Real-life cases

Where it has alreadyworked.

Four real-life engagements at large global firms, led by Abhinav Gaurav. Each one started with a ceiling the institution could not see past. Here is what changed, and which Antara solutions and instruments map to the work.

At a glance

The challenge, and the solutions behind it.

Problem	Solutions and instruments
Financial crimeGlobal bank	Sentinel · Secure Bearing · Prove HALO · Lead
Credit decisioningMajor UK lender	Keel · Govern Bearing · Prove HALO · Lead
ML platform and generative AIMajor UK retail bank	Keel · Govern Soundings · See HALO · Lead
Fraud intelligenceGlobal payments company	Sentinel · Secure Squall · Test HALO · Lead
Consumer goods and retailPerspective	Soundings · See HALO · Lead

All solutions →All instruments →

Financial services

Four engagements, four challenges.

Financial crime, credit decisioning, enterprise machine learning and real-time fraud. The problems differ in every respect except one: in each case a rules-based or fragmented estate had become the ceiling on what the institution could do.

Global bank · Financial crime

Replacing rules-based AML alerting with a risk-scored model

The problem. More than 95% of anti-money-laundering alerts were false positives. Investigators spent their time on noise while genuine crime stayed hidden.

60%Reduction in false-positive alerts

2 to 4xMore genuine financial crime detected than the previous system

The situation

A global bank processing billions of transactions each month relied on static, manually defined monitoring rules. Investigators spent most of their time on cases that led nowhere, and genuine criminal activity was harder to see against that noise.

The approach

Replaced rules-based alerting with a machine learning system that produces dynamic customer risk scores in place of binary alerts.
Trained the models on the bank’s own transaction history, customer records, network behaviour and confirmed criminal typologies.
Built explainability and an audit trail into every output, so that regulatory defensibility was a design requirement from the outset.
Deployed in phases, starting with the highest-volume markets, on an architecture designed for global rollout.

Major UK lender · Credit decisioning and AI governance

One decisioning layer across lending, with governance built in

The problem. Three lending platforms with no shared logic. A credit policy change took at least two weeks, and conduct risk was rising.

50%Faster policy change cycles

Real-timeCredit limit adjustments

5Products migrated within 12 months

ZeroOperational downtime

The situation

Mortgages, loans and cards each ran on a separate decisioning platform, with no shared logic or data layer. Credit limit adjustments were manual, and an ageing, fragmented estate was increasing conduct risk as regulatory expectations tightened.

The approach

Consolidated decisioning onto a single platform that unifies logic, data and risk models across lending products.
Designed the governance architecture alongside the platform: an AI council, a Responsible AI framework and an AI centre of expertise.
Migrated product by product, keeping the live service running throughout.

Major UK retail bank · ML platform and generative AI

Unifying hundreds of models on one platform, then scaling the value

The problem. Hundreds of models in production on a decade-old platform that capped what the bank could do next.

Days to secondsMortgage income verification

Minutes to secondsFrontline knowledge retrieval

£50MAI-attributed value in a single year

The situation

The bank had invested heavily in AI and had hundreds of models in production, but they ran on a decade-old on-premise platform nearing end of life. The data science community was working across fragmented environments.

The approach

Platform. Migrated the full machine learning estate to a modern cloud platform and brought the data science teams onto one environment, with no unplanned downtime.
Customer. Deployed ML-based income verification for mortgage applications, replacing a multi-day manual review.
Colleagues. Built a generative AI assistant drawing on thousands of authorised internal knowledge articles for frontline staff.
Scale. Established an AI Centre of Excellence combining engineers, data scientists and AI ethicists, supported by a bank-wide training programme.

Global payments company · Fraud intelligence

Real-time fraud decisioning at global card scale

The problem. Billions of fraud decisions a year, each due in milliseconds, with customer trust at stake on every one.

50xFaster than the CPU baseline

10%Engineering time saved

MillisecondsDecision latency at global scale

The situation

Billions of authorisation decisions each year, each required within milliseconds. A slow decision fails the transaction, frustrates the customer at checkout and costs the company trust as well as revenue.

The approach

A proprietary deep learning fraud model combined with gradient boosting, running on GPU infrastructure to meet latency requirements at global volume.
A generative AI developer copilot, grounded in internal codebases, deployed across software engineering teams.
A copilot for customer-facing advisers, so that the person on the other end of the call is better briefed and faster.

What the mandates share

Five patterns that appeared in every engagement that worked.

Absent in most initiatives that stall. Use them as a diagnostic on your own AI programme.

01Data over rulesRules set a ceiling. Learning systems keep raising the floor.

02Governance as a design requirementBuilt into the architecture before production code.

03Platform before productThe first investment was the foundation. The products followed.

04Human and machine togetherNo deployment removed people. AI worked best as a colleague, not a replacement.

05Speed of change is an AI outputPolicy cycles shortened from weeks to days, and verification from days to seconds.

Perspective · Consumer goods and retail

Five decisions shaping AI readiness.

This section is a diagnostic lens and does not describe a delivered engagement. It sets out decisions that consumer goods and retail organisations are taking now, often without framing them as AI decisions, and that fix how ready they will be.

Decision 1

Data foundation

Customer data typically sits across many disconnected systems. A platform selected today shapes the AI roadmap for the years that follow, so the platform choice and the AI readiness choice are the same decision.

Has the organisation mapped how fragmented its customer data is, and what that means for the AI use cases it plans next?

Decision 2

Platform renewal

CRM and marketing platforms are adding AI agents, generative content and individual-level personalisation as standard. Renewal commits an organisation for several years, so the AI question belongs in the renewal conversation and not after it.

Have customer journeys been redesigned for AI agents, or are the same processes simply being automated faster?

Decision 3

Product data and agentic commerce

AI shopping agents read product data as structured information. Incomplete attributes and inconsistent tagging make a brand invisible to them, and data quality has to be maintained across every channel an agent might read.

If an AI agent were shopping this category today, would the brand appear in the results, and who is accountable for what the agent chooses?

Decision 4

Workforce preparation

Most AI programmes stall because people are not ready, senior leaders included. The usual defaults are resistance or redundancy. The more productive path redeploys people towards judgement, relationships and creativity.

Are HR and internal communications part of the programme, or only technology and marketing?

Decision 5

AI governance and the EU AI Act

Many organisations lack a complete inventory of the AI systems already running in their business, including those embedded in vendor platforms for loyalty, personalisation, pricing and audience scoring. The timetable for high-risk obligations under the EU AI Act has been extended, and the obligations themselves remain. A short readiness and governance assessment is a low-risk place to start.

Does the organisation hold a full inventory of its AI systems, classified against the EU AI Act risk tiers?

Get in touch

Facing a similar challenge?

Most conversations begin with a diagnostic of what is already running. Start with Soundings, or get in touch directly to discuss advisory work or an AI readiness conversation.

Get in touch →Begin with the Soundings diagnostic →