This week, the region's tax chiefs did something the Caribbean rarely does in public: they put a number on their own failure and named the tool they intend to fix it with. The Caribbean Organisation of Tax Administrators, known as COTA, held its 27th General Assembly and Technical Conference in Georgetown, Guyana from July 27 to 31, hosted by the Guyana Revenue Authority under the theme Future-Ready CARICOM Tax Administration: Smart, Data-Driven and AI-Enabled for Sustainable Revenue. That is not a subtle title. It is a regional tax body telling every CARICOM member state, in the clearest language its own conference programme allows, that artificial intelligence is no longer an optional upgrade to how the Caribbean collects money it is already owed.
Artificial intelligence can close a meaningful share of the Caribbean's tax gap within a few years, not decades, using tools that are already proven in economies with less digital infrastructure than Jamaica, Trinidad and Tobago, or Barbados. Risk-based audit selection, automated invoice matching, and AI-assisted customs targeting have each been tested in real revenue administrations with real, measurable results. The Caribbean's obstacle is not that the technology is unready. It is that the data these tools need has not yet been built to a standard AI can use, and that gap has to close before the AI does its job.
The scale of what is at stake is not abstract. It runs through every under-resourced hospital, every delayed road repair, and every government that has had to borrow externally on unfavourable terms rather than collect what its own laws already say it is owed.
A Regional First: COTA's Georgetown Agenda
Regional tax conferences are not usually news. COTA has met annually for decades, rotating hosts among CARICOM member states, producing communiques that rarely reach beyond the finance ministries that send delegates. This year is different in one specific way: the theme itself made AI and data analytics the organising idea for the entire General Assembly, rather than a single afternoon panel buried in a longer programme about compliance and treaty coordination.
Alongside the Technical Conference, Georgetown also hosted a High-Level Regional Meeting on Tax and Development, run jointly with the Organisation for Economic Co-operation and Development. That pairing matters. The OECD has spent the past decade building the international architecture for automatic information exchange, base erosion rules, and digital tax cooperation that mainly serves larger economies with the technical capacity to use it. A CARICOM meeting that puts AI-enabled revenue administration on the same agenda as OECD cooperation is a signal that the region's tax administrators want a seat at that table, not just a subscription to whatever standard gets decided elsewhere.
Guyana is a fitting host for this conversation. Its oil-driven growth has turned the Guyana Revenue Authority into one of the fastest-scaling tax and customs operations in the hemisphere almost overnight, managing a tax base that barely resembles the one it administered a decade ago. An administration growing that quickly cannot rely on the manual, experience-based case selection that smaller, slower-growing Caribbean tax offices have used for generations. Guyana needs modern risk-scoring tools now, in a way that gives its neighbours a live test case rather than a theoretical one.
The Size of the Revenue the Caribbean Is Leaving on the Table
The Economic Commission for Latin America and the Caribbean, the United Nations body whose full mandate covers the region COTA represents, estimates that tax evasion across Latin America and the Caribbean costs $340 billion a year, equal to 6.7 percent of regional GDP. ECLAC's analysis further finds that corporate and personal income tax evasion exceeds 60 percent in some countries in the region, a figure driven by high informality, weak institutional capacity, and what ECLAC describes as low fiscal consciousness among taxpayers who see limited direct benefit from the taxes they might otherwise pay.
The Inter-American Development Bank has produced country-level estimates that make the Caribbean's specific exposure clearer. Using both the electricity consumption method and the currency demand method, IADB researchers estimate the informal economy at 35 to 44 percent of GDP in Jamaica, 29 to 33 percent in Guyana, and 26 to 33 percent in Trinidad and Tobago. Nearly a third to almost half of real economic activity in these three economies happens largely outside the reach of formal payroll records, registered invoicing, and the audit trails that conventional tax administration depends on to identify who owes what.
These two data points describe the same underlying problem from different angles. ECLAC measures the money that should have reached the treasury and did not. The IADB measures the share of the economy operating in a space where that money was never visible to begin with. Between them, they explain why Caribbean finance ministries routinely face a familiar bind: raising consumption taxes such as GCT and VAT on the households and businesses that already comply, because the portion of the economy that does not comply is, under current administrative methods, too costly and too slow to chase one taxpayer at a time.
What AI Tax Administration Actually Looks Like
Strip away the phrase artificial intelligence and what COTA's Georgetown theme is describing is a specific, well-documented set of tools that tax administrations elsewhere already run in production.
