Earlier this week, I had an opportunity to share my views about building agentic organisations. Thank you Spark – CIO Academy for hosting.
The gap has already opened. It is opening at the individual level, at the firm level, and at the sector level — and in each case, it is compounding.
Start with the individual. Anthropic’s 2025 internal study surveyed 132 engineers and researchers and found that, compared to twelve months prior, they were using Claude in roughly 60% of their daily work and reporting a self-assessed 50% productivity gain — up from 20% the year before.[^1] Microsoft’s 2025 Work Trend Index found that 82% of leaders expect AI agents to expand workforce capacity within the next 12 to 18 months.[^2] These are not predictions. They are reported measurements from organisations already deep into this transition.
At the extreme, a solo founder in Israel — Maor Shlomo — built an AI-native application platform called Base44. Six months after launch, with 250,000 users and zero employees, he sold the company to Wix for US$80 million in June 2025.[^3] AI agents handled the bulk of execution. The unit of scale in that company was not headcount; it was agents. Solo-founded startups, once a rounding error in the venture landscape, now account for 36% of all new ventures — the highest share in over fifty years of startup data, according to Carta.[^4] Dario Amodei, CEO of Anthropic, placed 70–80% confidence on the first billion-dollar one-person company emerging in 2026.[^5]
Now look at the firm level. DBS Bank — a regulated financial institution in Singapore, with 35,000 employees — delivered more than S$1 billion in AI-attributable economic value in 2025, up from S$750 million the previous year, a verified 33% year-on-year increase. The bank now operates more than 430 production AI use cases, running across more than 2,000 deployed models.[^6] The typical Singapore enterprise, operating under the same regulatory environment and drawing from the same talent pool and vendor catalogue, runs fewer than ten pilots. That is not a technology gap. It is an organisational learning gap — of roughly five years, compounding at over 30% annually.
This is the context in which I want to introduce a framework. Not a theoretical one. A practical roadmap for how an organisation actually closes that gap — from a handful of pilots to hundreds of production use cases. The framework has three pillars: Architect, Economise, and Resilience. Get all three right and you build what I call the 100x Organisation. Miss any one of them, and you stall.
Pillar 1: Architect — Build Agents Like New Hires, Not Like Features
The first and most consequential decision an organisation makes about agentic AI is architectural, and most organisations make it wrong.
The mistake is treating AI as a faster way to do existing tasks: better search, faster summarisation, smarter chatbots. These are features — scoped, static, and entirely dependent on the user to initiate and steer every interaction. They do not accumulate institutional knowledge. They do not improve with feedback. They cannot be reassigned. They do the same thing in month six as they did on day one.
The 100x organisation treats agentic AI as a new class of workforce. The principle is simple: build agents like new hires, not like features. New hires get owned workflows. They accumulate context. They get better with feedback. They can grow.
This principle implies three specific decisions every leader must make.
Decision 1: What workflows do you hand to agents?
The answer is not whatever is technically easiest. Most organisations start with chatbots and FAQ assistants — high visibility, fast to deploy, politically safe. But an agent that doesn’t own anything isn’t doing agentic work. It’s a search box with manners. Every interaction starts from scratch.
The right criterion: workflows an experienced new hire would own within their first six months. High-volume. Multi-step. Structured inputs. Clear success criteria. The agent owns the work end-to-end — not just the suggestion.
A local example makes this concrete. ScamShield, built by GovTech Singapore, is Singapore’s AI scam-triage agent. In 2025, it processed 2.11 million scam checks and 832,000 citizen reports. Total scam cases fell 27.6% year-on-year; total losses fell 17.9% to approximately S$913 million.[^7] That is an agent owning a population-scale workflow — not a chatbot, not a search tool.
The practical test for any leader: if you would write a job description for this work, you have already described the agent’s job. The criteria for a good workflow — multi-step, structured, success-defined — are the same criteria for a good job description. If you can articulate what success looks like for a new hire, you can specify an agent.
Decision 2: Who builds — centralised or distributed?
The default most organisations choose is centralised: an AI Centre of Excellence builds for everyone. Three years in, twelve use cases have shipped. The team has become the bottleneck on the very capability it was set up to create.
The 100x organisation distributes building. A small central team builds the platform and the guardrails. Every business unit builds its own use cases on top. The centre’s role is to enable, not to execute.
This works because building an agent’s capabilities is no longer a technical task. Agent instructions are written in natural language. They are, in effect, job descriptions — and anyone who can articulate what they want a new hire to do can author an agent. That is why distributed models scale: you do not need an army of engineers; you need the people who know the work.
Moderna is the clearest example. A small central team built the ChatGPT Enterprise platform and the governance guardrails. The result: 750 custom GPTs deployed across legal, R&D, manufacturing, and commercial — in approximately 60 days, owned not by the AI team but by the business functions themselves.[^8] Shopify went further: in April 2025, CEO Tobi Lütke issued a company memo declaring that teams must demonstrate why AI cannot do the work before requesting new human hires.[^9] Distributed by mandate.
