What problem does it solve?
A bank or insurer website holds many pages on accounts, cards, loans and cover, and a visitor often wants a short answer to one question: how do I open an account, what does excess mean, what do I need for a mortgage. ANZ's most common question to Jamie in her first 100 days was how to open a bank account, asked nearly 1,200 times. A text chatbot can answer such questions, but some customers avoid digital channels altogether.
ANZ New Zealand said it started its Jamie trial to see whether a digital human would appeal to customers who are less comfortable with its other digital channels, and NatWest said early testing of its Cora digital human suggested that customers who had avoided digital services might be more inclined to use one.
ANZ said that talking to the bank instead of pressing screens might be easier for certain groups of customers, and around 30% of its New Zealand customers do not use digital channels at all. Institutions also treat the avatar as one more channel: Suncorp described its Ava avatar as another way for customers to engage, next to chatbot, web chat and phone, and UBS tested whether clients accept avatars next to a human advisor in a wealth management meeting.
The risks are specific too. A lifelike face may make an answer more persuasive than the same text on a screen, which raises the cost of a wrong answer, and a face that looks like a real employee or executive raises its own disclosure questions. Most published deployments are pilots with no disclosed outcome, which is the reason to start narrow and measure.
How does it work?
- Greet and identify the topic. The avatar appears on a website, in an app, on a kiosk or on a screen in a meeting, greets the visitor and takes a question by voice or text. In the published pilots it listens through the device microphone and, in some, watches through the camera to react to the person. Reading a customer's emotions from the camera is high risk under the EU AI Act, so treat the camera as an optional feature that needs its own assessment.
- Answer from approved content. The avatar answers from the institution's own knowledge base of products, fees, policy wording and procedures, in the customer's language. ANZ limited Jamie to 30 frequently searched help topics and no customer data; Ajman Bank's avatar answers from the bank's knowledge base in Arabic and English.
- Explain and guide. It explains terms in plain language, shows where to find a form or feature and walks through a journey such as a mortgage inquiry or a motor insurance quote.
- Act only within limits. Avatars differ in what they do. ANZ's Jamie answered general questions only. UBS tested one in 2018 that could order a credit card for a client in a meeting. Shinhan Bank's AI Banker at its branch digital desks opens deposit accounts, takes loan applications and issues debit cards and certificates, with a video handover to staff. Anything that needs an account lookup or a money movement needs authentication and an allow list.
- Hand over. Complex, sensitive or unanswered questions go to a person with the conversation so far, such as a client advisor, a contact center agent or an appointment.
- Audience
- Customer facing
- Autonomy
- Supervised agent
- Adoption
- Emerging
- Channels
- Digital human, Web chat, Mobile app, Kiosk and branch
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Interactions handled | Not pooled | 12,000 to 13,000 | 2 | 1 organization, 1 vendor |
| Accuracy | Too few to pool | 94.5% | 1 | 1 vendor |
Value drivers: Customer experience, Inclusion and access, Lower cost to serve, Revenue growth.
Indicative value
A retail bank with 2 million help site and app sessions a year
USD 4800 to USD 86,400
Staff handled contact cost avoided per year
How this is calculated
Formula: sessions * engagement * answered * replaced * costPerContact. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Help site and app sessions per year sessions, sessions | 2,000,000 | 2,000,000 | The reference bank. |
| Share of sessions that start a conversation with the avatar engagement, fraction of sessions | 0.01 | 0.03 | Editorial assumption, replace with your own. For scale, ANZ New Zealand reported more than 12,000 conversations in the first 100 days of Jamie. |
| Share of conversations the avatar answers without staff answered, fraction of conversations | 0.4 | 0.6 | Editorial assumption, in line with ANZ New Zealand, which said Jamie could answer about 60% of customer queries in her first 100 days. ANZ counted all queries, including off topic ones, so that figure is not a resolution share. |
| Share of answered conversations that would otherwise have reached a staff handled contact replaced, fraction of answered conversations | 0.2 | 0.4 | Editorial assumption, replace with your own. Many visitors would have left or found the answer on the page. |
| Cost of a staff handled contact costPerContact, USD per contact | 3 | 6 | Editorial assumption for a blended chat and phone contact. Replace with your own fully loaded cost. |
What it leaves out: Gross avoided contact cost only. It leaves out avatar production and hosting, content upkeep, the revenue effect of more engaged visitors and any lift in leads or quotes, which published pilots have not reported.
