Mashreq runs IBM Bob across 4,000 agent requests a month in banking pilot

What the pilot covers

IBM Bob is positioned as an enterprise platform for software engineering rather than an assistant that answers questions inside a code editor. According to the pilot description, it works across the software development lifecycle: solution design, code review, deployment and production troubleshooting. At Mashreq the scope is integration engineering, the layer that connects core banking, payment rails and internal enterprise systems.

That scope explains why the numbers are measured in requests and hours rather than in lines of code. Integration work is mostly reading specifications, mapping fields between systems, checking edge cases and then repairing whatever breaks in production at two in the morning. The project was reported by Zawya on 12 October 2026, and the figures in that report are attributed to Mashreq and IBM; no independent measurement of the pilot has been published. It is one of the first publicly named deployments of the platform inside a bank.

The numbers, divided

4,000 agent requests a month sounds like heavy traffic until you split it up. Across 21 business days it comes to roughly 190 requests a day. Spread over an eight-hour working day, that is one request every two and a half minutes. For a pilot, that is steady background load rather than a burst, and it says more about how the gateway and the surrounding controls are provisioned than about how much engineering work the agents are finishing.

The time figure deserves the same treatment. If a task took two hours and the saving is 60%, the result is 48 minutes. At three hours and 70%, it is 54 minutes. Both are useful gains on work that is repetitive and easy to get wrong. Neither turns a task measured in hours into one measured in minutes, unless the original job was short to begin with. The line about tasks being completed "within minutes" and the 60% to 70% range describe different kinds of work, and the range is the safer planning number.

There is a third figure attached to IBM Bob: IBM's published average productivity gain of 45% across enterprise adopters of the platform. That number does not come from Mashreq, and neither IBM nor Mashreq has published a measured output figure for this deployment. The 60% to 70% time reduction and the 4,000 monthly agent requests should therefore be read as pilot-phase figures reported by the customer and the vendor, not as measured results. That is normal for a project in its first months, and it should be labelled as such.

What IBM changed during the third quarter

IBM Bob did not ship for this deal. Across the third quarter of 2026 it gained multi-agent parallel execution, cost tracking under the name Bobalytics, self-hosted and air-gapped deployment options, and specialised modernisation workflows. Multi-agent execution matters for integration work because the job splits naturally into independent sub-tasks, such as mapping one payment interface while another agent reads the specification for the next. Cost tracking matters for a less glamorous reason: agentic workflows make many model calls per task, and the bill can drift away from the estimate without anyone noticing until the invoice arrives.

Why the deployment model matters in Europe

Since 2 August 2026, the European Commission and the AI Office enforce obligations on general-purpose AI models directly. Regulators can run model evaluations, demand transparency, require risk mitigations and issue fines of up to €15 million or 3% of global turnover. The voluntary period, when providers operated on commitments rather than statutory oversight, has ended, and adherence to the Code of Practice is now audited as a compliance benchmark rather than a draft to comment on.

For a European bank, the deployment question goes beyond the model. DORA has applied to EU financial entities since January 2025 and covers ICT third-party risk, including the risk of leaning on one supplier. On top of that, the European discussion about residency has moved on: hosting on EU soil under a non-EU parent is no longer treated as a sufficient answer to extraterritorial access laws. An air-gapped deployment, where the agents and the source code stay inside the bank's own infrastructure, answers that question without a legal argument. It is also the version most likely to survive a procurement review.

What the models behind such agents cost

IBM has not published per-token pricing for Bob, so the closest comparison available is the API list price of the frontier models that agentic coding tools typically call. These are USD list prices published by the vendors themselves; European business customers pay in USD and add local VAT where applicable. Every row below carries its source and the date the price was checked.

Claude Haiku 5.5 pricing applies to requests up to 100k tokens. DeepSeek's peak rate is double the off-peak figure. These prices say nothing about quality on integration tasks, and a platform like Bob may route different steps to different models. The gap between the cheapest and the most expensive row is a factor of roughly 27 on input and 40 on output, which is why cost tracking inside an agent platform is not a decorative feature. We keep our own per-token and throughput measurements, including local models, in AI Arena.

What cannot be calculated from outside is the cost per agent request. IBM and Mashreq have published the number of requests, not the number of tokens each one consumes, and an agentic workflow of the kind described here may make several model calls before it produces an answer a human accepts. A cost-per-request figure would need both halves of that equation.