M&T Bank, a US regional bank, has rolled out internal AI copilots to more than 15,000 employees. The tools now cover call-centre conversation analysis, report drafting, code generation, spotting customer needs, and flagging portfolio risk. What makes the bank interesting is not how quickly it deployed the technology, but the eight years it spent rebuilding its technology organisation and data foundation before touching generative AI.
It started by banning public LLMs
M&T Bank's first move was restriction, not adoption. It blocked employee access to public large language models. Chief data officer Andrew Foster pointed to the risk of staff entering sensitive company information into consumer-facing services, a reasonable call for a regulated institution.
The bank then evaluated enterprise vendors and selected Microsoft Copilot. Rather than opening it to everyone at once, it began with a pilot of roughly 800 employees and expanded from there. The most concrete result disclosed so far is call-centre work: having generative AI summarise conversations saves about 6 minutes per call. Developers use GitLab tooling to generate code.
Responsibility for the accuracy of AI-assisted output stays with the employee. That principle is written into the bank's 2026 Code of Business Conduct and Ethics, which requires the use of approved AI tools and prohibits entering confidential, proprietary, customer, employee, or regulated information into unapproved systems.
Eight years of rebuilding came first
The generative AI rollout sits on top of a technology overhaul that began in 2018. Back then, more than half of the bank's technology specialists were external contractors. Today about 80 percent of that workforce is in-house. M&T has roughly 2,000 technologists spread across more than 300 agile teams, and it hired over 1,000 technology specialists during the programme.
Legacy platforms were replaced in parallel. The bank says technology outages have dropped by more than 80 percent since 2018, while the number of system upgrades completed each year has tripled. According to Wisler, who joined as chief information officer in 2018 and became senior executive vice-president for technology and operations in 2025, annual releases grew from about 15,000 in 2018 to 65,000 in 2025. Technology spending passed 1.2 billion USD (about 187 billion yen) in 2025, close to three times the 2017 level.
※1 USD = 156 JPY (as of September 8, 2026)
In other words, the platform was rebuilt before anything was placed on it.
Making data traceable to its source
The second piece of groundwork was on the data side. Foster, who joined in 2023, launched a data-lineage programme to track where information originates, how it is used, and how it moves between systems. He has said the work was not a reaction to generative AI, but a core capability for understanding the bank's data estate.
In practice, M&T uses data-lineage software from Solidatus and Monte Carlo to trace information as it passes through databases, applications, and business intelligence systems. It also built an internal repository called Edison that holds authoritative documents and bank policies. Retrieval-augmented generation runs against this governed data. A Data Academy covering data governance and data skills has drawn around 2,000 participants.
Source, meaning, quality, and governance are visible for individual data elements. That state was established first, and Copilot was layered on afterwards.
Three paths, and where peers stand
Wisler frames the bank's generative AI work as three paths: general employee use, AI capabilities embedded in applications the bank already runs, and proprietary systems built around its own data and processes. M&T operates more than 1,800 applications, many of them from third-party vendors, so finding AI features already shipped inside that software is a practical second path. On the proprietary side, early targets include repetitive operational work, software development, fraud prevention, and cyber defence.
Larger banks are further along. JPMorganChase opened its internal LLM Suite platform to more than 200,000 employees in 2024, and by 2025 more than 65,000 people in its Corporate and Investment Bank were active users, with over 90 percent of engineers using AI coding assistants. Its AI-based transaction screening reviewed more than twice the previous volume while halving manual operator checks. Bank of America has deployed EricaAssist to more than 18,000 customer service staff; as of July 2026 it delivers contextual guidance in under 3 seconds and has cut average call times by nearly a minute.
M&T employs about 22,000 people. In absolute numbers it cannot match those figures, but in terms of coverage it is not far behind.
Published user counts depend on how you count
One note on reading the numbers. American Banker reported in September 2025 that 16,000 of M&T's roughly 22,000 employees were already using Microsoft Copilot. The figure disclosed now is more than 15,000. It looks like a decline, but the answer changes depending on whether the number refers to licences issued, people with access, or people actively using the tool.
Coverage of enterprise generative AI often leaves that distinction unstated while the headcount travels on its own. Comparisons only work once you know what was counted.
Summary
M&T Bank has extended AI copilots to more than 15,000 employees and saves about 6 minutes per call on conversation summaries. Behind that sits a technology overhaul running since 2018: 80 percent in-house staffing, outages down by more than 80 percent, 65,000 annual releases, and 1.2 billion USD in technology spending. Data lineage and the Edison repository made the provenance of the data handed to AI traceable before the rollout began. If there is a lesson here for enterprise AI adoption, it is less about which model was chosen and more about what was put in order before the choice was made.
