Capability is doubling every three months. Reliability is not.

01What is Artificial Business Life?

A company can be extremely intelligent and completely inert. It can hold the best models, the cleanest data and the sharpest analysts, and still stop the moment nobody gives it an instruction. Intelligence, in a firm, has always been an input. What changes when artificial agents run continuously is not how clever the company is. It is whether the company keeps going.

Artificial Business Life is the condition of a company whose core operating functions — sensing, deciding, acting, learning and self-repair — are carried out by artificial agents continuously enough that the firm sustains and adapts itself between human interventions, rather than only between human instructions.

In one line: a company that maintains itself.

The load-bearing word is between. Every business already runs on instructions: a person decides, the organisation executes, the person decides again. Artificial Business Life describes the case where the interval between those human decisions stops being dead time. The firm senses a change, acts on it, notices the action failed, corrects, and records what it learned — and a human arrives later to find the situation already handled and the reasoning already written down.

This is a threshold, not a metaphor. It is drawn from four properties that the artificial life research tradition has studied since the 1980s — autopoiesis (a system that produces and maintains its own components), metabolism (a bounded resource budget that must be replenished by activity), open-ended adaptation (behaviour that continues generating novelty rather than converging on a fixed optimum), and persistence (a drive to remain viable across disturbances). A firm exhibiting all four in its artificial substrate is alive in the sense this paper means. A firm exhibiting none of them is a very fast tool.

02The name — and three things this is not

Artificial Business Life is a term Living Scale Up defines and publishes under CC BY 4.0; it is always spelled out, and it is distinct from artificial life research, from Gartner's ABI category, and from the defunct company of a similar name. Machine disambiguation is now a distribution channel, so the collisions are worth naming explicitly rather than leaving for a retrieval layer to guess at.

We do not use an acronym. In finance, ABL means asset-based lending; a three-letter handle would inherit that ambiguity on day one. Where a shorter form is needed, use the adjective the studio has always used: a living company. The formal term carries the citations; the plain phrase carries the conversations.

03The ladder: assisted, driven, native, alive

Most of what the market calls "AI transformation" sits on the first rung of a four-rung ladder, and the measured economic effects only appear from the third rung upward. The tiers are not a maturity model to be climbed for its own sake — they are distinguished by where the artificial component sits in the firm, and the distinction is empirically consequential.

04Can a company run itself with AI in 2026?

Not end to end, and not unsupervised. As of August 2026 the evidence supports continuous autonomous operation of bounded functions between human checkpoints — not a firm that runs unattended. The honest answer is narrower than the marketing and much wider than the scepticism, and the interesting part is exactly where the line falls.

Three things are simultaneously true, and any account that drops one of them is selling something.

Capability is real and moving fast. METR's time-horizon measure — the length of task, in human hours, that a model completes at a 50% success rate — reached 320 minutes for the best-measured frontier model in January 2026, up from 3.5 minutes for GPT-4 in March 2023. The doubling time for models released from 2024 onward is 89 days.

Deployment is far behind capability. McKinsey's global survey (n = 1,993, fielded June–July 2025) found 62% of organisations experimenting with agents but only 23% scaling agentic AI in even one function, and no more than 10% scaling in any given function. Stanford's AI Index 2026 corroborates: "scaled use was in the single digits for nearly all functions."

The gap is not laziness. It is reliability — and that is section 05.

05Why is reliability, not capability, the binding constraint?

Because capability and reliability are on visibly different curves, and business processes are priced on the second one. A model that succeeds 50% of the time at a five-hour task is a remarkable research result and an unusable employee. The distance between those two statements is the whole feasibility question for 2026 and 2027.

