Anti-Corruption AI — A 4-Layer Framework for Bangladesh

Publication Date: January 2026

Classification: Policy Research Paper

Note: This paper focuses on AI as a structural anti-corruption tool, not surveillance of individuals.

---

Executive Summary

Corruption is estimated to cost Bangladesh 2–3% of GDP annually — Tk 35,000–50,000 crore — through procurement fraud, tax evasion, bribery in service delivery, and regulatory capture. Traditional anti-corruption measures (transparency requirements, audit institutions, prosecution) have had limited impact because they are reactive, slow, and subject to political interference.

AI offers a structural alternative: systems that detect corruption patterns automatically, in real time, at scale — reducing the dependence on any individual's integrity and making corruption harder to conceal.

This paper proposes a 4-Layer Anti-Corruption AI Framework for Bangladesh, with specific use cases, implementation requirements, and safeguards.

---

Why AI for Anti-Corruption

The problem with current approaches:

Bangladesh's Anti-Corruption Commission (ACC) has 750 staff to monitor 1.5 million government employees and 25 million registered businesses. The Bangladesh Supreme Court has a 4-million-case backlog. The National Board of Revenue loses an estimated Tk 80,000 crore annually to evasion and corruption.

No human organisation can monitor this at sufficient scale and speed.

What AI changes:

AI systems can monitor every government transaction, every procurement tender, and every tax declaration simultaneously — flagging anomalies for human review. They don't get tired, don't accept bribes, and don't have political connections.

The key distinction: AI should flag; humans should decide. Anti-corruption AI does not replace due process, presumption of innocence, or human judgment. It makes human oversight effective at scale.

---

The 4-Layer Framework

Layer 1: Transaction Monitoring AI

Target: Financial flows in government accounts, procurement payments, tax collections, customs revenue.

Mechanism: Real-time anomaly detection on the IBAS++ government financial management system and NBR databases, identifying:

Payments to newly registered companies immediately after contract award
Round-number transactions (common in bribery schemes)
Payments to vendors with addresses matching government officials
Split purchasing (breaking large contracts into small ones to avoid approval thresholds)
Late-quarter spending spikes inconsistent with project progress
Payments to offshore entities from ministries without international procurement mandates

Technical requirements:

IBAS++ API access for real-time transaction feeds
NBR vendor registration database integration
Company registration data (RJSC) linkage
Rules engine + ML anomaly detection model

Safeguards:

No automatic action — all flags require human review by ACC or Finance Ministry
Audit trail of all flags and dispositions
Monthly false-positive rate reporting
No individual financial monitoring without judicial authorisation

Estimated implementation: 18 months; Tk 50 crore

Expected impact: Identify 500+ suspicious transactions per year for investigation; estimated Tk 500+ crore in prevented fraud annually.

---

Layer 2: Procurement Transparency AI (PTAI)

Target: Bangladesh's government procurement process — approximately Tk 2,00,000 crore annually.

The corruption problem: Tender manipulation (rigged specifications, bid coordination, winning vendor pre-selection), phantom work (payments for work not done), inflated valuations.

PTAI components:

2A — Specification Analysis Engine:

Analyses tender specifications for language patterns that are common in rigged tenders: unusually narrow specifications matching a single vendor's product catalogue, copied text from vendor brochures, unusual technical requirements excluding otherwise qualified suppliers.

2B — Bid Pattern Detector:

Monitors bid submissions across all government tenders for:

Bid rotation (same set of companies wins in sequence, as if coordinated)
Complementary bidding (firms bidding high to let a pre-agreed winner win)
Bid suppression (qualified firms not bidding in specific tenders)
Price clustering (bids suspiciously close together)

2C — Vendor Network Mapper:

Maintains a relationship graph of vendors, directors, and government officials. Flags when:

Winning vendors share directors with government officials
Multiple bidders share registered addresses (phantom competition)
Vendor ownership chains lead to offshore entities with no disclosed Bangladesh presence

2D — Work Completion Verifier:

Uses satellite imagery and field photos (submitted via mobile app by field engineers) to verify that infrastructure projects have actually progressed as claimed before stage payments are released.

Platform: Integrated with CPTU (Central Procurement Technical Unit) eProcurement system — already operational; needs AI module addition.

Estimated implementation: 24 months; Tk 80 crore

Expected impact: Reduce procurement corruption losses by 20–30%; estimated Tk 5,000–15,000 crore saved annually.

---

Layer 3: Whistleblower Protection and Intake AI

Target: Make it safe and effective for citizens and officials to report corruption.

The problem: Bangladesh has a Whistleblower Protection Act (2011) but it is rarely used — reporters fear retaliation, and the ACC's intake process is slow, non-anonymous, and perceived as politically influenced.

AI-enhanced system:

3A — Secure Anonymous Intake Platform:

End-to-end encrypted digital platform for corruption reports. AI pre-screens submissions:

Removes identifying metadata from uploads (photos, documents)
Classifies reports by type (procurement, tax, land, service delivery)
Routes to appropriate investigating unit
Sends encrypted status updates to anonymous reporter

3B — Report Corroboration Engine:

When a report is received, AI automatically checks:

Transaction records for patterns matching the allegation
Procurement records for the named tender
Land records for described property transfers
Cross-references with prior reports about the same entity

Provides investigating officer with corroborating data within 24 hours of report receipt — dramatically reducing investigation initiation time.

