ARTICLE · INVESTIGATION

Evidence No Longer Silent: Financial Investigation in the Age of AI, Deepfakes, and Instant Fund Flows

Valid predication, three-flow corroboration, and the limits of a public accountant's role in theft, embezzlement, and fraud matters

By Hendrawanto, SE., Ak., CA., CLI., ASEAN CPA, CPA., CFI. ·

Infographic on financial investigation in the age of AI: Articles 476, 486, 488, and 492 of the Indonesian Criminal Code in force on 2 January 2026 and amended by Law 1/2026, the SJI 5100–5500 investigation service standards effective 1 January 2022, and the ACFE 2026 finding that only 54% of cases are referred to law enforcement; illustration of money, document, and goods flows corroborating into a single established fact

Modern fraud moves through bank accounts, user accounts, synthetic identities, documents that look flawless, and communications that can be forged. The challenge is no longer a shortage of data, but telling authentic traces from digital noise, protecting the integrity of evidence, and resisting the urge to turn an anomaly into a verdict.

Technology changes the method, not the standard of proof. Precisely because artificial intelligence can now compose convincing emails, invoices, voices, and faces, machine speed must be answered with human care. An investigator's first question is not “who is guilty?” but “what can be proven, with what evidence, through which method, and how far may that conclusion be drawn?”

The 2026 Landscape: More Data, More Room to Manipulate

According to PPATK's Record of Strategic Achievements for 2025, Indonesia's financial intelligence unit received 43,723,386 reports during 2025, up 25.5% from 35,650,984 reports in 2024. The total value of funds analysed reached IDR 2,085.48 trillion, up 42.88% from IDR 1,459.65 trillion. Transactions analysed in connection with suspected fraud amounted to IDR 22.53 trillion, and those linked to hacking to IDR 1.03 trillion.

On 16 March 2026, INTERPOL published its global financial fraud threat assessment and estimated that AI-enabled fraud is roughly 4.5 times more profitable than traditional methods. INTERPOL does not publish how that figure was calculated, so it is better read as an indication of direction than as a computational parameter.

One finding in ACFE's Occupational Fraud 2026: A Report to the Nations — which analyses 2,402 cases from 143 countries with losses exceeding USD 3.4 billion — speaks directly to the limits of an investigator's role: only 54% of cases are referred to law enforcement, down from 69% in 2008. Investigative work more often ends as the basis for an organisation's internal decision than as a criminal case, and that calls for tighter documentation, not looser.

These figures must be read carefully. Reports received by PPATK are a reporting volume, and the value of transactions analysed is the turnover examined. Neither is proven loss. Statistics quoted without that distinction easily turn into case conclusions that were never actually tested.

Choose the Type of Engagement Before Examining Anything

Many investigations fail at the pre-engagement stage. A prospective client says “investigative audit” when what they need may only be data clarification, agreed-upon procedures, a loss computation, or expert testimony. Choosing the wrong standard sends everything that follows to the wrong address.

  • SA 200–800 — financial statement audit. To express an opinion on the fairness of financial statements as a whole; risk and materiality based, not an engagement to identify the perpetrator of a fraud.
  • SJT 4400 — agreed-upon procedures. To perform specified procedures and report factual findings without an opinion or assurance; no conclusions beyond what was requested.
  • SJI 5100–5500 — investigation services. For investigative examinations, financial loss computation, and expert testimony; predication and objectives must be clear, and evidence and report distribution are controlled.

According to IAPI, the Investigation Service Standards are effective for engagements on or after 1 January 2022 and comprise the Investigation Service Engagement Framework together with SJI 5100 (general standards), SJI 5200 (risk management), SJI 5300 (investigative examination), SJI 5400 (financial loss computation), and SJI 5500 (expert testimony). For examinations by the Audit Board, a different framework applies, discussed separately in SPKN 2026: Understanding Investigative DTT Examinations.

A Chain of Reasoning That Must Not Be Reversed

Strong conclusions come from a disciplined sequence. That sequence matters more when evidence can be machine-generated, because each stage acts as a brake on the next.