METR itself is the most important witness against the naive reading of its own headline number. In a note published 22 January 2026, the laboratory states plainly that the metric measures "the amount of serial human labor they can replace with a 50% success rate" — not how long an AI can work unsupervised — and adds four qualifications that matter more than the headline:

• Reliability-critical work needs 98%+ success, and "a 50% time horizon doesn't mean tasks under that duration should be automated."
• 99%+ reliability horizons cannot be measured at all with current benchmark sizes.
• Domain variance is enormous. Visual computer-use horizons run 40–100× shorter than software and research horizons. One frontier model's "make coffee" horizon is roughly two minutes.
• Error bars span a factor of about two in each direction. METR's own researcher: "I really have no idea whether Claude's 'true' time horizon is 3.5h or 6.5h."

Independently, a Princeton team decomposed reliability into consistency, robustness, predictability and safety across 12 metrics and 14 agentic models, and found that "reliability gains lag noticeably behind capability progress" — despite steady accuracy improvements across 18 months of model releases. Their conceptual point is the one that matters operationally: accuracy alone cannot distinguish an agent that fails systematically from one that fails unpredictably at the same rate, and only the second kind breaks a business process. A 5% failure rate you can characterise is a design constraint. A 5% failure rate you cannot characterise is an unbounded liability.

The question is no longer what an agent can do. It is what an agent can be trusted to finish.

Anthropic's field telemetry closes the loop with the largest published sample. Across 998,481 randomly sampled public-API tool calls and more than 500,000 Claude Code sessions, the 99.9th-percentile turn duration nearly doubled between October 2025 and January 2026, from under 25 minutes to over 45 — while 80% of tool calls included at least one safeguard, 73% involved human oversight, and only 0.8% involved irreversible actions. Internally, on hard tasks, the success rate doubled between August and December 2025 while human interventions fell from 5.4 to 3.3 per session. The trend is unmistakable. The destination has not been reached.

06The five vital functions — and what each can actually do

A firm stays viable by performing five functions continuously, and as of August 2026 artificial agents can carry two of them well, two partially, and one not at all. The five come from Stafford Beer's Viable System Model, the cybernetic account of what any surviving organisation must do — a framework with fifty years of standing that nobody has yet mapped rigorously onto agentic architectures.

07What does Artificial Business Life create that AI adoption does not?

The measured returns to AI come from changing what the company is, not from adding AI to what it already does — and the difference shows up in headcount structure, in valuation per employee, and in the shape of the cost curve. This is the single cleanest empirical finding in the field, and most corporate AI budgets are pointed away from it.

Kim and Koning matched Y Combinator and PitchBook cohorts from 2020–2024 against Revelio Labs employment records and found AI-native firms are structurally different, not incrementally better:

Two operating cases give the shape of tier-1 value at enterprise scale, and honesty requires publishing both directions. Salesforce reduced customer support headcount from 9,000 to about 5,000, with support costs down 17% and more than a million conversations handled by agents over six to nine months — its CEO on the record, with numbers. Klarna, the most-cited case in the entire literature, is also the most-misquoted: its February 2024 press release stated that its assistant handled 2.3 million conversations in one month, "equivalent to the work of 700 full-time agents". That is a workload-equivalence claim, not a headcount reduction, and it has been reported as the latter for two years. Klarna subsequently rehired experienced human operators for ambiguous, emotional and edge-case work.

08Why not? Seven constraints, dated

Seven documented constraints stand between an agentic operation and a firm that maintains itself, and five of them are getting better while two are getting worse. Every constraint below carries a named publisher and a date; where a figure has a weak denominator, we say so rather than laundering it.

09What does it cost to keep a company alive?

The cost of intelligence collapsed and the cost of autonomy did not, because agentic workflows consume tokens faster than tokens got cheap. This is the least discussed and most decision-relevant economics in the field, and it inverts the standard extrapolation.

The collapse is real and well documented: the cost of querying a model scoring 64.8% on MMLU fell from $20 per million tokens in November 2022 to $0.07 by October 2024 — a factor of more than 280 (Stanford HAI, AI Index 2025, 7 April 2025). We note, against the convenient reading, that the 2026 AI Index does not update this figure; neither its technical nor its economy chapter contains a cost-per-token decline number. Any "2026 update" to the 280× is circulating without a source, and we do not publish one.

