AI Is No Longer Just About Adoption: Why Indonesian Businesses Need to Start Measuring ROI

A company buys several AI licences. Employees begin using copilots. The technology team develops an internal chatbot. Another department starts automating document work.

Six months later, the CEO asks an apparently simple question:

What did we get from it?

The answer is often harder than explaining what the company purchased.

AI is moving from experimentation into a period of proof. IDC predicts that 45% of AI-fuelled use cases across Asia-Pacific/Japan in 2026 will fail to meet ROI targets, citing unclear definitions of value and weak data foundations among the reasons.[1] This is a regional forecast, not an observed failure rate for Indonesian companies, but the underlying message matters: enthusiasm for technology does not automatically translate into business value.

A Gartner survey provides a narrower cross-check. Among 782 infrastructure and operations leaders, only 28% of AI use cases fully succeeded while meeting ROI expectations, and 20% failed outright. Data quality, availability and persistent skills gaps were among the frequently cited causes of setbacks.[2] The survey applies to I&O use cases and should not be treated as a universal success rate for enterprise AI.

For Indonesian businesses, the important question is no longer whether AI is interesting.

It is which AI initiatives deserve to be funded, retained and scaled.

Using AI is not the same as transforming a business with AI

The number of AI accounts is not a transformation metric.

Neither is the number of prompts, chatbots or pilots.

An employee may use generative AI to summarise a document and save 15 minutes. That is useful. But it does not become ROI until the organisation understands whether that saved time turns into additional productive capacity, revenue, better quality or lower cost.

AI also costs considerably more than the subscription price.

Integration, data preparation, security, training, governance, change management, monitoring and human verification all carry costs.

In practical terms:

gross productivity gain is not the same as net business value.

If an AI tool saves 1,000 hours but requires hundreds of additional hours for verification, integration and error correction, management needs to measure the benefit after those costs.

Start with the problem, not the model

Many organisations begin by asking:

“What can we do with AI?”

A more disciplined question is:

“What business problem is expensive or important enough to justify using AI?”

Imagine a company processing 30,000 invoices every month, with manual validation slowing payments.

That is measurable.

There is transaction volume.

Processing time.

Error frequency.

Labour cost.

Working-capital impact.

Technology can then be evaluated against an observable problem.

An initiative called “AI assistant for everyone” is much harder to evaluate when the company has not established what behaviour or process is expected to improve.

Measure the business before introducing AI

ROI requires a baseline.

A company cannot credibly claim that AI reduced processing time by 30% if it never measured processing time beforehand.

For customer service, useful baseline metrics may include:

response time, resolution time, escalation rate, repeat contacts, cost per interaction and customer satisfaction.

For finance:

reconciliation time, days sales outstanding, processing errors, cost and manual hours.

For operations:

downtime, defects, throughput, inventory variance and forecast error.

Without a baseline, companies end up evaluating AI through impressions:

“The team feels faster.”

That may be a useful signal.

It is not yet ROI.

The GATICORP AI Value Test

Before scaling an AI use case, we recommend five questions.

1. Problem — What exactly is being fixed?

The issue should be specific enough that management knows when it has improved.

“Improve productivity” is vague.

“Reduce average invoice-validation time from eight minutes to four” can be tested.

2. Baseline — What does the problem cost today?

The cost may be financial, time-based, operational, risk-related or tied to lost sales.

3. Value — What measurable change should AI create?

Value does not have to mean new revenue.

Lower fraud, faster closing, reduced downtime or higher conversion may all matter.

But the metric should ideally be chosen before the pilot begins.

4. Risk — What happens when the AI is wrong?

A mistake in an internal recommendation is very different from an error in credit decisions, medical advice, legal documentation or public communication.

Higher consequences require stronger human controls, documentation and monitoring.

5. Scale — Does value remain positive when the pilot becomes an enterprise system?

A 50-user pilot can appear inexpensive.

Five thousand users introduce different inference, integration, security, support and governance costs.

This framework is a GATICORP editorial decision tool, not a regulatory standard.

