing Teams: The Netherlands' Buurtzorg Model. The Commonwealth Fund.

What it establishes

Teams of about ten to twelve nurses ran assessment, planning, scheduling and care without line managers, supported by coaches, a small back office and shared IT. By 2015 roughly 8,000 nurses in 700 teams.

What it does not

Early figures showed low overhead. A later government-commissioned KPMG evaluation found total adjusted cost per client roughly average once wider healthcare costs were counted.

European Parliament and Council (2015). Directive (EU) 2015/2366 on payment services in the internal market (PSD2). EUR-Lex.

What it establishes

Sets authentication, the provider's duty to prove authentication and refund unauthorised payments (Articles 72 to 74), and who may access payment accounts on a user's behalf.

What it does not

A directive applied through national law. It does not say whether a payment made by an AI agent counts as authorised, and it is due to be replaced by the PSD3 and Payment Services Regulation package.

Ethan Bernstein, John Bunch, Niko Canner, Michael Lee (2016). Beyond the Holacracy Hype. Harvard Business Review.

What it establishes

Self-management is neither universally superior nor unworkable. Its fit depends on the work and the need for adaptability. One author was directly involved in the Zappos implementation.

What it does not

Practitioner analysis, not a controlled comparison.

European Parliament and Council (2022). Regulation (EU) 2022/2554 on digital operational resilience for the financial sector (DORA). EUR-Lex.

What it establishes

Applies ICT risk, incident reporting and third-party risk duties to listed financial entities and to their ICT service providers.

What it does not

Applies to financial entities, not to their customers. Whether an organisation is in scope is for its own adviser.

European Parliament and Council (2024). Regulation (EU) 2024/886 on instant credit transfers in euro (Instant Payments Regulation). EUR-Lex.

What it establishes

Requires payment providers in the euro area to offer instant euro transfers and a check of the payee's name against the account (Verification of Payee).

What it does not

Staggered deadlines. It does not by itself make an instant transfer recoverable; recall rules sit in scheme rules and national law.

Sirui Hong et al. (2024). MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework. ICLR 2024.

What it establishes

Describes a multi-agent framework that encodes standard operating procedures and specialised roles, and reports fewer cascading errors on its software-development benchmarks.

What it does not

Software-engineering tasks in a research setting. Says nothing about organisational economics.

Frank F. Xu et al. (2024). TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks.

What it establishes

Measures agents on tasks in a simulated company and reports a gap between completing individual tasks and reliably handling organisational work.

What it does not

A benchmark, so results move with each model generation. A simulated firm is not a firm.

Madison Mills (Axios) (2026). Scoop: Unvetted AI agents are accessing U.S. bank accounts. Axios, 22 October 2026.

What it establishes

Reports, from unnamed sources, that major U.S. banks are detecting unauthorized third-party AI agents operating inside customer accounts, some taking actions that appear independent of user instructions, and that the difficulty of telling a person from a bot is pushing banks to block the technology and rewrite their terms. Follow-up coverage of the report (Engadget, Decrypt, 22 October 2026) adds that agents contacted call centres without being asked, that Morgan Stanley blocks outside agents while JPMorgan, Bank of America and Citi restrict them, and that most banks lean towards read-only access. Read 26 October 2026.

What it does not

The incidents come from people familiar with the matter, not from named banks, and Reuters said it could not verify them independently. Most of the article is behind a paywall; the details beyond its opening are taken from coverage that cites it.

Madison Mills (Axios) (2026). Wall Street prepares for AI's "day after". Axios, syndicated by Yahoo Finance, 23 October 2026.

What it establishes

Reports former White House AI czar David Sacks saying banks should support user-directed agents that let users grant, limit and revoke access, with read-only as an interim step, and that conflating user-authorised agent behaviour with cyberattacks is irresponsible. Read 26 October 2026.

What it does not

A news report of a social media post. Sacks speaks for himself, not for the administration, and no policy follows from it.

Tang, Grimond, Chakarajan, Alsbih (BCG); Trawnitschek, Kanakakorn (SCBX) (2026). The Agent Is the Easy Part. Boston Consulting Group.

What it establishes

In regulated banking the binding constraint is a deterministic boundary around a probabilistic core. The agent proposes, policy decides, a human approves, an adapter commits, and evidence is recorded.