Against that collapse runs the opposite movement at the workflow layer. Where a 2023 customer-service exchange consumed hundreds of tokens, a 2026 orchestrated agent session consumes hundreds of thousands — and EY's practitioner estimate puts the per-interaction cost at roughly 30× its 2023 level. The figure has no published sample and should be treated as directional, but the direction is corroborated by the structural fact underneath it: reasoning, tool use, retries and verification all multiply token consumption per unit of work delivered.

Autonomy is not a capability you buy. It is a reliability you earn — and the invoice arrives as supervision, retries and verification.

The honest budgeting model for an Artificial Business Life function therefore has seven lines, not one: tokens and API, subscriptions and licences, platform infrastructure, governance burden, organisational change, expected failure and recovery, and the evaluation harness required to know the failure rate at all. The sixth and seventh lines are the ones omitted from every business case we have reviewed, and they are the ones that decide whether the function is cheaper than the human it replaces.

10Who is liable when no one is in charge?

A named human or a legal entity, always — because no jurisdiction on earth confers legal personality on an autonomous agent, and the European Parliament's own commissioned study recommends against ever doing so. This is the constraint most likely to be assumed away in a strategy document and least likely to move.

The authoritative European position is a July 2025 study for the Parliament's Committee on Legal Affairs by Prof. Andrea Bertolini, which rejects electronic personhood on the ground that "existing and reasonably foreseeable technologies do not seem to require the attribution of legal personality," and recommends instead strict liability for high-risk AI attached to a single identified responsible operator, precisely to eliminate causal uncertainty. Meanwhile the AI Liability Directive was withdrawn by the Commission in February 2025, leaving the revised Product Liability Directive — which does expressly cover software and AI as products — plus national tort law, and the fragmentation the study warned about.

Wyoming is routinely and wrongly cited as the counterexample. Its DAO Supplement (W.S. Title 17 Ch. 31, effective 9 March 2022) and its DUNA Act (W.S. Title 17 Ch. 32, effective 1 July 2024) give legal form to associations of human members. Neither grants standing to software. There is no AI-personhood statute in Wyoming or anywhere else.

11Why embrace the disruption now rather than wait?

Because capability arrives on every competitor's doorstep at the same moment, and the record of having operated does not. This is the compounding argument, and it is the only one that survives the reliability critique intact — indeed it is strengthened by it.

Consider what an AI-native company actually holds. Model capability is rented: it is delivered by frontier labs on a release schedule, at a published price, to everyone simultaneously. Waiting eighteen months does not cost you capability; you will get the better model either way, and cheaper. The naive conclusion is that waiting is free.

It is not, because a second asset accrues only by operating. Call it the metabolic record: the accumulated corpus of what this specific business tried, what failed and how, which actions were escalated and why, which were refused, what the true failure rate is on each action class, and where the boundary sits between reversible and irreversible in this particular market. That record is what turns a general-purpose model into a function you can leave running. It has no vendor. It cannot be purchased at any price, because it is a record of decisions in a context that only exists inside one company.

The reliability evidence sharpens this rather than softening it. Precisely because agents cannot yet be trusted at 98% on unbounded tasks, the scarce asset is knowing — with evidence, for your business — which tasks are inside the boundary. A firm that has been instrumenting, failing and correcting since 2026 will know. A firm that waits for the technology to be "ready" will arrive in 2028 holding the same models as everyone else and no record of its own, and will have to run the experiments then, at higher stakes, against competitors who ran them cheaply when the stakes were low.

12Six numbers we are not publishing

Six widely circulated figures about autonomous AI failed our sourcing standard during the research for this paper, and we are naming them rather than quietly omitting them. Each would have made this page more quotable. None of them has a publisher, a date and a denominator we can stand behind.

