A Minute Disappears From 18,000 Bank Calls
At noon Eastern on July 21, Bank of America attached numbers to a customer-service call.
More than 18,000 customer-service representatives use EricaAssist, an internal AI assistant that works beside them during conversations. The bank says its new generative AI capabilities can surface contextual guidance in under three seconds. It attributes a reduction of nearly one minute in average call time to the system.
Picture the minute from the employee’s side. A customer calls about a transfer that does not look right. The representative has to listen, authenticate the caller, understand the account history, decide whether the issue is routine or urgent, explain the next step and leave a usable record for whoever may handle the case later. EricaAssist summarizes the reason for the call, retrieves relevant information and recommends a next action without asking the employee to leave the conversation.
If that works as described, the representative spends less time searching and more time listening. A client waits less. A queue moves faster.
Then the minute reaches the operating review.
At the operating review, the contact-center leader can multiply the minute by call volume and produce a large capacity number. Finance translates it into labor cost and puts the result beside Bank of America’s $14 billion annual technology budget. Across the table, human resources asks whether the job now requires stronger judgment because easy search work has been compressed. Compliance wants to know whether a faster call still recognizes a dispute and gives the customer an effective path to a person who can fix it.
The bank’s July 21 announcement does not disclose call volume, the baseline handle time, the distribution of savings across call types or a headcount result. It says the average call is nearly one minute shorter. That is a useful product claim, but it is not yet a workforce decision.
Citi makes the distinction harder to ignore. At its May investor day, the bank said its own Agent Assist reduced call handle time by 60 seconds. Two large banks, two employee-facing systems, the same round number.
That does not prove that every bank can remove a minute from every call. It does create a benchmark that executives will carry into budget meetings. Once a vendor, a bank and an investor presentation put 60 seconds on the page, the time becomes something a buyer can price.
The price still depends on where the minute goes.
It can become shorter customer waits, more calls handled by the same team, better explanations, more sales conversations, more time for a distressed client, additional review of suspicious activity or fewer scheduled shifts. It can also disappear into a higher volume target while employees inherit only the complex calls the software cannot resolve.
Bank of America calls EricaAssist “human-assisted AI.” That phrase describes a product architecture and a labor choice. The AI does not take the customer’s place in the conversation. It changes the information environment around the employee who answers.
The product may release a minute. Leaders decide whether the minute becomes capacity, quality, revenue or payroll.
July 21 put three seconds beside the caller
EricaAssist is not the consumer chatbot named Erica. It sits on the employee side of the call.
Bank of America says the assistant uses generative AI to summarize why a client is calling, pull together relevant information and recommend next steps based on the employee’s role and the client’s relationship with the bank. The guidance appears in under three seconds. The company plans to extend the system to more service scenarios and business lines later in 2026.
Where the screen sits changes who has to translate the problem. A consumer chatbot asks the customer to express a problem in terms the system can understand. An employee assistant listens to a conversation already being handled by a person with access, training and responsibility. If the system misses the issue, the representative is still present to notice. If the answer needs explanation, a person can adapt it. If the case needs escalation, the employee can move it.
Ashley Ross, Bank of America’s head of Consumer Client Experience and Business Transformation, describes the design as a combination of real-time guidance and human judgment. Tom Ellis, the bank’s consumer chief information officer, connects it to governance, transparency and accountability. Those are executive claims from the company selling the result, so buyers should test them. They also make the intended division of labor explicit.
The AI retrieves and proposes. The employee listens, interprets, explains and remains accountable for the service moment.
For a representative, three-second retrieval can remove several kinds of friction. A policy may be buried in a knowledge base. A customer’s relationship may cross products. A recent interaction may explain why the person is calling again. A new employee may know the rule but not which exception applies. Searching while a customer waits creates silence, repeat questions and cognitive switching.
An assistant can compress that search. It can also create a new review task. The employee has to notice whether the retrieved guidance belongs to this product, this jurisdiction and this client. A fluent recommendation can arrive faster than the employee can inspect its basis. Latency falls while the demand for rapid judgment rises.
