All Talk, No Return? What a Decade of Annual Reports Reveals About AI in Arab Banking

Gulf banks have learned to talk fluently about artificial intelligence. The evidence says they have not yet learned to earn from it. Closing that gap is a management job, not a technology one.

There is a sentence that now appears, in one form or another, in almost every annual report published by a Gulf bank: artificial intelligence is central to our strategy. Ten years ago, it appeared in roughly one in ten reports. Today, it opens with the chief executive’s letter, anchors the strategy section, and has its own board committee.

That shift is real and it matters. A sector that hesitated over cloud computing a decade ago has decided, collectively and quickly, that machine intelligence is its future. National agendas across the Arab world have supplied both the mandate and the political cover.

But beneath the first pattern lies a second, less comfortable one. We read ten years of annual reports from banks across the six Gulf states, 692 in all, extracting every passage in which AI was discussed and setting it against each bank’s financial results. The banks that talk most about artificial intelligence are not, so far, the banks earning the most from it. Across the sector, the relationship between what an institution says about AI and what it earns is effectively zero.

This is not a case against the technology. It is a reading of where the sector actually stands, which is early, and of the decisions bank boards would do well to take now, while the field is still open.

Fluent, not yet profitable

Start with the speed of the change. In 2016, about one Gulf bank report in ten discussed AI in any substantive way. By 2025, it was close to seven in ten. Few strategic themes in modern banking have traveled that far, that fast.

Performance has not traveled with it. Put any two Gulf banks side by side and the one that writes more about AI is not reliably the more profitable. That holds for return on assets and return on equity, and it holds after adjusting for size, asset quality and the economy each bank operates in.

Follow individual banks through time and a faint positive signal does emerge: institutions tend to do slightly better in years when they describe their AI work with more confidence. The effect is small, sits at the edge of statistical significance and weakens considerably once the region’s most advanced market is removed from the sample. It is a hint, not a result, and we report it as such.

Meanwhile, the variable that dominates everything else is the oldest one in banking. The quality of the loan book predicts Gulf Bank’s profitability better than any technology measure, better than growth, and better than the macroeconomic backdrop. Whatever AI eventually does for this sector, it has not yet displaced credit discipline as the thing that separates strong banks from weak ones.

Talking about AI: share of Gulf bank annual reports discussing artificial intelligence, 2016–2025

Source: Authors’ analysis of 692 annual reports published by banks in Bahrain, Kuwait, Oman, Qatar, Saudi Arabia and the United Arab Emirates, 2016–2025.

The league table

The regional average conceals a widespread. Taking only the three most recent reporting years and counting each banking group once, the share of reports discussing AI ranges from roughly 3 in 10 to nearly 8 in 10, depending on the market.

Kuwait and Saudi Arabia lead on breadth: just over three-quarters of their banks’ reports discuss AI. Qatar follows at about seven in ten, the United Arab Emirates at just under two-thirds, and Oman and Bahrain trail at around a third.

Breadth is half the picture, and the other half reorders the table. Measure depth, meaning how much of a report is given over to AI among banks that raise it at all, and the UAE leads by a clear margin, at more than double the Kuwaiti figure, with Qatar second. Banks in these two markets mention AI in fewer reports but say considerably more when they do. Report lengths are similar across all six markets, so this is a real difference in emphasis, not an artifact of verbosity.

Two caveats travel with the table. It measures what banks publish, not what they have deployed, and that limitation runs through this whole exercise. And the markets are not composed alike: Bahrain’s listed banks include several small wholesale and investment houses whose reports are naturally light on retail technology, which pulls down their figures for reasons that have little to do with ambition.

One correspondence stands out. Qatar and the UAE, whose banks write at greatest length about AI, are also the two GCC states whose central banks have issued dedicated AI guidance for the financial sector. The Qatar Central Bank published its Artificial Intelligence Guideline in September 2024; the Central Bank of the UAE followed in February 2026 with a guidance note on the responsible adoption of AI and machine learning by licensed financial institutions. Both arrived late in, or after, the period measured here, so neither caused the disclosure pattern. But supervisory attention and bank communication are clearly moving together in these markets, and banks elsewhere in the region should expect their own regulators to follow suit.

Who is ahead: AI in Gulf Bank annual reports, 2023–2025

Source: Authors’ analysis of Gulf Bank annual reports. Depth is the median number of sentences discussing artificial intelligence per thousand sentences of report text, calculated only across reports that discuss it at all. Each banking group is counted once. The table measures disclosure, not deployment.