Discretion about method is not the same as vagueness about evidence.

13Where Living Scale Up stands

Living Scale Up designs companies to the Artificial Business Life standard, publishes the standard, and holds itself to it before asking anyone else to. The brand line — we build living companies — is not decoration. It is this paper's thesis, stated in four words since before the paper existed.

What that means concretely, in the vocabulary of section 06: the studio designs ventures whose Operations and Coordination run agentically from day one, whose Control layer carries hard external limits rather than trusted internal ones, whose Intelligence function is machine-generated and human-ranked, and whose Identity function sits with a named CEO who bears the liability. That configuration is the one this paper argues is both feasible and lawful in 2026. It is also, deliberately, the configuration that accumulates a metabolic record fastest — which is the compounding asset of section 11 and the reason the studio exists.

Three positions we will state plainly, because each is checkable:

• We define the term and give it away. Artificial Business Life is published under CC BY 4.0 and is not trademarked. A definition that competitors can adopt is a definition that can become a category; one they cannot is a slogan.
• We publish the counter-evidence. Every favourable figure on this page sits beside its strongest available refutation, and section 12 names six numbers that would have helped us. This is the standard we will be held to, including on our own results.
• We have not yet published our own operating numbers, and we will not invent them. Living Scale Up launched in July 2026. We will publish our first operating figures, with methodology and denominators, when we have a denominator worth reporting. Until then the honest answer is that we are a new studio applying a published standard, and that claim is falsifiable on the day the Index appears.

Operators who want to run a company built this way start here: livingscaleup.com/lead.

FAQQuestions, answered

What is Artificial Business Life?

Artificial Business Life is the condition of a company whose core operating functions — sensing, deciding, acting, learning and self-repair — are carried out by artificial agents continuously enough that the firm sustains and adapts itself between human interventions, rather than only between human instructions.

Can a company run itself with AI in 2026?

Not end to end, and not without supervision. As of August 2026 the evidence supports continuous autonomous operation of bounded functions between human checkpoints, not a firm that runs unattended. Anthropic's field telemetry of 998,481 sampled agentic tool calls (18 February 2026) found 73% still involve human oversight, and its own internal measurements show human interventions falling from 5.4 to 3.3 per session — falling, but not to zero.

Why is reliability, not capability, the binding constraint?

Because the two are on visibly different curves. METR measures the length of task a model completes at 50% success doubling roughly every 89 days for models released from 2024 onward, while Princeton's reliability study across 14 agentic models (arXiv:2602.16666, February 2026) finds that reliability gains lag noticeably behind capability progress. METR itself states that reliability-critical automation requires 98%+ success rates and that its method cannot measure 99% horizons at all.

What is the difference between an AI-native business and Artificial Business Life?

An AI-native business is architected around artificial intelligence from birth; Artificial Business Life describes a firm that additionally maintains itself between interventions. AI-native is a question of design origin. Artificial Business Life is a question of whether the firm's vital functions continue to run, adapt and self-correct when no one is instructing them.

Does any jurisdiction give AI agents legal personality?

No. As of August 2026 no jurisdiction confers legal personality on an autonomous AI agent. The European Parliament's own commissioned study (JURI, PE 776426, July 2025) recommends against electronic personhood and in favour of strict liability attached to a single identified operator. Wyoming's DAO and DUNA statutes give legal form to associations of human members, not to software agents.

Why embrace the disruption now rather than wait for the technology to mature?

Because capability arrives on everyone's doorstep at the same moment and the metabolic record does not. Model capability is rented from the frontier labs and is available to every competitor simultaneously; the corpus of decisions, outcomes, corrections and refusals that makes agentic operation safe in a specific business accrues only by operating. A company that begins in 2026 holds a record a company beginning in 2027 cannot buy.

