The Decisions That Shape Everything
Every day, at every port of entry, officers make thousands of judgements in real time. Does this traveller pass or get referred? Does this cargo consignment clear or get held for inspection? Does this declaration accurately reflect what is actually being imported? Does this document authenticate or get escalated? Does this name warrant an alert or is it noise in an already noisy system?
These decisions are made in seconds, by people working under throughput pressure, with the data their systems can deliver to them at that moment. The quality of those decisions determines whether a threat is caught or missed, whether duty revenue is collected or lost, whether a trafficking victim is identified or passes unseen.
And the pressure on those decisions is growing. Air passenger volumes hit a record in 2024 and are projected to more than double by 2050. Cargo volumes are rising in parallel. The mandate these agencies are asked to fulfil is expanding — sanctions enforcement, technology export controls, forced labour supply chain checks, terrorism screening — while headcounts are flat or falling. The US Customs and Border Protection agency is already short nearly six thousand officers and expects thousands more retirements by 2028. In the UK, Border Force officials have warned that staffing pressures and budget constraints risk reducing border checks, increasing delays, and weakening operational coverage across already stretched frontline operations.
More decisions. Fewer people. Higher stakes. This is the operating environment border agencies face today — and it is not a temporary condition. It is structural and permanent.
The only lever available is making each decision smarter.
Why the Data Never Arrived in Time
For decades, transaction processing and analytics have operated as separate disciplines. Separate platforms. Separate teams. Separate timelines. It was not a design flaw. It was the architecture of its era.
Transaction systems were built to do one thing at extraordinary scale — capture, process, and record what was happening, continuously, without failure. Mainframe platforms became the foundation for this work precisely because nothing else could match their throughput, availability, and integrity at national scale. They were trusted because they earned that trust over decades of operational performance.
Analytics lived somewhere else. The compute required to run queries, aggregate records, and surface patterns across large datasets lived on different platforms, managed by different teams, on different timelines. So the data moved. Copied to a warehouse. Exported to a lake. Replicated to an environment where it could be interrogated away from the transaction system. An entire industry grew up to facilitate this movement — and the separation of transaction processing and analytical capability became not just an architecture but an assumption. The way things were done.
Border agencies built on the same foundations — transaction systems processing passengers, declarations, and watchlists on one platform, the analysis that could improve those decisions on another.
For most government functions that gap is a manageable inefficiency. A report arrives late. A trend is identified after the fact. A decision that could have been better informed was made on incomplete data. The cost is real but it is recoverable.
Border control is different. The decision that closes when a passenger clears primary cannot be revisited. The declaration that releases a consignment into the supply chain cannot be uninspected. The window for intervention exists once, briefly, and then it is gone.
When the analysis arrives after that window closes it does not improve the decision.
It describes what the decision missed.
What AI Actually Changes
What AI actually does, applied to physical border control, is specific and concrete. It changes what the officer sees before the window closes.
Pre-departure passenger screening
Governments screen passenger manifests before departure. Airlines transmit passenger data to border agencies in the hours before a flight departs. That data is scored against watchlists, risk profiles, and travel history. If the screening flags a passenger, the agency can instruct the airline not to board them — stopping the problem before it travels rather than managing it on arrival.
When that assessment reflects the current operational picture rather than the state of data at the last batch run, the passenger who should have been stopped is stopped before they travel — at the cheapest and most effective point in the entire enforcement chain.
Cargo declaration targeting
Current targeting models encode known risk patterns. They cannot detect what falls outside those patterns — the shipper whose behaviour has shifted, the declared value that has drifted below current market benchmarks, the cross-port pattern that is only visible across the full declaration population.
When declarations are assessed against current behavioural and market context before clearance, the consignment that should have been flagged is flagged before the goods clear. The window is still open.
Document authentication
Rules-based authentication detects known fraud patterns. It cannot detect techniques that postdate the last model update or inconsistencies that are only visible when a document is assessed against the full population of genuine examples.
Document authentication evaluated against patterns derived from the full population of genuine and fraudulent examples at scale improves detection of novel forgery techniques. Each fraudulent identity stopped at the border eliminates the downstream enforcement cost it would otherwise have generated across multiple agencies over time.
The Economics — Pre-departure Passenger Screening
The metric that matters: cost per enforcement action
Intervening before a passenger travels costs a fraction of managing them after they land. Most of the operational value derives from avoiding downstream detention, investigation, removal, and inter-agency enforcement activity after arrival. That differential is the entire value of better pre-departure screening.
| Today | With AI | Delta | |
|---|---|---|---|
| Pre-departure detection rate | Baseline | +15% | +15% |
| False positive referral rate | 6% | 4.8% | −1.2% |
| False positive referrals annually | 2.16M | 1.73M | −430,000 |
| Cost per enforcement action | $45,000 | $3,000 | −$42,000 |
At 50 million passengers annually, the shift from post-arrival to pre-departure intervention is worth $135M–$290M per year.
