The demo is over. Making AI work in production begins now – Opinion
Source: Business Times
Article Date: 23 Jul 2026
Author: Abhas Ricky
Competitive advantage belongs to companies that can sustain, govern and defend the tech, says the author.
The hard part of enterprise artificial intelligence is no longer the pilot, but instead, everything that comes after.
The Decentralized Architecture initiative by the Massachusetts Institute of Technology (MIT) found that 95 per cent of enterprise generative-AI programmes have produced no measurable profit and loss impact, despite spending US$30 billion to US$40 billion.
Moreover, 42 per cent of companies abandoned most of their AI initiatives in 2025, up from 17 per cent in 2024.
The implication is blunt: Access to frontier models is now commoditised, and the durable advantage belongs to whoever can sustain, govern and defend AI once it leaves the demo.
Doing that requires shifting the core unit of account from tokens to trust.
In an enterprise context, trust is the operational certainty that an AI system is predictable, accountable and legally defensible.
The signal for boards is unmistakable: Regulators are now asking whether organisations can prove how AI behaves.
The “why” for AI adoption is settled; the entire contest has moved to the “how”.
Sustain: price the useful work, not the tokens
Token pricing flatters the economics.
Goldman Sachs projects an increase of 24 times the current level of token consumption by 2030, as agents proliferate. This means any cost modelled for each token is a number that compounds against you.
The discipline that survives contact with production measures AI through four lenses – unit economics, control, performance and operating burden – and collapses them into one metric: the cost for each useful task completed, not the cost for each token or model.
Put plainly, if you cannot price a completed task, you cannot defend the budget that funds it.
That same metric should decide where work runs.
Public cloud wins on speed, experimentation and access to frontier models; private and sovereign inference win on high-volume, regulated or latency-sensitive workloads.
The winning strategy is to bring AI to the data, not the other way around. Where the work runs is now the line between a margin and a write-off.
MIT found that AI bought from specialised vendors and run-through partnerships succeeds about 67 per cent of the time, while internal-only builds succeed roughly a third as often.
Practically, enterprises should define the business task first, match it to the simplest model that performs it reliably, and track the cost for each completed task.
They should then route each workload to the public cloud for speed and experimentation, or to private and sovereign environments for cost, control and latency.
Sustaining AI in production requires clear financial viability beyond the demo phase; today, 95 per cent of programmes cannot demonstrate this.
Govern: autonomy without an audit trail is just unmanaged risk
Control fails the moment a leader cannot explain why a system acted, which rules were applied, what data it used or who is accountable for the outcome.
That failure is structural, not incidental: Cloudera’s Data Readiness Index 2026 found only 10 per cent of Asia-Pacific respondents had all their data fully governed.
You cannot govern an agent’s decision on top of the data you have not governed first. The audit trail breaks before the agent does.
If data governance policies do not follow the data across every environment, scaling securely becomes impossible.
The risk compounds as agents shift from answering questions to taking action across tools and workflows.
Research and advisory firm Gartner expects agentic AI to drive at least 15 per cent of day-to-day work decisions by 2028, up from effectively zero in 2024, embedded inside a third of enterprise software.
Where pilots multiply across teams without shared standards, policies or visibility, autonomy turns fragmented, costly and unreliable. Scale without standards multiplies the surface area for failure.
To govern effectively, organisations must define what each model or agent is allowed to do, apply common policies across teams, and create audit trails.
They should then add orchestration that routes workloads by cost, latency, value, sensitivity and governance need, with explicit human checkpoints on higher-risk actions.
The stakes are concrete: Gartner predicts more than 40 per cent of agentic AI projects will be cancelled by the end of 2027, driven largely by weak risk controls and unclear value.
The goal, then, should be to match autonomy with oversight.
Defend: design for the day the model is wrong
Defence means engineering for what happens when AI is wrong, ambiguous or operating in a high-stakes context.
Singapore’s Model AI Governance Framework for Agentic AI makes this explicit: assess risks upfront, limit agents’ autonomy and their access to tools and data, and define the checkpoints for which human approval is mandatory.
Maturity is no longer measured by how much you automate, but instead by how deliberately you decide what you will not.
A practical safeguard model has four parts: clearly defined decision rights, tiered autonomy, full traceability of inputs, actions and outputs, and kill switches with policy overrides.
The Monetary Authority of Singapore recently launched the Safeguards for Agentic Finance at Runtime (SAFR) framework, which provides a useful example of this logic in practice.
Developed with industry partners for financial services, SAFR proposes run time governance checkpoints that verify and record an AI agent’s proposed actions before execution, helping ensure they remain within predefined mandates, policies and risk boundaries.
However, oversight must be selective. Reviewing everything destroys return on investment, which is precisely the trap behind Gartner’s prediction that more than 40 per cent of agentic AI projects will be cancelled.
The mature posture automates low-risk work and routes only ambiguous, high-impact or exception-based cases to humans. Human oversight should be a scalpel, not a blanket.
So, decide which actions agents may recommend, as opposed to executing them.
Gate high-risk tasks behind human approval, and lean on sampling, confidence thresholds and exception handling.
Track business-quality metrics such as error rate, escalation rate, compliance score, trust score and rework burden, with the same seriousness as efficiency.
Rather than applying brakes on AI adoption, defence involves building a control loop that compounds performance, while containing risk.
The unit of account is changing from token to trust
The economics of enterprise AI are being rewritten, because the bottleneck has moved from access to operation.
To sustain it, measure useful work and place workloads where they pay.
To govern it, enforce common standards, policy controls, traceability and orchestration.
To defend it, apply selective oversight, hard safeguards and systems engineered as carefully for failure as they are for success.
Frontier capability is now a purchase, not a moat. What compounds is the infrastructure, the governance and the operational certainty wrapped around it.
The 95 per cent of companies that stalled in the MIT study treated AI as something to acquire; the 5 per cent that broke through treated it as something to operate.
That gap is the entire distance between a pilot and production and increasingly, between the enterprises that will lead the next decade and the ones explaining to their boards why the demo never paid for itself.
Trust, not tokens, is the unit of account now. Build for it.
The writer is chief business officer and general manager of applied AI at Cloudera
Source: The Business Times © SPH Media Limited. Permission required for reproduction.
461