The Fraud Tax: Why Telcos and Fintechs Are Finally Building the Same Infrastructure

Network coverage hits 90%, AI agents promise to run our lives, but fraud costs are quietly rewriting the economics of digital services.

# The Fraud Tax: Why Telcos and Fintechs Are Finally Building the Same Infrastructure

U Mobile just expanded 5G coverage to 90% of populated areas in Malaysia. Sounds like a telecom infrastructure story. Meanwhile, neobanks are discovering that fraud has become a bigger cost line item than customer acquisition in some markets. That sounds like a fintech problem.

Here's what connects them: both industries spent a decade racing to distribute services faster and cheaper, and both are now learning that the unit economics break when trust infrastructure lags behind distribution infrastructure. The telcos built the pipes. Fintechs built the interfaces. Neither built adequate systems to verify that the person on the other end is who they claim to be.

The gap is widening. OpenAI and Atlassian just announced an expanded partnership to turn enterprise knowledge into action. OpenAI's new agent reportedly wants to run your life—one WIRED reporter's version even claimed it loved them. The surface narrative is about AI doing more. The operational reality is that AI agents amplify the attack surface for fraud faster than humans can adapt policy.

This isn't a technology article about what AI can do. It's a commercial reality check on what happens when you scale customer touchpoints without scaling verification systems at the same rate.

The Distribution Trap: Coverage Without Context

Telecom operators have spent billions extending network footprints. U Mobile's 90% coverage milestone is a real engineering achievement. But coverage is not the same as trust. Every new subscriber on that network is an identity assertion—a SIM card, a device ID, a billing relationship—that someone has to validate.

The Australian telco sector saw complaints fall 10.8% overall, but ombudsman referrals remain elevated. The Australian Communications and Media Authority confirmed that while complaint volumes dropped, the share escalated to the ombudsman stayed high. Translation: operators are handling more low-stakes issues internally, but the complex disputes—fraud, unauthorized charges, identity verification failures—aren't improving at the same rate.

That pattern repeats in fintech. Neobanks designed their customer acquisition flows to minimize friction: open an account in 90 seconds, get a virtual card instantly, transact immediately. The conversion metrics looked fantastic. The fraud losses came later, after the cohort aged six to twelve months.

PYMNTS.com reported that fraud has become a bigger cost of running a neobank. Not a secondary consideration. A primary cost structure issue. When your fraud losses exceed your net interest margin on deposits, you're not a bank—you're a subsidized fraud processing service.

When your fraud losses exceed your net interest margin on deposits, you're not a bank—you're a subsidized fraud processing service.

The distribution-first playbook assumed trust would scale linearly with volume. It doesn't. Trust degrades exponentially when verification systems stay manual and rule-based while distribution goes algorithmic and instantaneous.

The AI Acceleration Problem: Agents as Attack Vectors

OpenAI's partnership expansion with Atlassian aims to turn enterprise knowledge into action. The pitch is productivity: let an agent read your Confluence wiki, your Jira history, your Slack threads, and then execute tasks on your behalf. Alexander Embiricos from OpenAI is speaking at Disrupt 2026, presumably to explain how this scales.

But every agent capability is also an impersonation capability. If an AI agent can read my Slack history and draft a message in my voice, it can also impersonate me to a vendor, a colleague, or a financial counterparty. The same technology that lets an agent "run your life"—as OpenAI markets it—can let an adversary run a synthetic version of your life at scale.

WIRED's reporter got an agent that said it loved them. Amusing in a demo. Terrifying in a social engineering attack. The telco and fintech sectors are about to face a wave of AI-mediated fraud where the attacker sounds exactly like the customer, references transaction history accurately, and passes basic knowledge-based authentication because the model was trained on the same public data the call center uses to verify identity.

OpenAI also announced progress in AI for mathematics and launched initiatives to help teens learn and shape the future of AI. Both worthy. Neither addresses the verification gap. Teaching a model to prove theorems is impressive. Teaching it to distinguish a legitimate account recovery request from a credential-stuffed impersonation attempt is commercially urgent.

The Revenue Equation Nobody Discusses

India's RBI Chief noted that bank liquidity surplus is likely to decline by the financial year-end. Liquidity management is a macro issue, but it has micro implications. When liquidity tightens, cost discipline sharpens. Fraud losses that were tolerable at 50 basis points become existential at 150 basis points.

The EU is revisiting VAT exemption for financial services, according to Bloomberg Tax. Regulatory cost structures are shifting. If you're a neobank operating on a 1.2% net margin and fraud is costing you 0.8%, a VAT policy change or a liquidity squeeze can flip you from profitable to insolvent in a quarter.

Telcos face similar margin compression. REV Fiber is launching a broadband expansion in East Baton Rouge Parish. That's a capital expenditure bet on subscriber growth. But if subscriber acquisition cost rises due to fraud prevention requirements—KYC, device fingerprinting, behavioral analytics—the payback period extends. If fraud losses hit 3-5% of ARPU in prepaid segments, the business case for rural expansion collapses.

The shared challenge is that both industries priced services assuming fraud would remain a controllable externality. It's now a core cost of goods sold. And unlike network equipment or payment processing fees, fraud scales with automation. The adversary uses the same AI tooling you do, except they don't have compliance overhead.