Onboarding changes with the screen. Traditional service training often teaches where information lives, which script applies and which screen comes next. When the system handles more retrieval, training has to spend more time on ambiguity, customer state, policy exceptions and escalation. The representative needs to recognize a wrong answer in seconds, not after a quality reviewer finds it a week later.
The interface should make that judgment possible. A recommendation needs a visible source, effective date and policy owner. The employee needs a quick way to reject it, report a mismatch and continue without losing the customer. High-risk call types should show a clear escalation path. When the employee chooses another action, the system should preserve the reason without turning every disagreement into a performance mark.
Three seconds is therefore only the first service-level measure. Leaders also need to know how often guidance arrives, how often employees use it, how often they override it, whether the override was correct and whether the call produces another contact.
Fast retrieval helps only when the employee can see enough evidence to decide whether the guidance belongs in this call.
One minute becomes a finance decision
Average handle time is an operating metric, not a cash result.
Suppose a contact center receives 100,000 assisted calls in a week and the average saving is 60 seconds. The arithmetic releases 1,667 hours. It does not say that payroll falls by 1,667 hours. Calls arrive unevenly. Schedules include breaks, coaching, meetings and idle intervals between peaks. Some saved seconds may reduce a queue without removing a shift. Some may let the same staff absorb growth. Some may be spent on documentation that used to happen after the call.
The first finance task is to separate theoretical capacity from usable capacity.
Theoretical capacity is the handle-time reduction multiplied by eligible interactions. Usable capacity is the portion that can be collected into real blocks of work or removed from a schedule without harming service. Financial value is the part of usable capacity that produces an observable outcome: fewer overtime hours, less outsourcing, lower repeat contact, higher retention, additional revenue or a deliberate staffing change.
Those numbers will differ by call type. A balance inquiry may have little room to save because it is already short. A fraud concern may become longer when the employee asks better questions. A bereavement or financial-hardship call should not be pushed toward the fastest possible close. An AI system may make the average fall by handling common searches while the remaining tail of difficult calls grows.
An average can hide that tail.
Finance should ask for a distribution: median handle time, upper-percentile time, transfer rate, hold time, after-call work and repeat contact by call reason. Operations should pair it with customer outcomes and quality review. Workforce management should show whether the released time changed schedules, occupancy or vendor demand. HR should show whether the work mix changed skill requirements, training time or attrition.
Without those links, a one-minute claim invites two errors.
The optimistic error books all released time as savings. The pessimistic error assumes the minute cannot have value unless jobs disappear. A shorter queue, lower abandonment, faster fraud resolution or less employee scrambling may be worth money even when headcount is unchanged. The value needs its own measure.
The bank has not published the denominator needed for an outside calculation. We do not know how many calls receive EricaAssist guidance, how many minutes the typical eligible call lasts or what proportion of the 18,000 representatives use the generative features on a given day. Multiplying 18,000 employees by an invented call count would produce a dramatic number with no reliable basis.
The absence is useful. It shows exactly what a buyer should request during a pilot.
Define eligible call types before launch. Record the baseline for each group. Compare assisted and unassisted work over enough time to include peak periods and unusual cases. Account for after-call work and repeat contacts. Track whether savings persist after employees become familiar with the tool. State which costs are included: licenses, integration, data preparation, model evaluation, quality review, training and manager time.
Then choose what the minute is supposed to fund.
If the stated goal is customer access, use it to reduce waits and abandonment. If the goal is capacity, show the avoided hiring or outsourcing decision. If the goal is employee experience, measure search burden, schedule pressure and attrition. If the goal is revenue, separate a better conversation from an aggressive sales prompt.
The pilot should end with a named operating decision that converts the calculation on a slide into financial value in the business.
Citi found the same sixty seconds
Pam Habner, head of U.S. Consumer Cards at Citi, put another 60 seconds in front of investors on May 7.