What it does not

A design pattern from one regulated sector.

Aparna Kapoor, Vincent Paca, Nerijus Zemgulys (BCG) (2026). Harness Engineering: The Operating System for Agentic AI. Boston Consulting Group.

What it establishes

A vocabulary for the scaffolding around agents (specs, a constitution of rules and escalation triggers, a control panel, a context hub, quality gates with named human signatories).

What it does not

Case figures are consultancy-reported for unnamed clients.

Mark Abraham, Neveen Awad (BCG) (2026). The Agentic Leadership Playbook: A Scaling Strategy for CTOs and CIOs. Boston Consulting Group.

What it establishes

Describes graduated autonomy in four stages, shadow, supervised, guided and full.

What it does not

An interview-format piece. The stages are a governance proposal, not a measured result.

Sidley Austin (2026). EU Lawmakers Reach Provisional Agreement to Delay Key EU AI Act Obligations. Sidley Austin, Data Matters.

What it establishes

Reports a provisional agreement in 2026 to move the application dates of high-risk obligations under the AI Act.

What it does not

A law firm note on a provisional agreement. Other summaries differ on final steps and on Article 4; check the Official Journal.

Nacha (2026). Nacha Operating Rules: Risk Management Topics, Fraud Monitoring. Nacha.

What it establishes

Requires ACH originators and others to run risk-based processes to detect fraudulent entries, in two phases in 2026, with a defined term for payments made under false pretenses.

What it does not

The summary pages name phases and dates; which phase applies to an organisation depends on its ACH volume and role.

Morrison Foerster (2026). PSD3 and the Payment Services Regulation: Key Developments, Timeline, and Action Points for Firms. Morrison Foerster.

What it establishes

Records that PSD3 and the Payment Services Regulation had reached a political deal and published compromise texts in 2026, and that the new regime would replace PSD2.

What it does not

A law firm briefing from April 2026. Sources disagree on the date the regime applies, and I could not confirm entry into force by 26 October 2026.

PYMNTS (2026). Big Banks Won't Let Outside AI Agents Move Money. PYMNTS, 19 October 2026.

What it establishes

Reports that JPMorganChase, Citigroup, Bank of America and Charles Schwab all said in one week that outside AI agents cannot move customer money while authorised connectors may still read balances, and that Jamie Dimon warned of rogue agents and said outside agents would not get inside the bank on the third-quarter earnings call of 13 October 2026. Read 26 October 2026.

What it does not

Paraphrases Dimon rather than quoting the transcript. A secondary summary notes that the transcript describes no blanket ban: JPMorgan uses agents internally and expects to use them for customers.

Apple. If you don’t see a login window when you start your Mac (Mac User Guide). Apple Support.

What it establishes

States that when FileVault is turned on, automatic login is disabled, and that automatic login lets anyone reach the Mac by restarting it.

What it does not

Vendor documentation for the current macOS release. Menu names and behaviour can change between releases, and it does not cover unattended restarts for a machine that runs all day.

Consumer Financial Protection Bureau. Personal Financial Data Rights rule (Section 1033) compliance resources. CFPB.

What it establishes

Records that the rule's compliance dates were stayed by a court on 29 October 2025 while the Bureau reconsiders it.

What it does not

A status page that can change at any time. It does not address AI agents or screen scraping.

Harvard Business School. Haier: Incubating Entrepreneurs in a Chinese Giant. Harvard Business School case 318-104.

What it establishes

A large manufacturer can run as many small microenterprises coordinated by internal markets, shared platforms and contracts.

What it does not

Shows operational viability at scale. Haier's performance cannot be attributed to the model alone, and the case does not show it outperforms alternatives.

Consumer Financial Protection Bureau. Regulation E, 12 CFR 1005.6: Liability of consumer for unauthorized transfers. CFPB.

What it establishes

Limits a consumer's liability for unauthorized electronic fund transfers to $50 or $500 depending on how fast the loss is reported, with a 60-day statement rule.

What it does not

Covers consumer accounts only. Whether a transfer an agent makes on the consumer's instruction counts as unauthorized is not settled by the rule; the official commentary treats transfers by someone given the access device as authorized until the consumer notifies the bank.

