Change Management for AI Adoption

The Technology Is the Easy Part

Every AI transformation has two components: the technology and the people. Boston Consulting Group's widely cited 10-20-70 rule puts numbers on the balance: roughly 10% of the value from AI comes from the algorithms, 20% from the technology and data, and 70% from changing how people and processes work. Most institutions spend their energy the other way round.

Large Language ModelsLarge Language Model (LLM)A neural network trained on vast amounts of text data that can understand and generate human language. LLMs power chatbots, document analysis, code generation, and many enterprise AI applications.See glossary are only valuable if people use them -- and in banking, where new technology touches customer data, compliance, and job security, adoption is anything but automatic.

Understanding Resistance

In banking, AI resistance comes from specific, often legitimate concerns:

Fear of Job Loss

This is the elephant in every room. Bank employees read the same headlines as everyone else, and many read every AI deployment as a sign their role is next.

The honest answer: Some roles will shrink -- especially in operations and back-office functions -- and large banks now say so openly:

• JPMorganChase: In his April 2026 shareholder letter, Jamie Dimon wrote that AI "will definitely eliminate some jobs, while it enhances others," and that the firm will have definitive plans to support and redeploy affected employees.
• Standard Chartered: In May 2026 the bank said about 7,800 roles would go by 2030, mostly in corporate functions such as HR, risk and compliance -- more than 15% of its back-office roles. CEO Bill Winters said staff who want to reskill are being given "every opportunity."
• The pattern in the numbers: American Banker reported in July 2026 that JPMorgan's operations headcount was down about 4% while its client-facing roles grew about 4%, and that Wells Fargo had reduced headcount for 23 consecutive quarters.

Client-facing and judgment-heavy roles are more likely to evolve than disappear: relationship managers freed from paperwork spend more time with clients. The leading banks pair honesty about reductions with explicit redeployment and reskilling commitments. Do the same -- and do not let people read about it in the press first.

Regulatory and Compliance Anxiety

Compliance officers and risk managers are being asked to accept a technology they may not fully understand into processes where errors have regulatory consequences. Their caution is not resistance -- it is professionalism.

The honest answer: GuardrailsGuardrailsSafety mechanisms that constrain AI model outputs, and the actions AI agents can take, to prevent harmful, off-topic, non-compliant, or unauthorized results. Critical in banking for regulatory adherence and brand safety.See glossary, human oversight, and governance frameworks exist specifically to address these concerns. Engage compliance and risk teams as partners in AI deployment, not as obstacles to overcome. Their input makes AI deployments safer and more durable.

Distrust of AI Accuracy

Employees who have used consumer AI tools have met confidently wrong answers. Asking them to trust AI where errors have real consequences is a significant ask.

The honest answer: Acknowledge the limitation. AI is not infallible, which is why every banking AI deployment includes human oversight, validation, and the ability to override. Position AI as a tool that assists human decision-making, not one that replaces it.

Loss of Expertise Value

Senior employees who have built careers on deep domain expertise may feel that AI diminishes the value of their knowledge. If an LLM can answer regulatory questions, what value does 30 years of compliance experience provide?

The honest answer: Enormous value. AI can retrieve information, but it cannot build relationships with regulators or make nuanced calls on edge cases. Someone has to validate AI outputs and make the judgment calls -- that is senior expertise.

Communication Strategy

Principle 1: Lead with Honesty

Do not promise that AI will not change anything. It will. Do not promise that every job is safe forever. They may not be. Instead:

• Acknowledge that AI will change workflows and some tasks will be automated
• Explain how the institution will invest in retraining and role evolution. Firm-wide access to AI tools is becoming the norm at large banks, which makes literacy training urgent: BNY, for example, has given all of its employees access to its internal AI platform, Eliza, and reported in April 2026 that about half use AI daily
• Commit to transparency about AI deployment plans and their impact on teams
• Provide concrete examples of how similar institutions have managed the transition

Principle 2: Show, Do Not Tell

Abstract promises about AI are less compelling than concrete demonstrations:

• Let employees try AI tools in a safe, sandboxed environment before any formal deployment
• Share specific examples of time saved, quality improved, and errors prevented
• Feature early adopters who can speak peer-to-peer about their experience

Principle 3: Involve People in the Design

The fastest way to build buy-in is to involve users in designing the solution:

• Include front-line employees in use case selection (they know where the pain points are)
• Ask power users to test and provide feedback during development
• Create feedback loops so user suggestions visibly improve the product

Designing Effective Pilots

AI pilots are change management tools, not just technology experiments: a well-designed pilot builds evidence that AI works, surfaces adoption barriers, and creates internal champions.