In the bank’s 2026 Investor Day transcript, Habner says Citi rolled out Agent Assist across customer service and reduced call handle time by 60 seconds. She places the result inside a wider cost-to-serve story: digital adoption rose 18%, cost per account declined 12%, digital collections rose 21%, cost per delinquent account fell 22% and digital Net Promoter Score increased five points.
The time window and methods behind those figures are not fully described in the transcript. The claims come from Citi, and several changes occurred together. Digital adoption, platform upgrades, process work and AI may all affect cost and satisfaction. Agent Assist cannot claim every improvement around it.
Still, Citi does something important with the savings. Habner says scale economics fund efficiency and reinvestment in growth. She describes AI as both expense reduction and a way to win customers. That gives the released minute at least two possible destinations.
One destination is a lower cost per account. The other is more capacity for acquisition, personalization and service.
Employees feel the operating choice in their daily queue. If a bank reinvests capacity, a representative may handle a growing customer base without a matching increase in staffing. The job remains, but the expected output per person rises. If easier interactions move to digital channels, phone employees receive a denser mix of disputes, exceptions and anxious customers. The average call can fall at the same time that the human work becomes more demanding.
Citi also says technology developers are managing teams of coding agents and seeing productivity gains of up to 40% in software development time. The comparison crosses functions, but the labor pattern is similar. AI produces or retrieves more work. A person moves toward orchestration, review and exception handling.
That shift is often described as elevation. It can be, if the employee receives training, authority and pay for the added judgment. It can also become work intensification. Removing routine tasks can remove recovery time and the cases through which a newer employee learned the system. More complex work arrives with the same schedule adherence target.
Two banks reporting the same minute should encourage comparison, not imitation.
A buyer should ask whether both are measuring talk time, total handle time or another interval. Does the measure include hold time and after-call documentation? Is the result an average across every call or a subset using the assistant? Did repeat contacts change? Did quality scores change? How did the employee mix, call mix and digital channel mix move?
The common number is a market signal. It is not a universal constant.
Once 60 seconds becomes a benchmark, vendors will design demonstrations around it and contact-center leaders will be asked why their result is lower. A responsible review preserves the context. A bank that spends longer on a complex dispute and resolves it once may perform better than a bank that ends two short calls without fixing the problem.
The meaningful comparison is not seconds removed. It is the customer and workforce outcome purchased with those seconds.
$33.8 billion funds two human systems
Bank of America says it now spends $14 billion a year on technology, with more than $4 billion allocated to new initiatives including AI. JPMorganChase gives a 2026 technology budget of approximately $19.8 billion. Together, the two disclosed figures total $33.8 billion.
That sum is not an AI budget, and it is not a fair estimate of what either bank spends on employee assistants. It covers infrastructure, cybersecurity, data, maintenance, product work and many other programs. The number is useful because it shows the scale of the operating system around a one-minute result.
Bank of America’s March shareholder letter reports $13 billion of technology spending in 2025, including more than $4 billion for new initiatives. The July EricaAssist release gives the current annual figure as $14 billion. Brian Moynihan writes that AI capabilities have been deployed to more than 213,000 employees. He also points to relatively flat headcount from 2007 through 2025 as the company grew.
That headcount statement changes the frame. AI does not need to trigger a named layoff to affect employment. A company can grow customers, products and interactions without adding workers at the rate an older operating model would have required. Attrition, internal movement, selective hiring and productivity can alter the workforce slowly.
Bank of America also says full-time U.S. employees now earn at least $25 an hour, and that low attrition accompanies its pay and benefits approach. The bank has committed to hire 10,000 people with military backgrounds and 8,000 community-college hires over five years. These facts do not cancel the productivity story. They show that a labor strategy can combine hiring commitments, higher pay, technology and slower aggregate headcount growth.
JPMorganChase makes its intended trade more explicit. In her April 6 letter to shareholders, Chief Operating Officer Jennifer Piepszak writes that productivity gains will free capacity for reinvestment in growth. She says the goal is not lower headcount, while acknowledging that fewer people may be the result in certain jobs. The company plans to upskill, reskill and redeploy talent as technology changes work.