3C — Retaliation Detection:

Monitors employment records, procurement awards, and land transactions involving named whistleblowers (with their consent) for patterns of retaliation — sudden dismissal, contract cancellations, regulatory actions — that may indicate suppression.

Safeguards:

Platform operated by ACC with external audit by civil society
No AI-generated conclusions used in prosecution; AI evidence is corroborating only
Reporter consent required for all monitoring
Data retention limit: 5 years

Estimated implementation: 12 months; Tk 20 crore

---

Layer 4: Integrity Risk Prediction (IRP)

Target: Proactively identify high-risk environments for corruption before it occurs, enabling preventive action.

Mechanism: IRP scores government functions (not individuals) on corruption risk using:

Budget size and discretionary authority
Absence of digital transaction records
Historical complaint rates from citizens
Audit finding patterns
Staff rotation frequency

Output: A quarterly Corruption Risk Map identifying the 100 highest-risk government functions — procurement offices, tax collection posts, land registration offices, customs checkpoints — for priority ACC audit and oversight.

Important distinction: IRP scores environments and processes, not individuals. A high IRP score means "this function needs more oversight," not "this person is corrupt."

Application:

ACC uses IRP to direct scarce audit resources
Finance Ministry uses IRP to prioritise digitisation (moving transactions online)
Anti-corruption reforms are targeted at highest-risk functions first

Estimated implementation: 24 months; Tk 30 crore

Expected impact: 40% improvement in ACC audit targeting efficiency (finding actionable cases in audited entities).

---

Cross-Layer Requirements

Data Infrastructure

All 4 layers require interconnected data access. Priority integrations:

IBAS++ (government financial management)
CPTU eProcurement
NBR tax and customs
RJSC company registration
Land records database
ACC case management system

A Government Anti-Corruption Data Hub (GADH) — a secure, access-controlled data integration platform — is the foundational requirement for all 4 layers.

Governance and Oversight

An AI Anti-Corruption Oversight Board should govern the framework:

Chair: Comptroller and Auditor General of Bangladesh
Members: ACC, NBR, Finance Division, TIB, civil society, BUET
Functions: Approve use cases, review false positive rates, investigate AI system misuse, publish annual transparency report

Legal Framework

Bangladesh needs a Fraud Detection Data Sharing Act authorising the data integrations described above under strict access controls — currently these integrations would require individual ministry consent for each use.

---

Bangladesh-Specific Calibrations

Challenge 1 — Political independence: Anti-corruption AI must be protected from direction by political authority. The oversight board must include civil society and judiciary representation with dismissal protection.

Challenge 2 — Capacity: ACC and Finance Ministry need 200+ AI-trained staff to manage and act on AI outputs. Training investment: Tk 15 crore over 3 years.

Challenge 3 — False positives: An AI system that flags 1,000 suspicious transactions per month but only 50 are real requires enormous human review capacity. Threshold calibration is critical — start conservative (high specificity) even at cost of lower sensitivity.

Challenge 4 — Counter-adaptation: Corrupt actors will adapt to avoid AI detection patterns. System must be continuously updated; this is not a one-time deployment.

---

International Precedents

Georgia: AI-assisted procurement transparency platform (e-Procurement) introduced 2010; estimated 20% savings in government procurement costs.

South Korea: Anti-corruption data integration system (ACRC) operational since 2002; credited with Korea's rise from 40th to 32nd in Transparency International CPI.

Kenya: IFMIS financial management AI introduced 2011; significant reduction in ghost worker payments (discovered 12,000 ghost workers in first year).

Ukraine: ProZorro procurement transparency platform (open data + AI monitoring); estimated €5B in savings since 2016.

Key lesson: Technology alone is insufficient. ProZorro succeeded because civil society had real-time access to procurement data. Georgia's system worked because political leadership genuinely supported it. Bangladesh's anti-corruption AI will fail if deployed as theatre while political protection for corruption continues.

---

Cost-Benefit Summary

|-------|-------------------|--------------------|-----------------------|

Return on investment: Even at the conservative estimate, every taka invested returns 30–85 taka in recovered or prevented losses.

---

What AI Cannot Do

Anti-corruption AI is not a substitute for political will. It cannot:

Prosecute cases — that requires judiciary independence
Protect whistleblowers from political harassment — that requires legal reform
Work when deployed systems are intentionally under-resourced
Function if the officials who receive AI alerts choose to ignore them
Prevent corruption enabled by legal means (regulatory capture, favourable legislation)

Bangladesh's corruption problem is not merely technical. AI addresses the detection gap; political and judicial reform must address the accountability gap.

---

Recommendations

1. Establish GADH by 2027 — the data foundation all layers require

2. Deploy Layer 1 (Transaction Monitoring) first — highest ROI, lowest political sensitivity, 18-month implementation

3. Integrate AI with eProcurement immediately — Layer 2 components 2A and 2B can begin in 2026

4. Pass Fraud Detection Data Sharing Act by 2026 — without legal authorisation, integrations cannot proceed

5. Ensure ACC independence before deploying any AI — AI anti-corruption tools in politically compromised institutions become tools for selective enforcement, not genuine anti-corruption

Research contact: research@bangladeshai.org | TIB collaboration on methodology: ongoing