Decision 3: Is the agent designed to compound?
An agent that performs identically in month six as it did on day one is automation, not a member of the workforce. Three design choices determine whether an agent compounds or stagnates: memory across runs (does it carry context from previous interactions?), feedback loops that change behaviour (when a human overrides it, does that correction change future outputs?), and reusable templates (can the same architecture be adapted across teams rather than rebuilt from scratch each time?).
The Moderna example illustrates the third: each of the 750 GPTs was built on a shared skeleton — extraction, validation, escalation — customised per function but never rebuilt from scratch. That reusability is what made the scale possible.
Pillar 2: Economise — Treat Compute as a Managed Budget, Not an IT Line Item
There is a reframe required before the economics of agentic AI make sense. For two decades, every web engineering team in the world optimised latency — because a 100-millisecond delay in page load killed conversion. Latency was the make-or-break KPI. The analogous metric for agents is cost per outcome. Slow pages lose users. Expensive agents do not scale.
The inversion that should land clearly: we used to ask whether AI could do something more cheaply than a human. The question now is whether we are spending our AI compute budget wisely enough. Tokens are the new headcount line.
The hidden bill
Pilot economics are systematically misleading. One hundred tasks per day at a small scale looks fine on a credit card. One hundred thousand tasks per day is a completely different line item — and nobody sees it coming because nobody is watching the right number.
The reason is compounding cost. A reasoning agent re-reads its own outputs step by step. A workflow agent fans out into sub-agents. A reliability layer retries on failure. A single user request can trigger five, ten, or twenty model calls behind the scenes. Multiply that by growing context windows — the agent loads the entire conversation history on every turn — and by model choice (frontier model used when a fast, cheap model would do), and you have a recipe for bills that surprise at scale.
The receipts are real. Uber’s engineers, using agentic coding tools, burned through the entire 2026 AI budget in approximately four months.[^10] The experience has forced the company to revisit budgeting from scratch. These are not edge cases; they are what happens when capable AI tools meet organisations that have not yet built the discipline to govern compute spend.
Three levers that separate the portfolio of 400 from the portfolio of 50
The 100x organisation pulls three levers, consistently:
Model routing. Most of what an agent does — classifying a query, formatting a result, summarising a tool output — does not need a frontier model. A smaller, faster model does it in milliseconds at a fraction of the cost. Reserve frontier capability for the moments that genuinely require it. If every call is a frontier call, the routing decision has never been made.
Context discipline. Every turn the agent re-reads, you pay for again. Prompt caching — available in most major model APIs — lets you pay roughly a fraction of the normal cost on repeated static context. Most teams do not use it. They do not know they have it. That single gap often explains the difference between a sustainable portfolio and an unsustainable one.
Output accountability. If you cannot state what one resolved ticket, one completed brief, or one closed case costs in tokens, you cannot manage the portfolio. Cost per API call is not the metric. Cost per outcome is the metric — and it needs a named owner who answers for it monthly, like any other operating expense.
The structural move
Name a person. Not a team — a person. Call the function AI FinOps, fold it into your AI platform lead’s remit, or pair it with your CFO. The label is secondary. What matters is that one human, with a name attached to a slide, answers monthly for what your agents spent and what they delivered.
Compute is the new headcount line. You would not run a department where no one knew the salary bill. Do not run a portfolio of agents that way. That is the difference between the organisation that reaches use case 400 and the one that explains, in retrospect, why it stalled at 50.
Pillar 3: Resilience — Governance Is the Engine, Not the Brake
Most leaders think of governance as the thing that slows AI down. The 100x organisation thinks of governance as the thing that makes AI durable. Governance is the new SRE — site reliability engineering. Software did not scale because developers got better; it scaled because operations got disciplined. The same shift is happening for agents.
The data make the case plainly. MIT’s NANDA research programme studied over 300 publicly disclosed AI initiatives and surveyed more than 150 senior leaders. The finding: 95% of generative AI pilots fail to deliver measurable P&L impact. Only 5% reach durable production with real value.[^11] Gartner, in a separate June 2025 report based on polling more than 3,400 organisations, predicted that over 40% of agentic AI projects will be cancelled by end of 2027, driven by escalating costs, unclear business value, or inadequate risk controls.[^12]
Critically, the MIT research found the problem is not the quality of the AI models. It is the absence of operating discipline.
Why pilots die
A demo in a sandbox tells you nothing about an agent in production. In a pilot, you control the inputs. You watch the outputs. Failures are interesting. In production, inputs are unpredictable, outputs propagate through real systems, and failures are someone’s bad afternoon.
Five failure modes appear consistently across organisations that cannot move from pilot to production:
- No named owner of outputs — when something goes wrong, accountability is diffuse
- Permissions too broad — the agent can act outside its intended lane
- Failure handling absent — agents guess instead of escalating when they hit uncertainty
- Reviewers untrained — when humans correct the agent, the corrections do not feed back into the system
- No incident playbook — failures hide instead of getting diagnosed and fixed
These are not model problems. They are operating problems. And they are entirely solvable.