Range reflects conservative to optimistic assumptions across 10,000 simulated scenarios. Figures are illustrative and based on published border agency benchmarks. Agencies are invited to substitute their own operational parameters.
The Economics — Cargo Declaration Targeting
The metric that matters: inspection yield
Yield — findings per inspection — is what border agencies report. The majority of operational benefit derives from improving findings per inspection without proportionally increasing inspection workload. Improving yield from the same resource is operationally equivalent to expanding the inspection function without adding headcount.
| Today | With AI | Delta | |
|---|---|---|---|
| Inspection yield rate | 12% | 21% | +9% |
| Additional findings annually | 36,000 | 63,000 | +27,000 |
| Undervaluation detected pre-clearance | 24% | 54% | +30% |
| Recovery rate on detected undervaluation | 23¢/$1 | 85¢/$1 | +62¢/$1 |
At 10 million declarations annually, improving yield and shifting detection pre-clearance is worth $280M–$430M per year.
Range reflects conservative to optimistic assumptions across 10,000 simulated scenarios. Figures are illustrative and based on published border agency benchmarks. Agencies are invited to substitute their own operational parameters.
The Economics — Document Fraud Detection
The metric that matters: downstream enforcement cost per undetected identity
The authentication decision costs nothing when it fails. The downstream cost — across immigration, law enforcement, benefits, tax — is where the value of getting it right accumulates.
| Today | With AI | Delta | |
|---|---|---|---|
| Detection rate at authentication | 60% | 77% | +17% |
| Undetected fraudulent identities annually | 16,000 | 9,200 | −6,800 |
| Downstream cost per undetected identity | $28,000 | $28,000 | — |
| Total downstream cost avoided annually | — | — | $190M |
At 50 million document presentations annually, reducing undetected fraudulent identities entering the system is worth $150M–$220M per year.
Range reflects conservative to optimistic assumptions across 10,000 simulated scenarios. Figures are illustrative and based on published border agency benchmarks. Agencies are invited to substitute their own operational parameters.
AI at the Moment That Matters
For border agencies that have run their transactions on mainframe platforms and their analytics elsewhere, that separation is now over.
IBM’s Telum processor, introduced in the z16 in 2022 and enhanced with Telum II in the z17 in 2025, places AI inference directly on-chip — inside the transaction processor itself. The inference engine runs simultaneously with the transaction. Not after it. Not against a copy of the data. Against the live transaction record, at the moment it is being processed, before it moves anywhere.
IBM’s chief architect described it simply: it is built for in-transaction inference.
For the first time, operational intelligence and transaction processing no longer need to exist on separate timelines. What that means in practice is a fundamental change in what is possible at the point of decision. Every passenger manifest can be assessed against a watchlist current to the moment of scoring — not the last batch run. Every cargo declaration can be benchmarked against commodity data reflecting this morning’s market — not last week’s extract. Every document can be evaluated against patterns derived from the full population of genuine and fraudulent examples — not a static library updated last quarter.
The z17 is designed to run inference across every transaction in a high-volume environment — not a sampled subset, not a rules-filtered fraction. At the volumes a national border system handles — tens of millions of passengers, millions of declarations, millions of document checks annually — every single transaction can carry a score.
This is not a faster version of off-platform AI. It is a different architecture entirely. One where the transaction and the assessment of that transaction are, for the first time, the same moment.
For border agencies already running passenger processing, cargo declarations, or watchlist matching on IBM Z, the platform that has processed these transactions without interruption for decades now carries this capability. The data does not need to move. The infrastructure does not need to change. What changes is what the officer sees before the intervention point closes — and the guarantee that what they see reflects the world as it is, not as it was.
The Value Was Always There — For Border Control
The data that would make border control decisions better already exists in the systems that process them.
Passenger histories. Declaration records. Watchlist correlations. Behavioural patterns accumulated across years of transactions on platforms that have never failed. It was never missing. It was never inadequate. It was simply on the wrong side of a gap that the architecture of its era made unavoidable.
That gap is now closeable.
Not by moving the data. Not by replacing the platforms that border agencies have trusted for decades with something new and unproven. By running assessment where the data already lives, at the moment the decision is made, against the live transaction record rather than a copy of it.
The value is not theoretical. It sits in every manifest assessed against yesterday’s watchlist. Every declaration cleared against last week’s benchmark. Every document checked against a fraud pattern the model hasn’t seen updated since last quarter.
It is leaving through decisions that could be better. Every day. At every port of entry.
The architectural constraints that once made this separation unavoidable are now materially changing. What exists now is a conversation about what becomes possible on infrastructure border agencies already own — and what it costs to leave that value where it is.
That is the conversation now facing border agencies.