Piepszak is describing a human system around the technology budget. More than 320,000 employees need tools, training, mobility routes and jobs into which released capacity can move. JPMorganChase recently rolled out an Employee Assistant and says it fills thousands of roles each year, many through internal mobility.
The words “reinvest” and “redeploy” need operational definitions.
Reinvestment can mean a new market, a larger customer base, a new product or better controls. It should identify the business unit receiving capacity and the result expected. Redeployment should name the employee population, destination roles, skill gaps, pay treatment, selection process and time window. Otherwise, both words can sit in an investor letter without changing an employee’s next assignment.
Large technology budgets leave room to fund the human layer. Whether money reaches that layer is a separate operating decision.
Every contact-center assistant has companion costs that rarely appear in the launch headline. Employees need time to learn the interface and practice overrides. Quality teams need samples large enough to find rare but serious failures. Knowledge owners need to keep sources current. Managers need to coach changed behavior. Workforce planners need new assumptions for call mix and staffing. Compliance teams need evidence about disputes, disclosures and escalation.
The $33.8 billion total puts those costs in perspective. The banks can afford sophisticated technology. The harder budget test tracks how much follows the released minute into training, quality and real mobility, and how much is absorbed by a higher productivity target.
Human-assisted AI keeps the employee in the call
The most consequential design choice in the Bank of America announcement is where the AI sits.
EricaAssist supports the representative. It does not require the customer to negotiate directly with a generative model. That structure resembles Morgan Stanley’s approach to advisor work. In 2024, the firm launched AI @ Morgan Stanley Debrief, a tool that can create meeting notes, action items, an email draft and a draft record for Salesforce when a client consents. Morgan Stanley estimated that the workflow could save about 30 minutes per meeting. The advisor reviews the output.
Both designs place AI inside a trusted relationship without removing the licensed or trained person who owns it.
Human presence is not a complete safeguard. An employee can over-trust a suggestion, miss a source error or follow a prompt because the metric rewards speed. A recommendation may be shaped by incomplete data. If the interface makes the AI answer look authoritative and the override cumbersome, nominal human oversight can become a click.
The quality of human assistance depends on authority and time.
The representative must be allowed to slow down a call. They need the authority to reject guidance and escalate a case. The system should distinguish a thoughtful override from noncompliance. Managers should review the reason and outcome, not reward agreement with the AI by default. A quality team should test whether the assistant changes decisions across customer groups, languages, products and vulnerability signals.
Human-assisted design also changes accountability. When a chatbot gives the customer an answer directly, the record can show the exact exchange. When AI guidance passes through an employee, the final explanation may differ from the prompt. The flexibility helps the conversation, but evaluation needs both layers: what the system showed and what the employee told the customer.
Recording those layers raises privacy and surveillance questions. A bank already monitors calls for quality and compliance. Adding model suggestions, acceptance, override and response time creates a richer behavioral record. That record can improve the system. It can also become a hidden employee score.
The purpose should be bounded before collection begins.
Use interaction data to test guidance quality, identify training needs and repair knowledge. Do not assume that frequent overrides indicate resistance or weak performance. A representative working on unusual cases should override more often. An employee serving customers with limited English proficiency may spend longer clarifying. Someone receiving more fraud or hardship calls should not be compared with a colleague handling routine requests without adjusting for case mix.
The employee needs a route to challenge the record. If AI-use telemetry influences coaching, performance or scheduling, the bank should tell workers which measures are used and let them correct a misclassified case.
Keeping a person in the call preserves context and an exit from automation. It also places a new burden on that person: inspect a fast answer, protect the client and remain efficient at the same time.
Human-assisted AI succeeds when the employee has enough power to make the assistance human.
The customer still needs an exit from the loop
The Consumer Financial Protection Bureau documented the other architecture in 2023.