Why the numbers lag the narrative

Three things explain the discrepancy between the story and the statement of income, and none of them is a reason for pessimism.

Annual reports are not operating manuals. One bank may run sophisticated models and give them a paragraph. Another may describe an ambition it has not yet built. The report reflects the communications function as much as the technology function. What our findings really establish is narrow and practical: disclosure is not a usable signal.

Technology pays late. Core banking transformations take years to reach the cost line. The benefits include shorter cycle times, fewer errors, staff redeployed to higher-value work, and losses that never occurred. None of those lands neatly in a single year’s return on equity. Judging by when the sector began discussing it seriously, the Arab deployment cycle gathered pace only around the turn of the decade. Expecting a clean profitability signal by 2025 was always optimistic.

Most of what has been deployed is the easy part. Chatbots, document capture, marketing personalization, and fraud alerting are useful and cheap to buy. They are also available to every competitor at the same price, which is precisely why they do not separate one bank from another in the results. The applications that would separate them sit deeper in the institution and are considerably harder to build.

Where the money is

If the customer-facing layer is being commoditized, the question is where durable value lies. Four areas stand out for Arab banks.

Credit. A better default prediction is worth more than all efficiency gains combined, for the reason established above: in this region, asset quality drives profitability. Models that draw on transaction flows, supply-chain relationships and sectoral exposures can see deterioration months before a payment is missed. For banks with concentrated corporate books, which describe much of the Gulf, six months of early warning on a single large exposure can be worth more than the entire annual technology budget.

Financial crime. Industry estimates routinely put the false-positive rate in rule-based transaction monitoring above ninety percent, and every one of those alerts consumes an analyst’s time. Machine learning applied to the same flows substantially reduces noise while capturing patterns that static rules miss. Arab banks with heavy correspondent banking relationships and cross-border remittance volumes have more to gain here than most international peers, in both costs and regulatory standing.

The back office. Trade-finance documentation, know-your-customer refreshes, collateral verification, reconciliation, and regulatory reporting are document-heavy, rule-bound, and repetitive. They are also where a surprising share of a bank’s headcount sits. Language models are unusually good at exactly this work, and unlike customer-facing deployments, it can be automated without a single client noticing.

Risk and treasury. Scenario analysis, liquidity forecasting and stress testing have always been constrained by how many scenarios a team could compute in the time available. That constraint has been largely lifted. Banks that use the new capacity to understand their own balance sheets more deeply, rather than to produce more reports, will be better prepared for the next shock than those that do not.

One thread runs through all four. In each case, the value comes from the bank’s own data, applied to the bank’s own decisions. None of them can be bought off the shelf, which is exactly why it is defensible.

The unglamorous obstacles

The obstacles are largely shared with the rest of the industry, but two have a regional edge.

Data is the real bottleneck. Almost every bank that has attempted a serious AI program has discovered that the hard part is not the model but the state of the underlying data. Customer records are stored across several systems that disagree. Contract terms live in scanned documents that nobody has structured. Historical data was never stored with analysis in mind. Fixing these takes years, produces nothing visible while underway, and is the precondition for everything else. It is also the part most often deferred in favor of a pilot that can be demonstrated.

Talent cannot simply be bought. Banks compete for machine-learning specialists against technology firms that can offer equity, and a hired specialist who does not understand credit or compliance produces models the business will not trust. The institutions making progress are training existing risk, operations and audit staff into hybrid roles alongside their specialist hires. That route starts slowly and lasts longer because the capability is tied to people who understand the bank.

Vendor concentration is a growing exposure. The Financial Stability Board warned in November 2024 that the financial system’s dependence on a small number of model, hardware and cloud providers could become a systemic vulnerability, and that institutions remain accountable for outsourced models they do not control. For Arab banks, the point is sharper, because most regional institutions are buying rather than building. Several competitors in the same market may be running on the same underlying provider, with the same blind spots.

Governance has not kept pace, and regulators have noticed. Most boards approved an AI strategy before they had a framework for overseeing model risk. The UAE guidance note makes the expectations concrete: a documented governance framework proportionate to the institution; a maintained inventory of every AI system in use; periodic testing for embedded bias; human oversight calibrated to the consequence of the decision; plain-language disclosure to customers, in Arabic and English, where decisions have high impact; accountability that does not transfer to a vendor; and the ability to stop a model through human intervention. Banks outside the UAE would do well to read that document as a preview rather than as somebody else’s problem.

