Tuesday morning at a clearing office off Umbilo Road, Durban. Three containers arrived on an MSC vessel over the weekend. One is on the Durban Container Terminal holding yard with a red-routed documentary stop against it, an FCL of auto parts from Shanghai where the commercial invoice shows a value SARS Customs does not believe for that tariff heading. The certificate of origin arrived late Friday and nobody has opened it. The importer in Pinetown is phoning at seven in the morning about Thursday’s production line. The second container cleared on paper on Monday evening, but the release order is still in the EDI queue because Transnet’s gate pass had a wobble overnight. The third is at sea, due Saturday, and the pre-cleared SAD500 is drafted but nobody has pulled final weights off the Bill of Lading.

That is the rhythm. Three containers in various stages, each with its own regulator, carrier, terminal and nervous importer. Clearing agents in South Africa are not short of software. They are short of time to feed the software the right data at the right moment, and short of time to answer the fifteen WhatsApps that come in before lunch asking “any news?”

AI will not clear your container.

Any provider selling you an autonomous customs agent in the SA market either has not read the Customs and Excise Act or is hoping that you have not. What AI will do, placed carefully, is take the paper layer and the chasing layer off your clerks’ hands. The parts of the job that eat time without directly moving a box.

Where a clearing week actually bleeds time

Open a clearing desk in Isando, Durban Harbour or Montague Gardens and the admin splits into roughly the same buckets, whether the shipment moves by sea, air or road.

Every one of those is high-volume, structured, and invisible to the person paying the invoice. The importer cares whether the container clears before the demurrage clock bites.

Document capture: the clean win

A typical FCL arrives with four to twelve documents from four different sources. The forwarder emails a Bill of Lading. The supplier WhatsApps a commercial invoice with a scanned stamp. The packing list is a spreadsheet with trailing empty rows. The certificate of origin is a PDF photographed from a desk in Guangzhou at an angle. Your clerk opens each in turn and keys the invoice number, gross and net weights, HS code, Incoterm and vessel into the clearing system by hand.

An hour per full shipment is the quiet norm. On fifty containers a month, that is fifty hours before anyone has looked at classification.

A capture layer with AI reads the pack as it lands, matches invoice to packing list to Bill of Lading, and flags the discrepancies the clerk would otherwise have caught two days later by way of a SARS stop: invoice weight against packing list weight, container number against Bill of Lading, consignee name against the importer code. It pulls the fields into the clearing system draft and leaves the clerk with a one-screen review instead of three browser tabs. In my experience the real gain is less keystrokes than catching mismatches early. A Chinese supplier billing in kilograms against a packing list in pounds is a cheap error on Tuesday and an expensive one at DCT on Friday.

What the capture layer does not do: decide. The clerk still eyeballs the pack before anything is lodged.

SAD500, tariff codes, and the lines AI should not cross alone

The SAD500 is the customs declaration. The HS code on it decides the duty rate, whether the shipment attracts anti-dumping, whether a rebate item applies, whether the import needs an ITAC permit, and whether SARS treats the valuation as credible. A wrong code is a penalty at best, a voluntary disclosure letter at worst. The licensed agent takes personal professional responsibility for the figure on the declaration, under the Customs and Excise Act and the terms of the licence itself.

An AI model can read a product description, say “centrifugal pump, cast iron, 15kW, three-phase”, suggest the HS code, cross-check against previous shipments the importer has lodged for the same product and flag where the current description reads differently to what has gone before. It can read the invoice value and compare it against a running band for that HS code and importer, flagging values that fall outside the window SARS has accepted without comment. On a repeat-importer book of several hundred line items a month, that catching layer earns its keep the first time it stops a mis-declaration from leaving the office.

Signing the SAD500 is not the model’s job. The licensed agent is personally accountable for the content of the declaration, and a bot that quietly reclassified a four-million-rand shipment of pumps from tariff heading 8413.70 into a lower-rate code is not something the Commissioner forgives on the day. The AI drafts and flags. The clearing agent signs. The audit trail shows a named human on every lodgement.

Calls that remain the clearing agent’s regardless of model suggestions: anti-dumping classification where the margin against a different HS code is thin; Rules of Origin under SACU or SADC-EU EPA; voluntary disclosures; written tariff determinations; and rebate items under Schedule 3 or 4.