Risk-based audit selection is the core of it. Instead of a tax officer choosing which returns to audit based on rotation, referral, or personal judgement, a machine learning model scores every filed return against dozens of variables: declared income relative to visible spending indicators, sector-specific benchmarks, prior audit outcomes for similar taxpayers, and inconsistencies between related filings. The model does not replace the auditor. It replaces the guesswork in deciding which of thousands of returns that auditor should look at first, out of a caseload no human team could review in full.
Invoice matching addresses a specific and costly form of fraud in VAT and GCT systems: missing trader fraud, where a business claims input tax credit for a supplier invoice that the supplier never actually reported or remitted. Manually cross-referencing every invoice against every supplier's filing is impractical at scale. An automated matching system flags the mismatch the moment both sides of a transaction are filed, closing a gap that otherwise persists for months or years before a manual audit happens to catch it, if one ever does.
Customs risk targeting applies the same logic to trade. AI systems compare a shipment's declared value, product classification, and country of origin against historical trade data, including, where available, corresponding export data from the partner country. A container declared at a suspiciously low value relative to comparable global shipments gets flagged for inspection. One declared consistent with market pricing does not, freeing scarce customs officers to focus physical inspections on the cases most likely to involve under-invoicing or misclassification rather than inspecting by quota or random sample.
Proof It Works: Senegal, Tanzania, and the Evidence Base
None of this is speculative for economies at the Caribbean's scale of development. A World Bank study of Senegal's corporate tax administration, led by economist Pierre Bachas, found that algorithmic risk-based audit selection meaningfully outperformed the discretionary methods Senegalese tax officers had relied on previously, identifying non-compliant firms the manual process routinely missed. In Tanzania, a risk-based case selection system introduced to strengthen firm-level tax examinations increased the corrected taxable income identified through resulting audits by approximately 20 percent in its first year of operation, according to research documented by the World Bank.
Neither Senegal nor Tanzania has more advanced digital tax infrastructure than Jamaica, Trinidad and Tobago, or Barbados currently operate. Both achieved these results with the kind of institutional constraints, limited audit staff, uneven record-keeping, and constrained IT budgets, that will sound familiar to any Caribbean tax commissioner. That is precisely why these two case studies matter more to CARICOM's conversation than examples from wealthier tax administrations in North America or Western Europe: they demonstrate that the return on AI investment in tax administration does not require first-world digital infrastructure to materialise. It requires a specific, achievable set of data foundations, which the IMF has been explicit about in its own guidance.
The IMF's 2024 technical note on artificial intelligence in tax and customs administration, and its companion guidance on analytics for compliance risk management, both make the same central point: the constraint on results is rarely the sophistication of the algorithm. It is the quality and structure of the underlying data. A model trained on inconsistent, incomplete, or poorly digitised filing records will simply learn and automate the blind spots already built into those records, rather than correcting them. This is the single most important caveat for any CARICOM administration reading Georgetown's theme and assuming AI alone is the fix.
Why This Is a Distinctly Caribbean Case
Three features of Caribbean economies make AI-enabled tax administration a sharper opportunity here than the generic global case for it.
First, Caribbean tax administrations are small by any regional comparison, and small administrations feel the benefit of automated case selection more acutely than large ones. A country the size of Jamaica or Trinidad and Tobago cannot staff an audit team large enough to manually review every filing anomaly across its economy. A risk-scoring system that concentrates a modest audit staff's attention on the highest-probability cases is not a marginal efficiency gain for an administration this size. It is close to the only realistic way to raise audit coverage without a proportional increase in headcount these governments cannot afford.
Second, customs and import duties carry disproportionate weight in Caribbean government revenue compared with larger, more industrialised economies that rely more heavily on domestic income tax. Small, import-dependent island economies collect a meaningfully larger share of government revenue at the border than economies with large manufacturing bases. That structural fact means AI-assisted customs risk targeting, the specific use case the IMF flags as one of the more mature and lower-risk entry points for administrations new to AI, may deliver the fastest, most defensible early win for CARICOM governments deciding where to start.
Third, the informal economy figures from Jamaica, Guyana, and Trinidad and Tobago describe a Caribbean-specific structural challenge that generic AI tax tools, built and tuned in economies with smaller informal sectors, will not automatically solve. A risk model trained mainly on formal-sector filing data will be good at catching formal-sector anomalies and comparatively blind to the 26 to 44 percent of economic activity that never generates a formal filing in the first place. Closing that gap requires deliberately incorporating non-traditional data, utility consumption records, point-of-sale and mobile money transaction data, and property and asset registries, the same categories of proxy data the IADB itself used to estimate informality in the first place. Caribbean administrations that want AI to reach the informal economy, not just audit the formal one more efficiently, need to build that data integration in from the start rather than treating it as a later phase.