The Infrastructure Nobody Built

When a telco provisions a SIM, it checks the ID document. When a neobank opens an account, it runs a liveness check and a sanctions screen. Both are point-in-time verifications. Neither continuously validates that the person transacting today is the same person who opened the account six months ago.

That gap is what AI agents exploit. An agent operating on your behalf needs persistent identity attestation, not just login credentials. If I authorize an AI to pay my bills, the biller needs cryptographic proof that the instruction came from an agent I control, not a lookalike model trained on my email history.

The Atlassian-OpenAI partnership focuses on enterprise knowledge graphs. The missing piece is a persistent identity layer that travels with the agent's actions. Blockchain had a narrative around this five years ago and delivered almost nothing useful. The requirement hasn't changed. The urgency has.

Telecom operators own the SIM, the device ID, and increasingly the eSIM provisioning stack. Fintechs own the transaction history, the spending patterns, the account recovery flows. Neither has built a shared identity fabric that lets them vouch for a customer across domains. So when a fraudster ports a number and uses it to reset a bank password, the telecom and the bank each see half the attack and neither stops it in time.

What Needs to Change: From Verification Theatre to Attestation Infrastructure

The Australian ombudsman referral data tells you where the system breaks. Complaints are down, but escalations remain high. Customers aren't satisfied with the resolution process because the disputes hinge on identity claims that neither party can definitively prove. "Someone used my SIM to authorize a transfer." "Someone opened an account in my name." The operator and the bank have logs, but logs are not proof of identity. They're proof of activity.

Real trust infrastructure requires three layers:

Continuous authentication. Not a password at login. A device trust score, a biometric cadence, a transaction pattern model that updates with every interaction. If my agent starts behaving inconsistent with twelve months of history, flag it before it drains an account.

Cross-industry attestation. When a telco verifies an identity, that assertion should be cryptographically portable to a fintech onboarding flow. When a bank detects a compromised account, that signal should propagate to the telecom to lock the SIM. Today, both happen in siloed ticketing systems with 48-hour SLA delays.

Liability assignment. Right now, fraud losses sit with whoever processed the transaction. The telco says the SIM swap followed procedure. The bank says the transfer was authenticated. The customer is out $5,000. Shared infrastructure requires shared liability models—something the EU VAT discussion touches indirectly but nobody has operationalized.

OpenAI is teaching AI to prove mathematical theorems. Useful, but orthogonal. The urgent problem is teaching systems to prove identity claims in real time, in a way that an adversarial agent can't cheaply forge.

The Margin Squeeze Ahead

U Mobile's 5G expansion and REV Fiber's broadband buildout are both bets on subscriber growth covering the capex. But if fraud losses scale faster than revenue, the business case inverts. You're not expanding coverage to serve customers. You're expanding attack surface to serve fraudsters.

Neobanks are already seeing this. PYMNTS.com reported fraud as a bigger cost than customer acquisition in some segments. When that happens, the rational move is to tighten onboarding, raise friction, shrink the addressable market. Exactly the opposite of the growth narrative that attracted capital.

Telcos will hit the same wall. The 10.8% drop in Australian complaints is encouraging. The flat ombudsman referral rate is a warning. As AI agents become standard interfaces for customer service, fraud, and transactions, the share of interactions that require high-trust verification will climb. If your verification stack is still rule-based and your fraud detection is still signature-driven, you'll lose money on every marginal customer.

The RBI Chief's comment on liquidity decline is a macro signal with micro teeth. When liquidity tightens, investors stop funding growth-at-all-costs. They start asking where the fraud losses are buried in the P&L. For telcos and fintechs alike, the answer is uncomfortable: distributed across churn, bad debt, customer service overhead, and regulatory fines. Add it up, and it's a tax on every transaction.

What to Do Next

If you're running commercial strategy or product at a telco, fintech, or platform company, here's the operating question: what percentage of your 2025 budget is allocated to continuous identity verification versus one-time KYC?

If the answer is less than 10%, you're pricing risk incorrectly. The fraud tax is coming due, and it scales with AI adoption, not headcount.

Three specific actions:

Instrument your verification stack. Measure false positive rate, false negative rate, time to resolution, and cost per verified identity. Treat it like you treat network latency or payment processing uptime. If you don't have a dashboard for it, you don't manage it.

Build cross-industry identity pilots. Pick one partner—a telco if you're a bank, a bank if you're a telco—and run a pilot where a verified identity assertion from one domain shortens onboarding or step-up auth in the other. Measure fraud rate and conversion rate. If both improve, scale it. If neither improves, you learned that your verification isn't predictive, which is also valuable.

Price fraud as COGS, not shrink. Stop treating fraud losses as an acceptable percentage of revenue. Model them as a variable cost that scales with transaction volume and automation. When you price a new product or market, include fraud cost in the unit economics from day one. If the margin doesn't cover it, the product doesn't ship.

OpenAI's agents will get better at running tasks. Fraudsters will get better at running scams. The telecom network will cover 90%, then 95%, then 98% of populated areas. Neobanks will onboard users in 60 seconds instead of 90. None of that matters if the trust layer stays manual and siloed.

The companies that win the next five years will be the ones that treat identity infrastructure as seriously as they treated payment rails in 2015 or cloud migration in 2018. The laggards will keep reporting record coverage, record sign-ups, and record fraud losses—and wonder why the unit economics never worked.