Its report on chatbots in consumer finance found that all ten largest U.S. commercial banks had deployed chatbots. It estimated that 98 million people, about 37% of the U.S. population, interacted with a bank chatbot in 2022. The Bureau saw value in basic inquiries and 24-hour access. It also collected cases involving inaccurate information, repeated loops, failure to recognize disputes and difficulty reaching a person.
The report’s clearest boundary concerns a system that does not understand the request. In that situation, the Bureau says a chatbot is not suitable as the primary service vehicle.
EricaAssist avoids part of that problem because a representative is already present. The customer does not have to discover the secret phrase that reaches a person. Yet an employee-facing assistant can still reproduce the underlying failure.
Consider an illustrative call that starts as a merchant question. The customer then says the charge was not theirs. A generated summary may keep the first label and retrieve transaction information. The representative has to hear the later sentence, recognize a dispute and open the required process. In that case, accepting the fast summary saves seconds and creates a failure. The employee’s interruption is the control.
The same problem can appear when guidance retrieves an old policy or recommends a routine path after the customer’s words indicate coercion, bereavement or hardship.
The service design needs stop conditions.
A potential dispute should trigger confirmation that a case was opened, a case identifier and a clear next step. A vulnerable-customer signal should slow the workflow and widen escalation options. A recommendation without a current source should not appear as settled guidance. Repeated contact about the same issue should raise priority and block the system from offering the same script again.
Customer outcome measures belong beside handle time. Did the problem resolve? Did the person call again within a defined period? Was a promised action completed? Did a complaint follow? Was the case transferred, and did the next employee receive the context? Did the customer understand the decision and the route to challenge it?
These measures can move against speed. An employee may spend another two minutes explaining a denial and prevent three later calls. A representative may recognize a dispute that an automated summary missed, lengthen the interaction and protect the bank from a legal failure. A hardship conversation may need patience that no average should punish.
The regulatory position is also moving. The OCC’s April revised model risk guidance emphasizes risk-based development, testing, validation, monitoring, governance and controls for models within its scope. The bulletin explicitly says generative and agentic AI are outside that guidance and points to a forthcoming request for information. It should not be presented as a prescriptive rule for EricaAssist.
The gap makes internal discipline more important, not less. A bank cannot wait for one document to define every evaluation step for a fast-changing employee assistant. Existing consumer obligations still apply to the service process, whether an answer came from a knowledge base, a script or a model.
A reliable exit from the loop is both a customer feature and a workforce feature. Customers need access to a person with authority. Employees need access to a specialist who can take the case. If every escalation returns to the same AI-generated suggestion, the human layer is decorative.
The best minute may be the one an employee chooses not to save.
A one-minute service ledger
A contact-center AI business case needs one table that product, operations, finance, HR, risk and employee representatives can read together.
The rows should be call types, not a single blended average. The columns should follow the released minute from suggestion to staffing decision.
| Call type | AI action | Human judgment | Time released | Quality and compliance check | Customer outcome | Employee skill | Capacity decision | Stop signal |
|---|---|---|---|---|---|---|---|---|
| Routine balance or payment question | Retrieves verified account and policy information | Confirms identity and explains the answer | Search and hold time | Source date, authentication and explanation accuracy | Resolved once, no repeat contact | Clear explanation | Reduce queue or absorb volume | Wrong source or authentication miss |
| Card transaction dispute | Summarizes the claim and proposes the dispute path | Recognizes a dispute, tests details and opens the right case | Data gathering and after-call notes | Case created, required notice and deadline | Case ID received and action completed | Dispute recognition | Keep time for investigation | Summary labels dispute as inquiry |
| Suspected fraud | Surfaces recent activity and escalation options | Distinguishes fraud from customer error and protects the account | Retrieval only | Escalation, security action and false-positive review | Exposure contained without blocking valid use | Risk judgment | Add specialist capacity | Model suggestion delays protective action |
| Financial hardship | Finds eligible assistance paths | Listens for context and explains tradeoffs | Policy search | Eligibility, disclosure and fair-treatment review | Customer reaches a workable plan | Empathy and exception handling | Reserve longer call time | Handle-time pressure shortens assessment |
| Bereavement or vulnerable customer | Retrieves sensitive-case guidance | Adjusts pace, language and handoff | Little or none expected | Consent, privacy and specialist referral | Fewer repeated explanations | Care and escalation | Protect a service floor, not a saving | Speed target changes employee behavior |
| Product inquiry with sales potential | Surfaces relevant products | Separates helpful recommendation from pressure | Preparation | Suitability, disclosure and complaint review | Appropriate choice and durable relationship | Needs diagnosis | Reinvest in advice | Conversion rises while complaints rise |
The values in the time column should come from measured pilots. The examples above are hypotheses, not Bank of America results. They show why one average cannot govern every row.