Four disciplines that keep agents alive in production
The organisations in the 5% that reach durable production are not luckier; they have, in many cases, informally reinvented a common set of disciplines. These can be codified:
Observability. Every agent action must be logged, queryable, and auditable. If you cannot see what the agent did, you cannot govern it. If you cannot govern it, you cannot trust it in production.
Least-privilege access. Write access only where the workflow requires it. Read-only by default. The most expensive incidents consistently come from agents operating with broader permissions than their specific workflow demands.
Failure design. The agent stops and escalates when it encounters a wall. It does not guess. The most dangerous agent is the one that fakes confidence — that produces a plausible-sounding answer to a question it cannot reliably answer. Building explicit escalation paths is not a concession; it is what separates a trustworthy system from a liability.
Accountability ownership. Every production agent has a named human owner — not a team, a person. If you cannot name that person, you have a governance gap. The owner reviews what the agent actually did, at a regular cadence. When the agent makes a bad decision, that person knows within the hour.
The structural move
Build the cadence. Weekly review for your highest-risk agents. Quarterly audit across the portfolio. A postmortem for every meaningful incident. This is not bureaucracy; it is the operating rhythm that the 5% who reach durable production have in common.
There is a useful analogy from software engineering: before SRE, every engineering team invented their own operations practices, inconsistently. After SRE, “SLO,” “error budget,” and “blameless postmortem” became vocabulary that any engineer at any company could walk in and apply. Agent governance is at the same pre-codification moment. The vocabulary is stabilising. The organisations that start building these practices now will not have to invent them under pressure later.
Governance turns pilots into a workforce. Without it, you scale failure faster than value.
The Homework
The organisations that will define their industries in 2030 are making their agent-infrastructure decisions this year. Every quarter of delay is a quarter of compounding you do not get back.
Three numbers to know by next week:
01 · ARCHITECT — Decide what your number of production agents will be twelve months from now. Then count what you have today and start closing the gap. The target sets the pace; the count tells you where you are starting.
02 · ECONOMISE — Set a token budget alongside that number. If you cannot tell me what your AI compute cost was last month, you cannot manage what you deploy next quarter.
03 · RESILIENCE — Name the human accountable for each production agent. Not a team — a person. If you cannot name them, that is your first governance gap and your first action on Monday morning.
The question is not whether agentic AI will reshape your organisation. That question has been answered. The question is whether you are the architect of that change — or subject to it.
References
[^1]: Anthropic. How AI Is Transforming Work at Anthropic. August 2025. https://www.anthropic.com/research/how-ai-is-transforming-work-at-anthropic
[^2]: Microsoft. 2025 Work Trend Index Annual Report: The Year the Frontier Firm Is Born. April 2025. https://www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born
[^3]: TechCrunch. 6-month-old, solo-owned vibe coder Base44 sells to Wix for $80M cash. June 18, 2025. https://techcrunch.com/2025/06/18/6-month-old-solo-owned-vibe-coder-base44-sells-to-wix-for-80m-cash/
[^4]: Carta. Solo Founders Report 2025. https://carta.com/data/solo-founders-report/ — “Solo founders now start 36.3% of all new companies.”
[^5]: Inc. Anthropic CEO Dario Amodei Predicts the First Billion-Dollar Solopreneur by 2026.https://www.inc.com/ben-sherry/anthropic-ceo-dario-amodei-predicts-the-first-billion-dollar-solopreneur-by-2026/91193609
[^6]: DBS Bank Annual Report 2025 — CIO Statement; Fintech News Singapore. DBS CEO Sees AI-Driven Revenue to Grow from S$750 Million to Over S$1 Billion This Year. https://fintechnews.sg/122167/singapore-fintech-festival-2025/dbs-ai-revenue/
[^7]: Singapore Police Force. Annual Scam and Cybercrime Brief 2025.https://www.scamshield.gov.sg/files/Scams%20and%20Cybercrime%20Briefs/2025_annual_scams_and_cybercrime_brief.pdf— ScamShield report card: 2.11M checks, 832K reports.
[^8]: OpenAI. Moderna Case Study. https://openai.com/index/moderna/ — 750 custom GPTs within approximately two months of ChatGPT Enterprise adoption.
[^9]: CNBC. Shopify CEO says staffers need to prove jobs can’t be done by AI before asking for more headcount. April 7, 2025. https://www.cnbc.com/2025/04/07/shopify-ceo-prove-ai-cant-do-jobs-before-asking-for-more-headcount.html
[^10]: Briefs.co. Uber Spends Full 2026 AI Budget in 4 Months. https://www.briefs.co/news/uber-torches-entire-2026-ai-budget-on-claude-code-in-four-months/
[^11]: MIT NANDA. The GenAI Divide: State of AI in Business 2025. August 2025. Cited in Fortune: https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
[^12]: Gartner. Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027. June 25, 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
Leave a Reply