The Risk Nobody at Georgetown Should Wave Away
An honest account of this opportunity has to include its sharpest risk, because getting it wrong would cost CARICOM governments public trust they cannot easily rebuild.
The Caribbean's informal economy is not primarily corporations hiding profit offshore. It is small vendors, market traders, gig workers, and household businesses who operate informally because formal registration is costly, slow, or simply disconnected from any visible public benefit. An AI system that is easiest to build first, because formal-sector data is cleanest and most available, risks training its sharpest enforcement attention on exactly this group: the taxpayers least able to absorb an unexpected assessment and least responsible for the $340 billion regional evasion figure ECLAC documents. If the first visible effect of AI-enabled tax administration in a CARICOM country is a wave of automated assessments against small traders while high-value corporate and cross-border evasion remains comparatively untouched, the technology will be experienced, correctly, as unfair.
Every credible deployment guide for this technology, including the IMF's own technical guidance, treats governance safeguards as inseparable from the technical rollout, not an optional add-on. That means publishing the general criteria behind risk scoring so it is understood, not feared. It means a transparent appeals process so a flagged taxpayer can contest a finding before an assessment becomes final, not after. And it means deliberately weighting early deployment toward high-value corporate, cross-border, and customs cases, where Senegal's and Tanzania's evidence base is strongest, rather than toward the informal, low-income sellers who are administratively easiest to model but socially costliest to target first.
What Every CARICOM Government Should Do Starting This Week
Georgetown's conference produces communiques. What determines whether this week changes anything is what individual finance ministries do in the months that follow it.
The sequence that Senegal, Tanzania, and the IMF's own guidance point toward is consistent, and it starts before any model is built. First, digitise and standardise the underlying filing and invoicing data, because a machine learning system is only as reliable as the records it learns from, and every study on this technology in developing-country tax administrations names data quality, not algorithm sophistication, as the binding constraint. Second, pick one narrow, well-evidenced use case to prove the model on rather than attempting a single sweeping national system: VAT or GCT invoice matching for missing trader fraud, or customs risk targeting at the point of import, both of which have documented results in comparable economies and clear, auditable outcomes a finance minister can defend publicly. Third, treat COTA itself as shared regional infrastructure rather than sixteen separate national projects. A risk-scoring platform designed once, with the data-sharing agreements and technical standards to let each member state adapt it to local law, is a fraction of the cost of Jamaica, Trinidad and Tobago, Barbados, and Guyana each building bespoke systems independently, and it is exactly the kind of coordination a regional tax body like COTA exists to provide.
Funding this is not the constraint it might appear to be. The High-Level Regional Meeting on Tax and Development held alongside COTA's General Assembly, run jointly with the OECD, signals active international engagement with exactly this agenda. The Inter-American Development Bank, the Caribbean Development Bank, and IMF Technical Assistance programmes have each funded Caribbean tax modernisation projects before and are natural partners for AI-specific pilots. Set against ECLAC's $340 billion regional evasion estimate and the IADB's informal economy figures for Jamaica, Guyana, and Trinidad and Tobago, the cost of a properly sequenced data and risk-scoring pilot is a rounding error relative to the revenue currently going uncollected every single year.
The theme chosen for Georgetown this week was not an accident. It was a regional tax body telling itself, in language plain enough for any finance minister to act on, that the tools to close a meaningful part of the Caribbean's tax gap already exist and are already proven at this scale of economy. The evidence from Senegal and Tanzania says the technology works. The IMF's own guidance says the data has to be right first. The choice in front of every CARICOM finance ministry now is whether Georgetown's theme becomes the year the region built that foundation, or another conference communique that reads well and changes nothing.
Frequently Asked Questions
What is COTA and why does its 2026 Georgetown meeting matter?
COTA is the Caribbean Organisation of Tax Administrators, the regional body that brings together tax officials from CARICOM member states. Its 27th General Assembly and Technical Conference ran in Georgetown, Guyana from July 27 to 31, 2026, hosted by the Guyana Revenue Authority under the theme Future-Ready CARICOM Tax Administration: Smart, Data-Driven and AI-Enabled for Sustainable Revenue. This is the first time the region's tax chiefs have organised an entire General Assembly around artificial intelligence and data analytics as the operating model for the next decade of Caribbean revenue collection, rather than treating AI as a side topic in a broader agenda.
How much tax revenue does the Caribbean actually lose to evasion and informality?