SRCSources and vintages

• METR, Time Horizon 1.1 (29 January 2026) and Task-Completion Time Horizons of Frontier AI Models (leaderboard, updated 8 May 2026) — 228-task suite; 50% time horizons; doubling 196.5 days across 2019–2025, 131 days post-2023, 89 days for 2024-onward models. metr.org/blog/2026-1-29-time-horizon-1-1/
• METR, Clarifying limitations of time horizon (22 January 2026) — 98%+ requirement for reliability-critical automation; 99% horizons unmeasurable; visual computer-use horizons 40–100× shorter; error bars ~2× each direction. metr.org/notes/2026-01-22-time-horizon-limitations/
• Rabanser, Kapoor, Kirgis, Liu, Utpala & Narayanan (Princeton University), Towards a Science of AI Agent Reliability, arXiv:2602.16666 (February 2026) — 14 agentic models, 2 benchmarks, 12 metrics; reliability gains lag capability progress. hal.cs.princeton.edu/reliability
• Anthropic, Measuring AI agent autonomy in practice (18 February 2026) — denominator: 998,481 randomly sampled public-API tool calls plus 500,000+ Claude Code sessions; 80% of tool calls carry ≥1 safeguard, 73% involve human oversight, 0.8% irreversible; internal interventions 5.4 → 3.3 per session.
• Anthropic Alignment Science, Agentic Misalignment in Summer 2026 (13 July 2026) — Petri auditing tool, n = 20 runs per model per scenario; record-tampering results spanning 20/20 to 0/20 across frontier models.
• Kim (INSEAD) & Koning (Harvard Business School), AI-Native Firms, HBS Working Paper 26-090 (9 June 2026) — n = 2,786 Y Combinator firms (2,233 matched) and 47,007 PitchBook firms (41,214 matched), 2020–2024 cohorts, Revelio Labs records as of January 2025; −25%/−12% headcount, +30%/+76% valuation per employee, product channel β = −0.144, internal tool use not significant.
• McKinsey & Company / QuantumBlack, The state of AI in 2025: Agents, innovation, and transformation (5 November 2025) — n = 1,993 respondents, 105 nations, fielded 25 June – 29 July 2025; 62% experimenting, 23% scaling agentic AI, ≤10% in any function, 39% any EBIT impact. And State of AI trust in 2026 (25 March 2026) — ~500 organisations, fielded December 2025 – January 2026; security and risk the top barrier for nearly two-thirds.
• Stanford HAI, The 2026 AI Index Report (April 2026) — Chapter 2 benchmarks (OSWorld 66.3% vs 72.35% human baseline; SWE-bench Verified ~76.8%; GAIA 74.5% vs 92% human), Chapter 4 economy findings. And The 2025 AI Index Report (7 April 2025) — inference cost $20 → $0.07 per million tokens, November 2022 → October 2024, at MMLU 64.8%. The 2026 edition publishes no cost-per-token figure.
• Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (press release, 25 June 2025; analyst Anushree Verma) — analyst judgement, no published denominator; supporting webinar poll n = 3,412 self-selected attendees, January 2025. No 2026 reaffirmation published.
• Fortune (2 September 2025) — Salesforce customer support 9,000 → ~5,000 heads, support costs −17%, >1M agent-handled conversations, per Marc Benioff on The Logan Bartlett Show. Klarna press release (27 February 2024) — 2.3M conversations in month one, "equivalent to the work of 700 full-time agents" (a workload-equivalence claim, not a headcount reduction); subsequent rehiring of human operators reported by Forbes (16 July 2026) and Entrepreneur — secondary sourcing, no primary Klarna filing located.
• EY, Agentic AI enterprise token cost (1 June 2026) — ~$0.04 per chat (2023) → ~$1.20 per orchestrated interaction (2026), ≈30×; seven cost components. Practitioner observation, no published sample size or methodology — directional only.
