At the border, the assessment is fast. After the border, it has to hold.
Customs assessments operate on a different rhythm from domestic taxes. Decisions are made in hours, sometimes minutes, with the importer's cargo physically present and the trade clock running. The same assessments must, months later, defend themselves at the Tax Appeals Tribunal under the same Section 56 burden. AI in customs has to operate inside both timeframes at once.
The KRA Customs and Border Control Department processes a meaningful share of the authority's total collections through the operations at Mombasa and the inland container depot at Embakasi. The Nigeria Customs Service, the Tanzania Revenue Authority's Customs and Excise Department, and the South African Revenue Service Customs and Excise Division operate at parallel scale. The economics of every African economy with a deep-water port run through these departments. The compliance question that runs underneath them is the same.
The customs assessment has a structural feature that distinguishes it from domestic tax assessments. It must be made fast. A container sitting at the port is paying demurrage. A truck waiting at a border crossing is paying drivers and fuel. The importer is paying the cost of capital on inventory they cannot yet sell. The customs system that takes too long to make a clearance decision destroys trade. The customs system that makes clearance decisions too fast lets through under-declared, mis-classified and prohibited goods.
The classical response has been risk-based clearance — green, yellow, red channels. Green-channel consignments clear automatically. Yellow-channel consignments are subject to documentary review. Red-channel consignments are physically inspected. The classification is driven by a risk model built on the importer's history, the goods type, the country of origin, and the routing. Most African customs administrations now use some version of this. The model assigns channels. The officers act on the channel assignments. The audit trail is the channel decision.
AI extends this. A modern customs risk model can use document parsing — reading the bill of lading, the invoice, the packing list, the certificate of origin — alongside the structured customs declaration to identify discrepancies a rule-based system would miss. It can use cargo scanner imagery, integrated into KRA's plan from June 2026, to flag containers whose contents do not match the declaration. It can use trade-pattern analysis to identify mis-classification — declaring HS code 7308.90 (steel structures) when the goods are HS 7228.50 (steel bars) because the duty rate is lower.
A customs assessment that was made in twenty minutes at the port must still hold up in twenty months at the Tribunal. The reasoning the officer did has to be the reasoning the file shows.
The defensibility problem at customs is sharper than in domestic taxes because the assessment is so often made under speed pressure. The officer at the wharf, reviewing the documents against the scanner image and the AI flag, does not have the time to compose a full reasoning memo before signing the entry. The reasoning is in their head; the file gets the channel assignment and the released-or-detained decision. Months later, when the importer's representative appeals to TAT, the file is what the Commissioner's legal team has to defend with. The reasoning that was real at the wharf is gone.
The other defensibility problem is trade-based money laundering — TBML. The customs assessment is also, increasingly, an AML control. Trade is one of the primary vehicles for moving value across borders outside the banking system. The FATF's guidance on TBML, the Wolfsberg Group's typologies, and African experience under the ESAAMLG and GIABA evaluations all point to the same patterns: over-invoicing, under-invoicing, phantom shipments, multiple invoicing on a single shipment, and misuse of trade finance instruments. The customs officer is now asked to spot these alongside the duty-related risks, with no more time per consignment.
Where each sits.
Akki is the substrate the customs system extends to. The integrated customs management system, the bill of lading data from the shipping lines, the cargo scanner output, the importer's history, the country-of-origin risk feeds, the AML watchlists, and the third-party data from the bank reporting feeds all converge on Akki. The customs officer is not switching between systems. The picture the AI presents is one structured picture.
Solva does the reasoning that produces the channel assignment and the assessment basis. When a consignment arrives, Solva structures the analysis the way the file must look at TAT, even though the decision will be made in minutes. It restates the question — is the declaration consistent with the documents, the scanner imagery, the importer's history, and the trade patterns the system recognises. It surfaces what is known about each — the declared HS code, the document discrepancies, the scanner anomalies, the importer's prior compliance, the country and counterparty risk. It triangulates and produces a recommended channel and, where applicable, the basis for an assessment of additional duty or for detention. The reasoning is generated and captured in the file before the officer signs. Solva's refusal behaviour does specific work at customs. When the data does not support a detention or an additional duty assessment, Solva refuses to recommend one. This matters operationally because customs is the part of the authority most exposed to the political and economic cost of holding goods that should have cleared. A consignment held for two weeks that turns out to have been correctly declared damages the importer, the consumer, and the authority's relationship with the trade community. Solva's refusals reduce the false-positive cost of the risk model.
SyniSense governs the data layer for customs, which has its own twist. Customs data is shared internationally — through the World Customs Organization's frameworks, through bilateral exchange-of-information agreements, through the recently-signed FIRS-DGFiP MoU and equivalents at other authorities. The sensitivity is real. Importer data, beneficial ownership data, shipper data — all are sensitive under the data protection framework and under bank secrecy where banking data is part of the customs picture. SyniSense anonymises identifying fields when external models do the reasoning, logs the purpose of each cross-reference, and produces the audit receipt that demonstrates compliance with the bilateral and multilateral frameworks the authority operates under.
The clearance process changes. The customs officer is not making a channel decision and then composing a reasoning memo. The reasoning is generated alongside the decision, captured in the file, and available the moment the decision is made. The clearance time is unchanged. The defensibility downstream is transformed.
The TBML detection capability changes. The customs department is no longer running TBML detection as a separate workstream by a separate team. The AML reasoning is integrated into the standard clearance picture. Trade-based laundering cases are flagged when the patterns are present in the live consignment, not reconstructed months later from historical data.
The TAT customs docket changes. Cases that escalate go with files that show the reasoning end-to-end. The win rate rises. Cases that should not have been pursued — where the file shows the original decision was thinly supported — are conceded earlier, often through ADR.
And the trade community's relationship with customs changes. The importer whose consignment is detained receives, from the system, the basis. The basis is specific. The compliant importer learns that the system can tell the difference between their consignment and the one being held next to it. The trade-facilitation argument — that customs is meant to enable trade, not obstruct it — is met by a system that obstructs only when it can show why.