Who already uses it?
8 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
ANZ New Zealand
New Zealand · Banking · 2018
ANZ New Zealand put a digital assistant called Jamie, built with Soul Machines, on its help site in July 2018. Jamie has a human face and voice and was set up to answer general banking questions on the 30 most frequently searched topics, without asking for any customer details. ANZ ran it as a trial to see whether it would appeal to customers who are less comfortable with its other digital channels. After the first 100 days ANZ reported more than 12,000 conversations, said Jamie could answer about 60% of the queries put to her, and extended the pilot.
- Interactions handled: at least 12,000, first 100 days of the pilot
"Jamie has already had more than 12,000 conversations with existing and potential customers."
Claimed by: organization
UBS
Switzerland · Wealth and asset management · 2018
In 2018 UBS ran a three month experiment called UBS Companion at a branch in Zurich. Client advisors in Wealth Management Switzerland brought two avatars into meetings with a set number of clients, shown on an LCD screen: Fin, a digital assistant that can handle simple banking transactions such as ordering a credit card, and Daniel Kalt, a digital version of the bank's Regional Chief Investment Officer Switzerland that supports market topics and provides CIO House View content on demand. UBS built them with IBM and FaceMe to test how clients accept digital assistants and how human like they should be. The page reports no results.
No outcome disclosed.
Ajman Bank
United Arab Emirates · Banking · 2026
Ajman Bank, an Islamic bank in the UAE, unveiled a generative AI avatar at GITEX Global in October 2025 and runs it with SESTEK across its mobile app, in branch kiosks and HoloBox holographic displays. Customers speak to the avatar in Arabic or English, and it answers from the bank's own knowledge base using speech recognition, a localized Emirati text to speech voice and generative question answering. SESTEK reports more than 13,000 customer interactions in the first four months, with content generation accuracy of 94.45% in Arabic and 96.7% in English.
- Interactions handled: at least 13,000, first four months
"Within the first four months, the avatar supported over 13,000 real customer interactions."
Claimed by: vendor - Accuracy: 94.5%, first four months, Arabic content generation
"It achieved a 94.45% content generation accuracy in Arabic and 96.7% in English, alongside a 4.3 customer satisfaction score."
Claimed by: vendor - Accuracy: 96.7%, first four months, English content generation
"It achieved a 94.45% content generation accuracy in Arabic and 96.7% in English, alongside a 4.3 customer satisfaction score."
Claimed by: vendor
Kuwait Finance House
Kuwait · Banking · 2025
Kuwait Finance House (KFH) and NCR Atleos announced in November 2025 a conversational AI avatar named Fahad, shown as a video avatar on the account service kiosks in KFH branches. The avatar answers frequently asked questions about KFH services and gives branch specific information, and from the second phase of the project it is meant to guide customers through transactions on the kiosk. It can hand the conversation to a live agent. KFH also plans AI speech analytics and automated quality management over these conversations. NCR Atleos describes the work as a field test of a product that is not yet commercially available. No usage or results have been published.
No outcome disclosed.
Shinhan Bank
South Korea · Banking · 2024
Shinhan Bank, in South Korea, put an AI Banker, a conversational digital human, on the video consultation "digital desks" of its unmanned Digital Lounge branches from September 2021. At the desks it handles dozens of consultation tasks, such as opening deposit and savings accounts and taking credit loan applications, and hands the customer to a remote staff member over video when it cannot finish a task.
In June 2024 Shinhan raised the task list from 56 to 64, added debit card, security card and certificate issuing, and extended the AI Banker to smart kiosks, where it guides customers who are not used to operating them. A July 2024 version uses five real branch employees as models, generative AI for the voice, and covers guidance on about 70 tasks at the kiosks. Shinhan said it had about 150 digital desks in branches nationwide in June 2024.
No results such as share of tasks completed or customer satisfaction have been published.
No outcome disclosed.
KB Kookmin Bank
South Korea · Banking · 2022
KB Kookmin Bank, in South Korea, installed AI Banker kiosks in selected branches from early 2022, including its digital test bed branch in Yeouido, Seoul. The AI Banker is a digital human built with DeepBrain AI from the voice and video of two branch employees.
