Why AI Agents Stall in Southeast Asian Enterprises, and the Trust Layer That Gets Them Live

The demo worked and the pilot saved hours, yet the project stalls at approval. In Southeast Asian enterprises this is rarely a model problem. It happens because nobody can show the business owner, IT and security what the agent may do, who approved it, and how to undo it.

This post explains that "trust layer", gives a 7-question check for your own project, and describes a 4-week pilot that builds it. A free check for 5 companies is at the end.

Interest is high, production is rare

Adoption looks strong on paper, but most firms cannot move from pilot to production.

• IDC reports that 42% of Southeast Asian organisations already use agentic AI and 79% are exploring use cases (Techgoondu).
• Adobe’s 2026 study of 461 large Southeast Asian firms found 57% stuck in a "frozen middle" of pilots that do not scale. Only 15 of every 100 AI ideas reach production (e27).
• In the same study, 36% of AI proposals fail IT evaluation and 18% are rejected at security review. Only 12% of firms have an AI owner with budget authority.
• Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing cost, unclear value and weak risk controls (People Matters).

The IDC respondents name the same worries: data-privacy breaches (51%), security flaws from autonomous actions (48%), and unintended consequences (47%). These are trust problems, not accuracy problems. Note that the Adobe sample covers only firms with at least US$100 million in revenue.

Regulators are asking for the same controls

Rules differ by country, but they point the same way: bounded access, human approval for risky actions, and a record of what happened.

• Singapore. IMDA released a draft Model AI Governance Framework for Agentic AI on 22 January 2026. It is voluntary. It asks firms to assess risk by the agent’s scope, reversibility and autonomy, limit tool access, require human approval for high-stakes or irreversible actions, test before launch, and monitor after (Davis Wright Tremaine).
• Vietnam. The AI Law took effect on 1 March 2026 and sorts systems into high, medium and low risk. High-risk systems need human oversight mechanisms. Existing systems must comply by 1 March 2027, or by 1 September 2027 in finance, healthcare and education (Duane Morris).
• Thailand. The Personal Data Protection Act already applies to agents that touch personal data, including a 72-hour deadline to notify a breach.

A company that builds the controls now does not have to rebuild when a voluntary framework becomes a customer or auditor requirement.

Six controls that answer the questions managers ask

A trust-first agent can answer each question below before anyone asks it.

flowchart TD
    A["Work request"] --> B["Agent proposes an action"]
    B --> C{"Within its permissions?"}
    C -->|"Yes"| D{"High-risk action?"}
    C -->|"No"| X["Do not execute"]
    D -->|"Low risk"| F["Execute"]
    D -->|"High risk"| E["Named approver reviews"]
    E -->|"Approved"| F
    E -->|"Rejected"| X
    F --> G["Write to audit log<br/>(searchable by internal audit)"]
    X --> G

The diagram shows where approval, permissions, audit trail and rollback act on an action. A kill switch stops the agent at any point and undoes its recent actions.

Control	What the manager or auditor asks	What a trust-first agent does
Approval	Who approved this action?	High-risk actions are proposals until a named approver signs off. Low-risk actions run automatically within set limits.
Permissions	What can it touch?	Least privilege per tool. Read-only by default, with write access granted action by action.
Audit trail	Can we reconstruct what happened?	Every input, step, tool call and cost is logged where audit can search it.
Rollback	Can we undo it?	Reversible actions first, a dry-run mode, and a kill switch that stops the agent at once.
Evaluation	How do we know it is right?	A test set of real past cases is run before launch and after every change.
Language	Can every team read it?	Approval screens, logs and manuals in English and the local language.

None of this is exotic. It is ordinary engineering discipline applied to a system that can take actions.

Seven questions to check your project

Tick what is already true. Fewer than five ticks means the project will probably stall at approval.

• [ ] You can name one task with a clear owner, a monthly volume and a cost per case.
• [ ] You know what a correct result looks like and have a few dozen past cases to test with.
• [ ] You have listed every system the agent needs, with read and write access separated.
• [ ] Someone with authority has decided which actions need human approval.
• [ ] Every agent step can be logged where internal audit can see it.
• [ ] You can stop the agent and undo its recent actions.
• [ ] IT, security and compliance have seen the plan, or a date is set.

A 4-week trust-first pilot

1. Week 1: pick the task and the test set. Choose one task, name its owner, and write the success metric (time per case, error rate or cost). Collect 30 to 50 past cases, including hard ones, and agree the correct result for each.
2. Week 2: read-only run. The agent proposes actions and people execute them. Score the proposals against the test set.
3. Week 3: approved actions. Turn on execution with narrow permissions, full logging, a named approver and a kill switch.
4. Week 4: live and report. Run on real work, review every exception, and give leadership a one-page report on results, incidents and controls. Then decide whether to continue.

Free 45-minute AI agent check for 5 companies

Simplico is a software engineering company based in Thailand. For over ten years we have built business systems for manufacturing, logistics and energy companies, many of them Japanese-affiliated. We will run a free 45-minute online check for 5 companies in Southeast Asia that are planning their first AI agent.

We look at one task and give you a one-page assessment: which of the six controls are missing, and whether the task is a good first candidate. There is no cost and no obligation.

To apply, contact us through simplico.net with:

• Your company, industry and rough size
• The task you want an agent to help with
• The systems it touches (ERP, MES, spreadsheets, email)
• Who approves changes to that task today

Sources

• Southeast Asia organisations to pick up on agentic AI in 2026: IDC study (Techgoondu)
• Adobe study: 57 per cent of SEA firms stuck in AI’s frozen middle (e27)
• Gartner: more than 40% of agentic AI initiatives to fail within three years (People Matters)
• New governance frameworks for agentic AI (Davis Wright Tremaine)
• Vietnam: the first law on artificial intelligence (Duane Morris)