The ledger forces several decisions.
Product owns the assistant action and source lineage, while operations owns the workflow and queue. The employee owns the final conversation within clear authority. Risk and compliance define cases that need extra review; HR and workforce management cover training, scheduling and role changes. Finance decides which outcomes count as value, and a customer-experience leader checks whether the person calling can feel the result.
Each row needs a baseline and a review date. Time saved should be split into talk time, hold time and after-call work. Quality should use calibrated samples, including cases where employees rejected the suggestion. Customer outcomes should include repeat contact and completed actions, not satisfaction alone. Workforce outcomes should cover overtime, hiring, outsourcing, attrition, internal movement and case complexity.
The ledger also needs an uncertainty column in practice. Vendor-reported claims should be labeled. Results from a limited pilot should not be projected across business lines without adjustment. A headcount scenario should distinguish positions eliminated, positions not added, vacancies left open and people moved. Those are different employee and financial outcomes.
The stop signal is the most important field.
A call type should leave the assistant or return to a narrower configuration when guidance creates repeat errors, prevents dispute recognition, changes outcomes across customer groups, overwhelms specialists or makes employees rush sensitive conversations. A system can keep its average minute while failing a high-consequence minority of calls.
The ledger prevents that minority from vanishing inside the mean.
It also gives employees a practical role in evaluation. Representatives know which searches waste time, which policy language confuses customers and which recommendations arrive without context. Their overrides and explanations can reveal where the product needs repair. At 18,000-person scale, the review group should include people from different queues, languages, shifts and tenure levels, with paid time to inspect cases. The review should treat them as operational evidence, not merely users to be pushed toward adoption.
At quarterly review, leaders should choose the next action row by row: expand, narrow, retrain, change the source, add specialist capacity, alter the metric or stop. A green average does not make every row green.
One minute is a useful entry in the ledger. It is not the ledger.
Training decides where the saved time goes
Bank of America has evidence that broad employee AI adoption is possible.
In April 2025, the bank said more than 90% of employees used Erica for Employees, its internal assistant. It reported that IT service-desk calls had fallen by more than 50% and that more than 19,000 developers saw efficiency gains above 20% from coding tools. These are company-reported measures, but they show a deployment strategy that reaches far beyond a small pilot.
The bank’s AI workforce announcement also points to the Academy and an employee-led AI program. Moynihan’s 2026 shareholder letter describes an AI catalyst effort that collects employee ideas and asks for real business cases. Representatives can identify the workflow that needs help before a central team sees it.
Adoption is only the first training outcome for EricaAssist.
A representative needs to understand what the system can access, how current the source is, when the recommendation is uncertain and which cases require another person. New hires need practice with intentionally wrong or incomplete suggestions. Experienced employees need a way to teach the system team about exceptions without becoming unpaid product testers.
New hires also need calls in which they can learn the structure of the bank’s work. If the assistant supplies every screen and next step from day one, a beginner may become fast without building a map of the underlying process. Training should sometimes hide the recommendation, ask the employee to reason through the case and then compare the answer with policy and an experienced reviewer.
Managers need their own curriculum. They have to coach judgment without reducing the review to tool usage. They should compare cases with similar complexity, inspect customer outcomes and distinguish a justified override from a knowledge gap. They also need to detect when the assistant changes the job faster than the role description and pay structure change.