It greets customers by voice, recommends savings products, explains financial terms, tells customers how to use the nearby automated teller machines, including issuing a debit card and changing a PIN at the bank's intelligent automated machines, and points to nearby facilities. A camera detects when the customer walks away and ends the session.
The reporter of a hands on test found that loans and account opening were not available at the kiosk and that detailed rates appeared on the screen instead of being spoken. No usage or satisfaction results have been published.
No outcome disclosed.
Suncorp Group
Australia · Insurance · 2021
In June 2021 Suncorp launched Ava, a digital employee with a face, as a pilot for some motor insurance customers of its AAMI brand. Customers talk to Ava through their microphone and camera while they scope a motor policy on the AAMI site, and she was trained to answer questions on coverage, price, discounts, excess, optional extras, claims and payment options. A selection of desktop users was randomly placed in the new experience with the option to interact with her. Suncorp built Ava with Soul Machines in four months and had set no end date for the pilot. No results were disclosed.
No outcome disclosed.
NatWest Group
United Kingdom · Banking · 2018
In February 2018 NatWest announced advanced testing of a digital human version of its Cora assistant, built with Soul Machines. The prototype lets customers hold a two way spoken conversation on a computer screen, tablet or phone and answers basic questions such as how to log in to online banking, how to apply for a mortgage and what to do about a lost card. NatWest said it would deploy the technology only if the pilot succeeded, and early testing suggested that customers who had avoided digital services might be more willing to talk to a digital human. No results were published.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Approved product, fee and policy content with an owner and a review date per document
- The top help topics from site search and contact center reasons, to scope the first topics
- A style guide for tone, name, appearance and what the avatar never says
- Consent and recording rules for microphone and camera use
Systems to integrate
- Website or app embed with microphone and optional camera permission
- Knowledge base with retrieval over approved content
- Contact center, chat or appointment booking for handover
- Authentication and core system APIs, only for later transactional phases
- Analytics for conversations, unanswered questions and satisfaction
Complexity: Medium
A general question avatar needs approved content in a knowledge base, a speech and avatar rendering service, a web or app embed, and a handover to chat or a branch. Anything that touches an account adds the bank's authentication and APIs, and a kiosk or in meeting screen adds hardware, microphones in a noisy room and accessibility checks.
- 1
Choose a narrow first scope
Start with general questions only, as ANZ did with 30 help topics and no customer data. Pick the topics from site search and contact reasons and leave account actions for a later phase.
- 2
Design the persona and the disclosure
Decide the name, appearance, languages and voice, and say on screen that the customer is talking to an AI. If the likeness is a real employee or executive, get that person's written consent and legal review before launch.
- 3
Ground every answer
Load approved content into a knowledge base, make the avatar refuse when the content does not cover the question, and review the questions it could not answer each week to fill gaps.
- 4
Design the handover
Define which topics and signals go to a person (complaints, vulnerability, advice that needs a licensed advisor, repeated misunderstanding) and pass the transcript to the advisor, contact center or booking.
- 5
Test, then release to a share of visitors
Run a set of real customer questions in every language before launch, then show the avatar to a random share of visitors, as Suncorp did with desktop users, and compare conversion, satisfaction and contacts with a control group.
- 6
Widen the channels and the actions
Add the app, kiosks and advisor meetings once the web pilot holds, and add transactions one at a time, each with its own authentication level and limits.
Guardrails
- On screen disclosure that the customer is talking to an AI, at the start of every session
- Answers only from approved content, with a refusal and a handover when the content does not cover the question
- No financial or insurance advice beyond what the institution's licensed advisors have approved
- No emotion recognition or other biometric inference from the camera feed unless it has passed a separate legal assessment
- Microphone and camera input processed only for the conversation, with a clear consent prompt and short retention
- Written consent and legal review for any avatar that uses the likeness or voice of a real person
- PII masking before text reaches a model, and an allow list with step up authentication for any account action
KPIs to instrument
- Conversations per week and the share that end with a staff handover
- Share of questions answered from approved content, from a weekly sample check
- Customer satisfaction after the conversation, and complaints that mention the avatar
- Conversion, quote or appointment rate against a control group without the avatar
- Cost per conversation including rendering, hosting and speech services
Human in the loop
Staff own advice, complaints, vulnerable customers and every action that moves money or changes a policy. A content owner reviews the avatar's unanswered and low rated conversations each week, and compliance approves every new topic and every new action before it goes live.
