← All case studies

    Case Study

    Capgemini: building enterprise voice agent prototypes on OmniDimension, from expert matching to insurance service

    How Capgemini's architects used OmniDimension to show banking, insurance and capital markets clients what a voice agent does inside their own processes, and what the 2ANT expert-matching project taught them about routing calls on intent.

    Published November 16, 2026

    Capgemini logoOmniDimension logo

    At a glance

    Customer
    Capgemini, a global consulting and technology services company
    Who built it
    Capgemini architects and engineers, building client prototypes
    Domains
    Banking and cards, insurance service, retirement plan data, wealth and relationship management, contact center operations
    Flagship project
    2ANT, a voice agent that finds the right volunteer expert for a caller and transfers the call
    Features used
    Custom APIs, caller authentication, dynamic call transfer, knowledge base, web search, webhooks

    The problem

    Capgemini builds systems for banks, insurers and capital markets firms. When those clients evaluate voice AI, a scripted demo does not answer their real questions. They want to see an agent check who is calling, read their own records, follow their rules, and hand the call to a person when it should.

    Each client works in a different domain with different data. So Capgemini's teams needed a platform where an architect could build an agent for one client's process, connect it to that client's data through an API, and put it on a phone line, then do the same for the next client.

    What Capgemini built

    Capgemini's teams built prototypes for a range of client scenarios on one OmniDimension account:

    • Credit card service for several demo banks. The agent verifies the customer against a customer database and logs issues to Airtable.
    • Insurance service. Agents handle policy details, payments and renewals, claims and document requests.
    • Life insurance policy changes. The agent verifies the policyholder, reads the policy, quotes rider options from a premium table, and works out the new annual premium.
    • Retirement plan data. "Unified Data Chat" answers questions about plans, participants, loans and contributions through five live API lookups.
    • Wealth and relationship management: a wealth advisory associate and agents that book relationship manager appointments.
    • Contact center operations: a metrics assistant, a contact detail updater, and agents that search the web for current answers.
    • 2ANT, the expert-matching agent described below.

    2ANT: routing a caller to the right expert

    2ANT connects a caller with a volunteer expert who knows their topic. It works like this:

    1. The caller describes the problem in their own words. The agent asks follow-up questions until the need is clear.
    2. The agent sends the conversation to a custom API, "Find Volunteer By Transcript".
    3. Capgemini's matchmaking server extracts the intent and turns it into an embedding. It compares that with embeddings of each volunteer's skills, stored in Supabase.
    4. The server ranks volunteers by how close their skills are to the request, then by rating and availability, and returns the best match with a phone number.
    5. The agent tells the caller who the expert is and their rating, and asks if they want to be connected.
    6. If the caller says yes, OmniDimension's dynamic call transfer dials the number the API returned.
    7. If no volunteer is available, the agent offers to help from its own knowledge base on data engineering topics such as dbt and Data Vault 2.0.

    The matching logic stays in Capgemini's own service. OmniDimension handles the conversation, the API call and the transfer. Read more on connecting voice AI to your systems with APIs and webhooks.

    What the projects taught us

    1. The caller's words can do the routing

    2ANT has no menu. The caller explains the problem, and the matchmaking service picks the person. Routing on meaning removes the step where a caller has to guess which option fits.

    2. Ask before you transfer

    The agent names the expert and their rating, then waits for a yes. A caller who knows who they will speak to is ready for the conversation.

    3. A fallback is a prompt change

    One version of 2ANT tells the caller that no expert is free. Another version offers help from a knowledge base instead, but only after it tries to find an expert. The difference between them is a prompt section and a knowledge base, not new infrastructure.

    4. Verification is a gate, not a question

    The insurance agents verify with a policy number first, or with a full name and the last four digits of the SSN. After three failed attempts the agent stops and offers a human specialist. No policy detail is shared before verification passes.

    5. Refusing beats guessing

    One insurance agent answers only from verified data. If it cannot confirm an answer, it says one fixed sentence instead of a guess. For regulated clients, a predictable "I can't help with that" is safer than a confident wrong answer.

    6. Voice has its own details

    2ANT keeps replies under 20 words. It says a rating of 4.5 as "4 point 5", and it says volunteer names slowly. After the volunteer hangs up, the agent comes back on the line to close the call.

    7. A new use case is a new prompt and a new API

    The same building blocks served cards, insurance, plan data and expert matching. Each new client scenario needed a prompt and a connection to that client's data, not a new platform.

    Limits worth knowing

    These were prototypes, and the banks and insurers in them are demo brands. This case study describes how the agents were designed and how they behave on a call. It does not report production results.

    Lessons for enterprise teams evaluating voice AI

    1. Prototype on your own data. Connect the agent to a real system through an API before you judge it.
    2. Keep business logic in your own service. A custom API lets the agent use your rules, such as your ranking logic.
    3. Put verification first. Set a retry limit and a human path for callers who fail it.
    4. Decide what the agent says when it does not know. Write the exact sentence.
    5. Plan the fallback for every hand-off. If no person is free, the caller still needs a next step.

    Related reading

    FAQs

    See what OmniDimension can do for your leads

    Book a 15-minute walkthrough with our team.

    Book a Demo