• SecurityScorecard STRIKE team (9 February 2026) — internet-wide scan: 42,900 unique IPs hosting exposed agent control panels across 82 countries, 15,200 vulnerable to remote code execution; CVE-2026-25253, CVE-2026-25157 and CVE-2026-24763 patched 29 January 2026.
• European Parliament, Committee on Legal Affairs, Artificial Intelligence and Civil Liability, study PE 776426 (July 2025), author Prof. Andrea Bertolini — rejects electronic personhood; recommends strict liability and a single responsible operator. AI Liability Directive withdrawn, European Commission Work Programme (February 2025). Wyoming DAO Supplement W.S. Title 17 Ch. 31 (effective 9 March 2022) and DUNA Act W.S. Title 17 Ch. 32 (effective 1 July 2024) — human associations, not software agents.
• Council of the EU (29 June 2026) and European Parliament (16 June 2026) — AI Digital Omnibus: Annex III high-risk obligations deferred to 2 December 2027, Annex I to 2 August 2028; Article 50 transparency unchanged at 2 August 2026. Analyses: DLA Piper (30 June 2026), Gibson Dunn, Addleshaw Goddard (11 May 2026).
• Swiss Federal Council (12 February 2025) — decision to ratify the Council of Europe Framework Convention on AI; signed 27 March 2025, not yet ratified. Chambers and Partners, Artificial Intelligence 2026 — Switzerland — sectoral rather than horizontal regulation; consultation draft expected end-2026. Lenz & Staehelin (2 August 2025) — EU AI Act extraterritorial application to Swiss providers.
• Beel, Kan & Baumgart, Evaluating Sakana's AI Scientist: Bold Claims, Mixed Results, and a Promising Future?, arXiv:2502.14297 (20 February 2025) — $6–15 per generated paper, ~3.5 hours human involvement; 5 of 12 proposed experiments failed on coding errors; iterations adding ~8% more characters; median 5 citations per paper, 29 of 34 predating 2020. Original system: Lu, Lu, Lange, Foerster, Clune & Ha (Sakana AI), arXiv:2408.06292 (12 August 2024) — "less than $15 per paper", no hours figure.
• Masumori, Doi, Maruyama, Takata & Ikegami (University of Tokyo / Alternative Machine Inc.), OpenLife: Toward Open-World Artificial Life with Autonomous LLM Agents, arXiv:2606.31046 (30 June 2026) — n = 6 agents, single deployment, no control condition; budget-based metabolism, editable identity files, emergent per-contact trust levels. Cite as case study, not as measured effect.
• Peng, Wang, Liao, Lin, Yang & Zhang, A Survey of Zero-Knowledge Proof Based Verifiable Machine Learning, arXiv:2502.18535 (v2, 29 March 2026) — no production deployment listed. Gold, Freiberg, Isah & Shahabi (Inference Labs), JSTprove, arXiv:2510.21024 (23 October 2025) — ~1.6M-parameter models, 13.9–16.8 s proof generation.
• Fritsch, Müller & Wattenhofer (ETH Zurich), Analyzing Voting Power in Decentralized Governance: Who controls DAOs?, arXiv:2204.01176 — under 10% of total tokens participate; denominator is tokens, not holders. Counter-evidence: Okutan, Schmid & Pignolet, Democracy for DAOs, arXiv:2507.20234 — ~64% average participation across 14 SNS DAOs, 3,000+ proposals.
• Xinhua (28 May 2026) — China's humanoid robot lifecycle management platform and 29-digit identifier (2-digit country + 4-digit enterprise + 6-digit model + 17-digit serial), operated by the MIIT Humanoid Robot and Embodied Intelligence Standardization Technical Committee; launch enrolment 100+ companies, 28,000+ units, self-reported.
• Stafford Beer, Brain of the Firm (1972) and the Viable System Model literature — the five-system framework underlying section 06. The feasibility verdicts mapped onto it are Living Scale Up's, dated August 2026.