Common failure modes
- A confident face on a wrong answer
- A lifelike avatar may make an incorrect answer more persuasive. Ground answers in approved content, refuse outside it and sample conversations every week.
- Novelty without outcomes
- Interest in the avatar is high at launch and the pilot never measures conversion or contacts avoided. Set the control group and the KPIs before the launch.
- Uncanny or unwanted face
- Some customers dislike talking to a lifelike face or have no microphone. Keep a text option and a human route, and test the persona with real customers, as Suncorp did with the name Ava.
- Vendor and platform risk
- Several early avatars ran on one vendor's platform, and Soul Machines was reported in April 2026 to be in receivership. Keep the content, prompts and persona assets under your own control so you can move to another platform.
- Likeness and consent
- An avatar modeled on a real employee or executive needs consent, a plan for when that person leaves, and a clear label that it is AI.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
Article 50(1): people must be told they are interacting with an AI system unless that is obvious from the context. Camera based emotion recognition of customers, which some of the avatars described here do, is high risk under Annex III point 1(c) and needs the Article 50(3) notice, so leave it out of the design or put it through a separate assessment. An avatar that mimics a real employee or executive may also be deepfake content under Article 50(4), which requires disclosure that the content is artificially generated. General product and service answers are not an Annex III use; the tier would rise if the avatar assessed creditworthiness or priced life or health insurance.
Rules that apply
Guidance
- Article 50, transparency obligations for providers and deployers of certain AI systems (European Union, Europe). People must be informed that they are interacting with an AI system, and deepfake content must be disclosed as artificially generated.
- Chatbots in consumer finance (Consumer Financial Protection Bureau, North America). Warns that chatbots which cannot resolve a problem and block access to a human can harm customers and lead to violations of consumer financial law; the same applies to an avatar front end.
Controls to put in place
- AI disclosure at the start of every session and a visible label on the avatar
- Inventory entry for the avatar with an accountable owner and a documented list of topics and actions
- Consent record for any real person whose likeness or voice is used
- Audit trail of conversations with a defined retention period and no biometric profiling of customers
- Change control and regression tests for every new topic, language or action
When it went wrong elsewhere
- Soul Machines placed in receivership. NBR reported on 13 April 2026 that Soul Machines owes creditors more than $19.5 million (NZ$) in receivership. Soul Machines was the platform for ANZ's Jamie, NatWest's Cora and Suncorp's Ava, according to the deployment records. The report does not say what happened to those deployments.
Frequently asked questions
- What is a digital human in banking and insurance?
It is an AI agent shown as a realistic person on screen, with a voice, facial expressions and the ability to answer questions in a conversation. ANZ New Zealand, Suncorp's AAMI and Ajman Bank have each run one to answer customer questions on a website or in an app.
- Does a digital human perform better than a text chatbot?
The published evidence does not show it yet. ANZ reported more than 12,000 conversations in the first 100 days of its pilot, and Ajman Bank's vendor reports more than 13,000 interactions in four months, but none of the pilots compares the avatar with a chatbot on resolution, satisfaction or conversion. Measure that against a control group before you scale.
- Does a digital human have to say it is AI?
In the EU, yes: Article 50 of the AI Act requires people to be told they are talking to an AI system unless that is obvious, and an avatar that imitates a real person may also need a deepfake disclosure. Show the label on screen from the first second.
- What can a digital human not do yet?
It depends on the avatar. ANZ's Jamie answered general questions only and did not ask for any customer details. UBS tested an avatar in 2018 that could order a credit card, and Shinhan Bank's AI Banker at branch digital desks handles account opening and loan applications, with a video handover to staff. Advice, credit decisions and payments belong with licensed staff or with systems that authenticate the customer.
How to cite this page
Blits.ai AI Use Case Library, "AI digital human guide for banks and insurers", last verified 14 October 2026, https://www.blits.ai/ai-use-cases/financial-services-digital-human-guide. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 14 October 2026: First published