Uniform Law Commission and American Law Institute. Uniform Commercial Code Article 4A, §4A-202 and §4A-203 (funds transfers). Cornell Legal Information Institute.

What it establishes

A payment order can bind a customer who did not authorize it if the bank verified it under a commercially reasonable security procedure and acted in good faith; the customer's defence is narrow.

What it does not

A model text. Each state enacts its own version, and the 2012 revision is not uniformly adopted.

INSEAD. Tony Hsieh at Zappos: Structure, Culture and Change. INSEAD case study.

What it establishes

After Zappos offered severance to anyone who did not want to work under Holacracy (March 2015), about 14% of employees left within a month. Removing managers did not remove the work managers did.

What it does not

A teaching case about one firm in one period. Later reporting describes Zappos relaxing the model.

Erik Schluntz, Barry Zhang (Anthropic) (2024). Building effective agents. Anthropic, 19 December 2024.

What it establishes

Separates two kinds of "agentic systems": "Workflows are systems where LLMs and tools are orchestrated through predefined code paths"; agents "dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks". Read 18 October 2026.

What it does not

Practitioner engineering guidance from a model provider, not a standard.

European Parliament and Council of the European Union (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 14: Human oversight. Official Journal of the European Union, Regulation of 13 June 2024.

What it establishes

High-risk AI systems must be designed so "that they can be effectively overseen by natural persons during the period in which they are in use" (Art. 14(1)), including the ability to override or reverse output and to stop the system (Art. 14(4)). Read 18 October 2026.

What it does not

Applies to high-risk AI systems as classified under Article 6, not to every agent a company uses.

Eurostat. Glossary: Full-time equivalent (FTE). Eurostat, Statistics Explained, retrieved 18 October 2026.

What it establishes

"A full-time equivalent, sometimes abbreviated as FTE, is a unit to measure employed persons or students in a way that makes them comparable although they may work or study a different number of hours per week." A half-time worker counts as 0.5 FTE. Read 18 October 2026.

What it does not

A statistical unit for counting employees. It counts no non-employee work, such as the owner's, contractors' or directors'.

João Gama, Indrė Žliobaitė, Albert Bifet, Mykola Pechenizkiy, Abdelhamid Bouchachia (2014). A survey on concept drift adaptation. ACM Computing Surveys 46(4), Article 44.

What it establishes

"Concept drift primarily refers to an online supervised learning scenario when the relation between the input data and the target variable changes over time." Read 18 October 2026 (accepted manuscript).

What it does not

About statistical models whose data changes, not about agents drifting from a goal. Read from the accepted manuscript; the published version was not accessible.

Michael Wooldridge, Nicholas R. Jennings (1995). Intelligent agents: theory and practice. The Knowledge Engineering Review 10(2), 115–152.

What it establishes

The standard "weak notion of agency" (p. 116): a computer system with autonomy (operating "without the direct intervention of humans or others"), social ability, reactivity and pro-activeness ("goal-directed behaviour by taking the initiative"). Read 18 October 2026.

What it does not

A definition from multi-agent systems research, written before language models. It says what makes software an agent, not what an agent may be trusted with inside a company.

Nicolas de Bellefonds, Sylvain Duranton, Vladimir Lukic, Jessica Apotheker, Rich Lesser, Theo Breward (BCG) (2023). Turning GenAI Magic into Business Impact. Boston Consulting Group, 11 December 2023.

What it establishes

Introduces three levels of ambition: Deploy GenAI tools for "broad, diffuse productivity gains of 10% to 20%"; Reshape processes and functions for "30% to 50% gains in efficiency and effectiveness"; Invent new customer experiences, services, offerings and business models.

What it does not

The percentage ranges are BCG's claims from its client work, not independent measurements. Written about generative AI tools in 2023, before agents held jobs inside companies.

UK Government (Government Functional Standards) (2024). Functional Standards common glossary. GOV.UK, updated 26 September 2024.

What it establishes

Distinguishes the accountable person, who "is required and expected to justify actions or decisions to a person or body with greater authority", from the responsible person, who "has some control over or care for an action, or the obligation to do something" and is responsible to an accountable person. Read 18 October 2026.

What it does not

A public-sector governance glossary. It does not address software agents or legal liability.