The Bureau of Labor Statistics projects that U.S. customer-service representative employment will decline 5.5% from 2024 to 2034, a loss of 153,700 positions. Its AI and employment table cites automation as one pressure on the occupation. That national forecast does not predict Bank of America’s staffing, but it gives employees a reasonable reason to ask how the company will use the minute.
Training should answer more than “how do I use the assistant?”
It should show which work is expected to grow, which duties are changing, how performance will be measured and where an employee can move. If routine retrieval shrinks, the company can create paths into fraud operations, complex servicing, quality, knowledge management, model evaluation, complaint analysis or relationship roles. Those paths need openings, selection rules and paid practice.
The destination should be specific. “Higher-value work” is not a job. A redeployment promise without a role, pay band and manager is not a plan.
The saved minute can fund this transition. A team can reserve a portion of released capacity for weekly case calibration, shadowing specialists and reviewing assistant failures. Finance may resist counting that time as a benefit because it does not immediately lower cost. Yet it protects the quality of a system whose business case depends on employee judgment.
Training also shapes fairness. Faster employees may appear more productive because they receive simpler calls or trust suggestions quickly. Experienced workers may take longer because they recognize hidden risk. Language, disability accommodation and customer population affect call patterns. Performance measures should adjust for the work before they influence pay, schedules or promotion.
Training built only around speed will produce faster calls. Practice in explanation, verification and escalation gives the released minute a chance to improve service.
The employee will understand the strategy through the calendar. If every released minute is filled with another call, the message is throughput. If some of it becomes practice and expanded authority, the role is being redesigned. If vacancies quietly disappear, the workforce result is consolidation.
Training reveals which destination leadership chose.
Another call still reaches an employee
Bank of America’s release ends with an expansion plan. More service scenarios and business lines are expected to receive EricaAssist later in 2026.
Before that expansion, the bank has an unusually clear unit to inspect. Three seconds is the retrieval promise. Nearly one minute is the reported average saving. More than 18,000 employees are the current human surface.
The next review should make the missing units equally clear.
How many eligible calls used the system? Which call types changed? Did repeat contact, transfer, complaint and completed-action rates improve? Did employees spend less time searching and more time explaining? Were complex calls protected from a blunt speed target? Did staffing change through hiring, attrition, outsourcing or internal moves? How much released capacity went to training and quality?
Citi’s matching 60-second claim makes those questions relevant beyond one bank. JPMorganChase’s language about reinvestment, redeployment and selected job reductions shows the range of workforce choices available. Morgan Stanley’s advisor tool shows another design that saves time while keeping a person responsible. The CFPB’s chatbot record shows what happens when the customer cannot reach one.
The pattern is wider than banking. Any employer putting an AI assistant beside a service worker will face the same conversion problem. Software can measure a second more easily than an explanation. Finance can count theoretical capacity before operations can collect it. A manager can raise the target before HR defines the changed skill. A launch can promise human oversight without giving the human time to exercise it.
Bank of America has made a strong case that employee-facing AI can remove search friction at scale. It has not claimed that the result requires 18,000 fewer people, and the available evidence does not support that conclusion. The company still offers clients access to tens of thousands of professionals by phone or in person. It also celebrates operating a much larger business with headcount near its 2007 level.
A bank can keep people in service while using technology to avoid adding them at the old rate. It can pay more, hire selectively and redesign work while aggregate staffing grows slowly. The employment effect arrives through budgets and vacancies as much as announcements.
That is why the one-minute service ledger belongs in the operating review before the next rollout.
For each call type, leaders should name the human judgment that remains, the customer outcome that matters, the capacity decision that follows and the failure that stops expansion. Employees should see how the data affects them. Customers should always have a route to someone with authority.
Then the next call comes in.
EricaAssist may summarize it in under three seconds. The representative still has to hear the part the summary missed, decide whether the recommended action fits and explain what happens next. The customer will judge the bank by that moment, not by the benchmark.
One minute has disappeared from the average. Its value depends on what the employee is allowed to do with the time left.
Published July 25, 2026.