Watching the portals so your clerks do not have to

A shipment moves through roughly eight systems between booking and gate-out: the forwarder’s platform, the carrier’s track-and-trace, the terminal operating system (Navis N4 at the big SA container terminals), the SARS EDI gateway, Transnet’s gate pass and ePortal, the empty depot booking, the trucking job board, and the deferment account balance. A clerk’s afternoon is clicking through those to see where each container has moved since yesterday.

AI handles this well because it is deeply boring work.

A watcher process polls each portal on the frequency the carrier publishes, reads the state (on vessel, discharged, on holding, released, gate-out), reconciles against the SAD500 lodgement, and surfaces only the ones that moved. More usefully, the ones that should have moved and did not. A container that discharged on Monday and is still in holding on Thursday with no stop against it is the alert the clerk actually needs. The twenty-three others that moved through on cue are not.

The same layer catches the surety running close to the deferment limit before SARS blocks the next lodgement, which is the sort of thing nobody tracks until the Friday-afternoon lodgement bounces and three clients phone within the hour.

Client comms: the parts importers actually chase you for

Importers want three things from their clearing agent, and they want them answered on WhatsApp without having to ask twice. Where is my container. When will it clear. What will I pay in duties and VAT. The first and third are mechanical. The second is where judgement earns its fee.

A compliant comms layer sends the importer a message when the vessel is confirmed at pilot station, another when the container is discharged, another when the SAD500 is lodged, another when the release lands, another when the truck is booked and the gate pass is pulled. The messages trigger off real portal state changes, not off a human remembering to update anyone. Routine replies, like “what’s my duty bill?”, go to a canned answer with the pro-forma figures off the draft SAD500. Harder replies route straight to a named clerk. In my experience, importers stop phoning twice a day in the first week this is live.

A second thing worth saying out loud. The clearing agent should tell the importer in writing when a shipment is going to be late. SA port congestion at Durban and Cape Town in the first quarter of 2026 did not disappear. The best use of AI there is the honest early message. Something like: “Your container discharged Tuesday morning, but DCT North is working a reduced labour pattern this week and we are forecasting a 48-hour hold against the usual turnaround.” Not a reassuring one.

And one restraint. The bot does not quote delivery dates and does not confirm releases that have not come back from SARS. A message that said “cleared” before the EDI response landed is a message you spend the afternoon walking back.

What AI handles poorly in this space

Honesty about limits is where expert advice earns its weight. There are parts of the clearing cycle where the obvious AI approach is wrong, and buying the demo without thinking about them has cost SA offices real money this year.

Tariff determinations in grey cases. If the HS code is genuinely contested, say on a composite product that could sit under two chapters, the next step is a written determination application to SARS, not a confident model answer. The agent needs that letter on file for the next audit anyway.

Rules of Origin on preferential agreements. The arithmetic of regional content under SADC-EU EPA, SACU-EFTA or AGOA is not what AI handles reliably without a human unpicking the bill of materials. We have seen a model assert origin on a shipment that would not have survived five minutes of a verification visit.

Voice calls to SARS Customs. The agent who argues a stop with the officer on the ground is building a relationship, not reading a script. The person who holds the licence holds the call.

Where to start: narrow and provable

Offices that try to automate the whole clearing cycle on day one end up with a system that does seven things poorly, and clerks are back in the old spreadsheets by week six. The cost of that attempt sits in the P&L for the rest of the year.

The right first project for most SA clearing offices is two steps, in order. First: document capture into a draft SAD500. One file lands in your mailbox, the AI reads the pack, the clerk reviews a single screen, the draft is lodged under the clerk’s name. Second: the portal-watching layer that pings on real state changes and feeds the client-comms stream. Both are high-volume. Both leave the regulated decision, the signature on the declaration, with the licensed person. Both pay back inside a quarter in the clerk-hours they free.

Classification suggestions, deferment-balance watching and WhatsApp capture of release-day instructions follow once those two are running cleanly. In the Durban and Isando offices that have built it in this order, the pattern is consistent. Container cycle times inside the office drop by a working day. Clerks spend less time clicking through portals and more time arguing with SARS on the files that need an argument. Importers who used to phone twice a day phone twice a week, because the system has told them what they wanted to know before they thought to ask.