Productivity is not automatically ROI

Suppose an AI-enabled designer completes an early concept 40% faster.

That is a productivity gain.

For it to become financial return, management still needs to ask:

What happened to the saved time?

Did the designer complete more projects?

Did customer lead time decline?

Was outsourcing reduced?

Did revisions increase because quality deteriorated?

Did project margin improve?

When there is no direct financial outcome, the productivity benefit may still be strategically valuable. It should simply not be presented as financial ROI without evidence.

The distinction leads to better investment decisions.

Bad data can help AI accelerate the wrong thing

Gartner found that 38% of surveyed I&O leaders who experienced AI setbacks cited poor data quality or limited data availability as a direct cause.[2] Indonesia's Komdigi has similarly identified data quality and governance as major implementation challenges.[6]

AI does not repair a duplicated or inaccurate customer database merely because the underlying model is sophisticated.

It may process bad information faster.

Companies need to know:

what data are being used, who owns them, whether they are accurate, whether the intended use is appropriate, who has access, how long information is retained and how outputs will be verified.

An AI budget that focuses almost entirely on models while neglecting data governance may therefore be built on the wrong foundation.

Human in the loop is not evidence of failure

Another misleading assumption is that the best AI strategy always minimises human involvement.

It does not.

Komdigi has highlighted human-in-the-loop approaches and the need to develop people alongside AI systems.[6]

For low-risk work, full automation may be appropriate.

For financial, legal, safety or reputational decisions, human review may be part of a responsible system design.

The business question should become:

How much human review is required to make this AI output sufficiently reliable?

If an AI system needs 30 minutes of checking to save 20 minutes of manual work, the economics deserve reconsideration.

Do not allow pilots to live forever

A pilot exists to answer a question.

It should not become a permanent organisational status.

Every AI pilot should eventually lead to one of three decisions:

Scale. Modify. Stop.

Scale when value is proven and risk is manageable.

Modify when the business problem is worthwhile but the workflow, data or model needs improvement.

Stop when expected value does not justify the cost and risk.

Stopping a weak pilot is not strategic failure.

Allowing it to continue for two years because nobody wants to acknowledge the result can be considerably more expensive.

Agentic AI raises the standard for risk measurement

The next stage is more demanding.

AI systems are increasingly being designed not only to answer questions but to take actions: create tickets, alter records, send communications, initiate workflows or interact with other systems.

A Gartner survey of 469 CEOs and senior executives found that 80% expected AI to force a medium-to-high degree of change in operational capabilities.[3] That means governance must evolve from “who may use AI?” toward “what may AI do, using which data, under whose authority?”.

For agentic systems, ROI cannot be separated from the cost of control.

Management may need:

permission boundaries, approval thresholds, audit logs, rollback mechanisms, exception handling and human escalation.

The greater the autonomy, the more important these controls become.

Indonesia is entering a more formal governance phase

As of the government's latest official update on 28 July, cross-ministerial discussions of Indonesia's national AI roadmap and AI ethics framework had been completed and the instruments were awaiting presidential establishment. Komdigi has described the presidential-regulation stage as an initial step before a possible more comprehensive AI law.[4][5]

Companies do not need to wait for final regulations to develop internal governance.

They can already map:

Where is AI being used?

What data enter each system?

Which vendors are involved?

Who owns each business use case?

Where does a human retain final decision authority?

What happens when the output is wrong?

That documentation will remain useful regardless of the final regulatory architecture.

Good AI should show up in the business—not merely in the demo

Not every AI benefit needs to appear as cost savings.

AI can improve speed, customer experience, decision quality, operational resilience or the ability to develop new products.

Those benefits still need to be stated specifically.

A healthy AI strategy is not a competition to accumulate the most tools.

It is a process of allocating capital and management attention to use cases whose value is large enough to justify their cost, organisational change and risk.

The most useful question in 2026 may not be:

“How advanced is our AI?”

It may be:

“What became better after AI arrived—and can we prove it?”

When a company can answer that question with metrics, workflow evidence and clear accountability, AI stops being a technology experiment.

It starts becoming a business investment.