The investigative chain of reasoning: seven stages, two most often skipped
  • 1AllegationDugaanUntested; not yet a fact
  • 2PredicationPredikasiFactual basis to start and to limit scope
  • 3HypothesisHipotesisIncluding non-fraud alternatives
  • 4EvidenceBuktiRelevant, with integrity preserved
  • 5AnalysisAnalisisCorroboration and testing against criteria
  • 6FindingTemuanTested facts and their limitations
  • 7ConclusionSimpulanProportionate to the engagement objective
Stages marked in red are the ones most often skipped. The diagram simplifies the flow for educational purposes; the detailed sequence for each engagement follows the applicable standards and circumstances.

Predication is skipped when an examination begins on an accusation alone, so the scope is never genuinely bounded. Alternative hypotheses are skipped when an investigator tests a single explanation and stops at the first supporting item — the classic confirmation bias that persists even among experienced examiners.

Three Flows Must Meet Before a Fact Holds

Corroboration: testing one event from three independent sides
Money flowAccounts, journals, ultimate recipients, circular transfers, layering, transaction splitting, related-party links.
Corroborated factIt holds when all three flows reinforce one another, not when one flow merely sounds convincing
Document flowContracts, approvals, invoices, handover records, metadata, correspondence, chronology.
Goods or services flowExistence, volume, quality, time of receipt, work output, benefit, user confirmation.
In an era of synthetic documents, the document flow is the easiest to forge and therefore the weakest on its own.

Tidy invoices, consistent emails, even voice recordings of approval can now be produced without visible traces. What saves a conclusion is not a cleverer detection tool but an old principle: preserve the evidence first, compute hash values over copies of digital evidence, record the chain of custody, and make sure the traceback can be repeated by someone else.

One distinction is often missed: authentication is not attribution. An authentic document or account does not by itself prove who used it. Device, credentials, timing, context, and linking evidence must still be tested separately.

Six Logical Corrections That Often Slip Through

  1. “An investigation is just a deeper audit.” The objective, users, methods, documentation, and output all differ. An audit gives an opinion on financial statements; an investigation answers a specific allegation.
  2. “A transfer to a personal account is a loss.” A transfer is an initial fact. Ultimate use, benefit, authorisation, repayment, and causation must still be tested transaction by transaction.
  3. “If it was approved, it was used for the company.” Approval of the need, approval of the transfer, knowledge of the ultimate recipient, and benefit are four different propositions.
  4. “An analytical anomaly is evidence of fraud.” Duplicates, Benford, split payments, or outliers only set the direction of testing. An anomaly shows where to look, not what has been proven.
  5. “The largest loss figure is the best one.” An expert's task is a figure that is traceable and defensible. Object, period, benefit, comparators, recoveries, and double counting must all be controlled.
  6. “A disclaimer can cover weak procedures.” A limitation-of-liability clause does not substitute for competence, objectivity, access, appropriate evidence, and adequate investigation working papers (KKPI).

Theft, Embezzlement, and Fraud: The Transfer Differs

The three must not be used interchangeably merely because each ends in a loss. The key distinctions are where the property sat before the act, how the perpetrator obtained control, the victim's role, and when the unlawful intent to take ownership arose.

One note before the articles themselves. The mapping below is an analytical framework for building and testing hypotheses, not a legal qualification. A public accountant performing an investigation engagement does not name suspects, does not determine which article applies, and does not conclude on guilt. What the engagement produces is facts, mechanisms, and supporting evidence; legal qualification and the assessment of criminal responsibility rest entirely with investigators, prosecutors, and judges.

Three ways property moves under the Criminal Code in force since 2 January 2026
Theft
Article 476
The perpetrator takes property without right; the intent to own exists at the moment of taking. Up to 5 years' imprisonment or a category V fine.
Embezzlement
Articles 486 & 488
The property is already lawfully held, then unlawfully appropriated. Article 486: 4 years or a category IV fine; Article 488 (employment, profession, or paid custody): 5 years or a category V fine.
Fraud
Article 492
The victim hands the property over after being moved by a false name, deception, or a series of lies. Up to 4 years' imprisonment or a category V fine.
The blue dot shows where the property sits between owner and perpetrator. The diagram is an analytical aid, not a determination of which article applies.

