Training reach starts the count
If your board asks whether AI training worked, a completion figure alone won’t answer it. Citi AI adoption measures are useful as a sequence: training reach, employee usage, developer capacity, then evidence of delivered work.
Banking Dive reported in 2025 that Citi required prompt training for roughly 180,000 employees with access to its AI tools. That describes the population covered by the requirement, not how many completed it, what they learned or whether training changed outcomes. We think required prompt training is a sensible starting point, but leaders should report participation separately from demonstrated skill.
Sources: Citi rolls out AI prompt training requirement to most staff | Banking Dive
Citi employee AI usage needs a denominator
Banking Dive reported in 2025 that Citi employees had written more than 6.5 million prompts on company AI platforms. The figure describes prompt volume, not unique people, repeat use or useful work; a small group of frequent users could contribute many prompts.
In July 2026, Banking Dive reported that Citi CEO Jane Fraser said nearly 9 out of 10 Citi employees were using the tools. That is a reported share, but the account doesn’t define usage frequency or measurement method. For your own reporting, state the population, activity window and whether the unit is people, sessions or prompts.
Sources: Citi rolls out AI prompt training requirement to most staff | Banking Dive, Banks report operational changes driven by AI adoption | Banking Dive
Citi staff prompt training isn’t usage
Citi’s June 2025 employee message said thousands of colleagues were using Citi AI every day after launch to approximately 150,000 colleagues. Citi also said more than 2,000 colleagues were serving in AI Champions and AI Accelerator programs to support adoption and provide feedback. These are distinct signals: reach, daily use as described by Citi, and program participation.
The figures are company statements reported in different contexts, and they don’t establish that training or champion programs caused usage. We’d track training assigned and completed separately, then pair both with a defined active-user measure and a skill assessment if you need evidence of learning.
Sources: Strengthening Our Commitment to AI Innovation
Citi AI coding productivity needs a work measure
In 2025, Global Banking & Finance Review reported Citi CEO Jane Fraser’s statement that AI freed 100,000 hours of developer capacity each week. Treat that as reported capacity: the supplied reporting gives no calculation method, baseline, quality measure or realized cost saving.
Capacity is a claim about time made available, not proof of completed work or productivity. These are leadership-reported figures, not causal evidence that AI produced the hours or improved results. Track the estimate alongside delivered work, cycle time and quality, and document how you calculate each measure.
Sources: Citigroup's AI Usage in Banking Saves 100,000 Developer Hours
Measure the work before the spend
The supplied reporting doesn’t establish a limit on Citi employees’ access to premium models, so don’t use that claim to explain usage or productivity. If access or cost control matters in your organization, report your own policy and spending data alongside adoption measures.
Use a staged scorecard to keep each claim in its lane:
- Training: report assigned and completed participation by team and date; assess skills separately if learning matters.
- Breadth of use: define active use and the observation window, then report the share of the intended population using AI.
- Capacity: state the baseline and calculation behind any hours estimate; label it capacity until redeployment is shown.
- Productivity: name the workflow, compare defined periods, and track delivered work, cycle time and quality.
- Cost: report spend and any access limits directly; claim savings only when your cost evidence supports them.
Frequently asked questions
How often should we repeat these measures?
Repeat them on a cadence that fits the initiative and keep the definitions stable across rounds. Record the date, population and activity window each time. If teams or roles change, preserve those segments where possible so shifts in the overall figure don’t hide differences in access or use.
Who should own the measurement plan?
Assign one accountable owner for the definitions and reporting, then involve the teams that hold each signal. L&D can report training participation, IT can clarify access and policy, and engineering leaders can define capacity and work measures. Agree on the population and terms before collecting results.
What evidence supports a productivity claim?
Use a defined workflow, a documented baseline and comparable periods, then examine delivered work, cycle time and quality together. Record relevant changes that could also explain the result. A leadership statement or estimated hours freed can motivate that analysis, but doesn’t establish that AI caused an improvement.
Sources: Citigroup's AI Usage in Banking Saves 100,000 Developer Hours