The Economic Commission for Latin America and the Caribbean estimates the wider region loses $340 billion a year to tax evasion, equal to 6.7 percent of regional GDP, with corporate and personal income tax evasion exceeding 60 percent in some countries. Separately, the Inter-American Development Bank estimates the informal economy at 35 to 44 percent of GDP in Jamaica, 29 to 33 percent in Guyana, and 26 to 33 percent in Trinidad and Tobago. Every dollar earned inside that informal share sits largely outside the reach of national tax systems built around registered payroll and formal invoicing.
How does AI actually improve tax collection in practice?
AI improves tax collection mainly through risk scoring: machine learning models rank taxpayers and transactions by the probability of non-compliance, using variables such as declared income against spending patterns, invoice mismatches, sector benchmarks, and historical audit outcomes. This lets a small tax administration with limited audit staff focus its scarce human attention on the cases most likely to yield unpaid revenue instead of selecting audits at random or by rotation. The same technique also powers automated invoice matching for VAT and GCT, which flags missing trader fraud and duplicate claims that manual review would rarely catch in time.
Has AI-based tax audit selection actually worked in a developing country like those in the Caribbean?
Yes, in comparable middle-income settings. A World Bank study of Senegal's corporate tax administration found that algorithmic risk-based audit selection meaningfully outperformed the discretionary case selection tax officers had used previously. In Tanzania, a risk-based case selection system increased the corrected taxable income identified through audits by approximately 20 percent in its first year of operation. Neither country has more advanced digital infrastructure than Jamaica, Trinidad and Tobago, or Barbados, which is precisely why these results matter for CARICOM administrations weighing whether the technology is proven enough to invest in now.
Can AI help Caribbean customs authorities as well as domestic tax agencies?
Yes, and for small Caribbean economies this may be the higher-value application. Import duties make up a larger share of government income for small island states than for larger economies with bigger domestic income tax bases. AI-assisted customs risk targeting compares declared invoice values, product classifications, and shipment histories against trade data and flags under-invoicing and misclassification for physical inspection, rather than inspecting every container or relying purely on manual profiling. The IMF's 2024 technical guidance on AI in tax and customs administration identifies this as one of the more mature, lower-risk starting points for administrations with limited AI experience.
What are the risks of using AI in tax administration in small Caribbean states?
The primary documented risk is the underlying data, not the AI model itself. Research on machine learning adoption in developing-country tax administrations consistently identifies inconsistent, incomplete, or poorly structured data as the main constraint on results, ahead of algorithm choice or computing power. Beyond data quality, there are governance risks: AI systems trained mainly on formal-sector data can under-flag informal-sector evasion while over-targeting small registered businesses that are easiest to model, and any risk-scoring system needs a transparent appeals process so a flagged taxpayer can challenge the finding rather than simply accepting an automated audit.
What should CARICOM governments do before deploying AI tax systems?
Three things come before any AI model. First, digitise and standardise tax filing and invoicing data, because a risk-scoring system trained on incomplete records will simply automate old blind spots. Second, start with narrow, well-evidenced use cases such as VAT or GCT invoice matching and customs risk targeting, where Senegal, Tanzania, and IMF guidance already provide a template, rather than attempting a single sweeping national AI system. Third, use COTA itself as shared infrastructure: a regional risk-scoring platform built once and adapted by each member state is far more affordable for small administrations than each Caribbean territory building its own system independently.
Will AI tax enforcement hurt small businesses and informal workers in the Caribbean?
It does not have to, and getting this wrong would undermine trust in the entire initiative. The informal economy in Jamaica, Guyana, and Trinidad and Tobago is dominated by low-income vendors, small traders, and gig workers, not the corporations and high-net-worth individuals responsible for the bulk of the region's revenue leakage. Responsible deployment should weight risk scores toward high-value corporate and cross-border transactions first, pair any expanded formal-sector enforcement with simplified registration and lower compliance costs for small operators, and publish the general criteria behind risk scoring so the system is understood as fairer than manual, opaque enforcement rather than more punitive.
Who is funding AI adoption in Caribbean tax administrations?
COTA's Georgetown conference included a High-Level Regional Meeting on Tax and Development held jointly with the OECD, reflecting the international institutions already engaged with Caribbean revenue reform. Beyond OECD technical cooperation, the Inter-American Development Bank, the Caribbean Development Bank, and IMF Technical Assistance programmes have all funded tax administration modernisation projects in the region previously and are natural funding partners for AI-specific pilots, since a modest data-infrastructure and risk-scoring project is inexpensive next to the hundreds of millions in evasion and informality each of these governments currently absorbs every year.