The Islamic exception

Islamic institutions account for a substantial share of Arab banking, and their position deserves separate mention for two reasons.

The first is encouraging. In the Gulf, Islamic banks are no less profitable than conventional ones. The performance gap reported in global studies simply does not appear here. Gulf Islamic institutions operate at scale, in developed markets, with the same access to infrastructure as their conventional peers, and on the measures that matter, they are not a different asset class.

The second is less so. They talk about AI considerably less. Where roughly seven in ten conventional bank reports in the Gulf now discuss it, the Islamic figure is closer to five in ten, and unlike in the wider global sample, that gap has not narrowed over the decade. Either these institutions are investing less and ceding ground in a capability well-suited to their model, or they are investing comparably and communicating poorly, paying the cost without collecting the benefit. Both are fixable. Neither fixes itself.

The irony is that Islamic banking is unusually well-suited to what this technology does best. Its contracts are document-intensive and rule-bound by design, with sequencing requirements that carry real significance. Shariah audit today relies on sampling because no human team can read every contract. Machine reading makes testing the full population feasible for the first time, allowing an institution to demonstrate compliance across an entire portfolio rather than a slice of it. For a business built on adherence to a body of rules, that is a meaningful upgrade, and one that the conventional sector has no equivalent need to build.

The governance dimension cuts the other way and deserves candor. A Shariah supervisory board is responsible, under both the AAOIFI and IFSB frameworks, for the institution’s compliance. It cannot discharge that responsibility over a process whose logic cannot be examined. Explainability is therefore not merely a prudential preference for Islamic banks; it is a governance requirement. Institutions that treat Shariah review as a final approval stage rather than a design constraint will find themselves rebuilding.

“The banks that look strongest ten years from now will not be the ones with the most ambitious narrative today.”

A short agenda for a long job

The practical to-do list is shorter than the volume of commentary suggests.

Stop accepting narrative as progress. Disclosure does not distinguish between banks that convert AI into earnings and those that do not. Boards should demand operational measures instead: transactions completed without manual intervention, false-positive rates in monitoring, time to credit decision, and hours returned to the business. These are auditable and, unlike a strategy slide, cannot be produced by the communications department.

Fund the data work as infrastructure. Structuring the institution’s information is a multi-year program that shows nothing for a long time. It should be funded like a core system, protected from the annual budget cycle and judged on completion rather than on quarterly demonstrations.

Build capability, do not only buy it. Every bank that outsources its models entirely is renting its own judgment. A mixed approach, buying commodity capability while building where the institution’s own data creates advantage, is the only version that produces something a competitor cannot replicate by signing the same contract.

Put model governance in place before scale, not after. Validation, accountability, bias testing and customer recourse are far cheaper to design in at the outset than to retrofit across a dozen live deployments. Banks that do this deploy more slowly at first and considerably faster afterward.

Pool what should be pooled. Fraud typologies, model-validation standards, disclosure conventions and regulatory reporting formats are shared problems that every institution is currently solving separately and expensively. This is exactly the work industry bodies exist to convene, and the Union of Arab Banks is well placed to lead it. A shared approach would lower the cost for every participant and reduce dependence on a handful of global vendors.

Duller reports, better banks

Arab banking is being reinvented, but it is earlier in that process than the volume of discussion implies. Ten years of annual reports show a sector that has learned to describe artificial intelligence fluently and has not yet, on the evidence, learned to convert it into returns.

Read that as timing, not verdict. The technology is genuine, the applications that matter are identifiable and the deployment cycle is young. What the evidence argues against is not investment but complacency: the assumption that having an AI strategy is the same as having a capability, and that describing the future is the same as building it.

The banks that look strongest ten years from now will not be the ones with the most ambitious narrative today. They will be the ones who spent these years fixing their data, training their people, governing their models, and measuring what was actually running. It is slower work, it makes for a duller annual report, and it is the only version that compounds.

Ahmet Faruk Aysan and Hassnian Ali

Prof. Dr. Ahmet Faruk Aysan is a former member of the MPC and the Board at the Turkish Central Bank and is a Professor and Associate Dean for Research at Hamad Bin Khalifa University; Dr. Hassnian Ali is at Hamad Bin Khalifa University in Doha.