Citi’s most revealing AI number is not a benchmark score or investment figure. It is the roughly 4,000 employees participating in the bank’s internal AI Champions and AI Accelerators programs. They represent a distribution network for practical knowledge inside an organisation of about 182,000 people.
More than 70 percent of Citi employees now use firm-approved AI tools in some form, according to reporting cited by AI News. The scale did not emerge from giving a specialist team exclusive ownership. Citi trained volunteers across technology, operations, risk, and customer support to help colleagues use approved systems in everyday work.
The result is an adoption model built around confidence, proximity, and controls rather than one dramatic application.
Local Champions Translate Tools Into Work
Employees could volunteer as AI Champions, receive training and early access, then become approachable contacts inside their own teams. This matters because a central AI office may understand the technology but not the small frictions of thousands of individual roles.
A colleague in the same function can demonstrate when a tool is useful, explain what information must stay out of a prompt, and recognise when an output requires expert review. That contextual guidance is difficult to reproduce through a generic course or company-wide announcement.
Citi reinforced the network with internal badges awarded for completing training or showing how AI improved a task. The badges did not carry direct promotions or pay increases, but they created visible credibility. Employees could identify peers with experience, while participants gained recognition for teaching rather than merely experimenting.
This approach distributes adoption without distributing all governance. Local teams explore relevant uses, while the bank determines which tools and data practices are acceptable.
Guardrails Make Broad Access Easier to Defend
Citi employees use AI to summarise documents, draft internal notes, analyse data, and assist software development. These are familiar applications, yet their deployment inside a regulated global bank changes the risk calculation.
Firm-approved tools place boundaries around the information employees can provide and the way outputs are handled. Such restrictions may slow some experiments, but they also give managers a defensible basis for widening access. In banking, a fast answer has limited value if its provenance, confidentiality, or use cannot be explained.
The program therefore treats trust as an adoption mechanism. Employees are more likely to incorporate AI into routine work when they know where the boundaries are. Managers can evaluate a governed workflow more readily than an informal collection of public accounts and copied data.
Small improvements also compound at Citi’s scale. Saving minutes on a common task can matter more than an impressive pilot used by a few specialists. The relevant measurement is not novelty but reduced friction, better consistency, and controlled time savings.
Peer Networks Still Need Active Maintenance
The model carries risks. Volunteer enthusiasm can fade, and some departments may develop much stronger support networks than others. Informal expertise can also become stale as models, policies, and approved practices change.
Citi has responded by rotating Champions and refreshing training. Leaders will still need to monitor participation, compare adoption across teams, and ensure that underrepresented functions receive support. A peer network should complement professional risk, security, and learning teams rather than quietly replace them.
The bank’s experience also challenges the idea that AI transformation begins with a sweeping redesign. Its rollout starts with existing work and asks where controlled assistance removes repetition. Employees do not need to become machine-learning specialists; enough of them need the confidence to apply tools responsibly and help others do the same.
That may be the quiet work other enterprises underestimate. Buying access creates availability, not adoption. Central mandates create urgency, not fluency. Citi’s 4,000-person network supplies the social infrastructure between the two.
As organisations try to escape pilot mode, their decisive asset may not be a proprietary model. It may be a credible person nearby who can turn a broad capability into a safe, useful habit—and knows when the machine should not be used at all.