Pilot Design Principles

1. Select enthusiastic participants. Do not force reluctant employees into the pilot. Volunteers who are curious about AI will give fairer feedback, work through initial friction, and become advocates if the experience is positive
2. Define success metrics before launch. Time saved, error reduction, user satisfaction, output quality -- agree on them before the pilot begins so results are credible
3. Provide adequate training. Most pilot failures are training failures. Invest in hands-on workshops for people's specific tasks, not just documentation
4. Allow a learning curve. Performance may dip before it improves. Do not judge the pilot on its first week
5. Collect structured feedback. Weekly surveys, one-on-one check-ins, and usage data provide the evidence base for scaling decisions

Scaling from Pilot to Production

The transition from pilot to production is where many AI initiatives die. Common failure modes:

• Pilot paradise: The pilot succeeds on disproportionate support; when that support is withdrawn, adoption collapses
• Forcing functions: Mandating AI use without training. Employees comply superficially but do not change how they work
• Champion dependency: One enthusiastic champion drives adoption; when they move on, usage drops

Mitigation: Scale gradually. Expand from the pilot group to adjacent teams. Train new users using existing users as mentors. Maintain support infrastructure (help desk, prompt libraries, best practices) throughout the scaling period.

Measuring Adoption

Deployment is not adoption. A tool that is installed on every desktop but used by nobody has zero value. Measure what matters:

Usage Metrics:

• Daily/weekly active users (not just accounts provisioned)
• Queries per user per week
• Feature utilization (which capabilities are used, which are ignored?)

Impact Metrics:

• Time saved per task (measured, not estimated)
• Output quality (does AI-assisted work meet or exceed prior quality standards?)
• Error rates (do AI-assisted processes produce fewer errors?)
• User satisfaction (regular surveys on tool usefulness and frustration points)

Cultural Metrics:

• Voluntary adoption rate (how many non-mandated users choose to use AI?)
• Internal referral rate (are users recommending the tool to colleagues?)
• Feedback submission rate (are users engaged enough to suggest improvements?)

What good looks like: At the three banks ranked highest by Evident, more than three-quarters of staff use AI assistants regularly and more than 80% of engineers use AI tools (Evident, October 2026). Bank of America reported in July 2026 that more than 18,000 of its customer service representatives use its EricaAssist tool, cutting average call time by nearly a minute. Morgan Stanley reported in 2024 that more than 98% of its wealth advisor teams used its AI assistant.

Handling Resistance When It Persists

Despite best efforts, some resistance will persist. Distinguish between:

• Constructive skepticism: Legitimate concerns about accuracy, compliance, or workflow disruption. These people make AI deployments better -- listen and incorporate their feedback
• Wait-and-see caution: Not opposed, but want proof first. Give them peer testimonials and concrete pilot results
• Active resistance: Refusal to engage regardless of evidence or support. This is a management issue, not a technology issue -- handle it through normal performance management, not by making AI a battleground

Managing AI Agents as Part of the Workforce

AI agentsAgentsAI systems that can autonomously plan and execute multi-step tasks by calling tools, querying data sources, and making decisions without human intervention at each step -- typically within defined permissions and human approval checkpoints.See glossary -- systems that carry out multi-step tasks rather than just answering questions -- raise a new question: how do people and agents work on the same team? Banks are already finding out:

• BNY reported about 140 "digital employees" in April 2026. According to press reports, each has its own system login and a human manager who is accountable for its work.
• Goldman Sachs was developing agents with embedded Anthropic engineers for trade accounting and client vetting and onboarding, CNBC reported in February 2026 -- described then as in development, not yet in production.
• Citi piloted agentic features in its Stylus Workspaces tool with about 5,000 employees in September 2025, capping inputs to keep costs under control.

Treat an agent like a new team member with unusual strengths and blind spots:

• Name a supervisor. Every agent needs a named human who owns its output, just as every desk has a manager who signs off.
• Define escalation. Decide in advance which errors and exceptions the agent must hand to a person, and how quickly someone responds.
• Rewrite job descriptions. When staff move from doing the work to reviewing an agent's work, their role, training and performance measures change. Say so explicitly; reviewing well is a skill that has to be taught.
• Be transparent. Tell employees what agents do autonomously and where human approval remains mandatory.