The elucidation of Article 476 states that “taking” is not limited to a physical act but also covers functional or non-physical taking, including stealing money by transferring it. The elucidation of Article 486 sets out the distinction from theft: if the intent to own already existed when the property was taken, the act is theft; in embezzlement, that intent arises only after the property is already in the perpetrator's hands.

In a corporate setting the more commonly relevant form of embezzlement is Article 488: embezzlement by a person whose control over the property arises from an employment relationship, from their profession, or from being paid to hold it. An employee who receives funds to pay a vendor and then diverts them for personal benefit is more properly tested against Article 488 than Article 486.

It is also worth noting that a single course of conduct may touch several provisions at once. Breaking into credentials and then moving funds, for example, may also engage Article 332 on unlawful access to another person's computer or electronic system. That is the practical reason an investigator should not lock onto a single article from the outset.

Comparison of the three property offences under the Criminal Code in force since 2 January 2026
DimensionTheftEmbezzlementFraud
Initial position of propertyStill held by the owner or another party.Already in the perpetrator's lawful control.Still held by the victim.
How it movesTaken by the perpetrator without right.Lawful control converted into unlawful ownership.Handed over by the victim after persuasion or deception.
When intent arisesPresent when the property is taken.Arises after the property is in the perpetrator's control.Present when the deception is used.
Points of proofThe act of taking, another's ownership, access without right, intent to own.Basis of initial control, mandate, act of conversion, benefit, concealment.False statement, causation, the victim's handover, unlawful gain.
Principal penaltyArticle 476: 5 years or a category V fine.Article 486: 4 years or a category IV fine. Article 488: 5 years or a category V fine.Article 492: 4 years or a category V fine.

The shorthand: theft = taken; embezzlement = lawfully held, then misused; fraud = handed over by a deceived victim. In real cases a single series of transactions can raise several hypotheses at once. The investigator describes facts, mechanisms, and evidence without pre-empting which article applies.

Law Number 1 of 2023 on the Criminal Code came into force on 2 January 2026 and, on the same date, was amended by Law Number 1 of 2026 on Penalty Adjustments. That amendment touches Articles 3, 76, 100, 120, 121, 243, 251, 332, 609, and 610 among others, but does not amend Articles 476, 486, 488, or 492. For cases involving state finances, the computation framework is discussed in Computing State Financial Losses After the National Criminal Code and Standards for Computing State Losses in Constitutional Court Case 206/PUU-XXIV/2026.

Six Practice Updates for an Era of Synthetic Evidence

  1. Strengthen the acceptance gate. Test the client's motive and legal standing, the intended use of the report, the status of any legal proceedings, conflicts of interest, and the ability to obtain primary sources.
  2. Put digital access rights in the engagement letter. Access to systems, logs, metadata, devices, cloud data, confirmations, physical inspection, and a preservation notice must be agreed at the outset, not negotiated once evidence starts disappearing.
  3. Separate authentication from attribution. Test device, credentials, timing, context, and linking evidence before tying one document to one person.
  4. Make the analytics replicable. Use read-only data, control totals, transformation logs, script versioning, layered output, and re-performance by a second person.
  5. Build working papers for a sceptical reviewer. Every material sentence must be traceable to evidence, procedures, results, judgements, the responses of those concerned, and the limits of the conclusion.
  6. Test the figures against an opposing scenario. Document alternative methods, comparators, valuation dates, benefits, recoveries, tax, retention, and an anti-double-counting matrix. The computation approach is discussed further in How to Compute State Financial Losses.

Conclusion

A defensible investigation leaves a complete trail: why the work began, which allegation was tested, what evidence was obtained, how its integrity was preserved, which alternative explanations were considered, how the figures were computed, and why the conclusion stops where it does.

In the age of AI, the best investigator is not the one who spots a pattern fastest, but the one most disciplined in proving that the pattern is real, relevant, and not an artefact of the data or of their own bias. Not the courage to accuse, but the steadiness to keep suspicion and proof apart.

Official References

This article is general information and educational material. It is not a legal opinion on any particular case and not a determination of criminal qualification. Please re-check the latest regulatory texts and professional standards against the date of the conduct and the circumstances of the engagement.