Enterprises are Writing Off AI Proofs of Concept. Not So Fast, Says Axis Bank

At a time when enterprises are moving beyond AI proofs of concept (POCs) and focusing on scaling successful deployments, Axis Bank is making a different argument: POCs remain indispensable, provided they are tied to a clear path to production.

In an interaction with AIM, Prasad Lad, who heads the bank's Business Intelligence Unit, says the lender has already moved several AI initiatives into production. However, he cautions against abandoning the experimentation stage altogether.

"Our philosophy is that enterprise is the place for this to go. But there is no path to enterprise scaling till you have done the POC right," Lad notes.

His remarks come amid a growing debate in the industry over whether companies should stop running large AI pilots and instead concentrate on scaling a smaller set of proven use cases. Lad asserts that the challenge is not POCs but the failure to convert successful pilots into production.

"There is no joy in doing thousands of POCs and not taking anything to production. But there is also no joy in just productionising stuff without doing the POC," he shares.

According to Lad, Axis Bank follows a structured framework under which every POC is evaluated for efficiency, efficacy, and accuracy before it is scaled.

"You need a healthy pipeline of POCs with very clear measurements on saying when is the POC ready for production," he notes. 

For a regulated institution such as a bank, Lad argues that validation cannot be skipped. "We are not sitting in Silicon Valley. We are not in the business of token maximisation. We are in the business of completely optimising the use case and making it live only when it makes economic sense."

Three-Level AI strategy

Axis Bank has established an Enterprise AI Centre of Excellence (CoE) that develops the bank's AI roadmap, oversees governance, and works with business units to identify and prioritise use cases. The CoE has around 80 members and works alongside the bank's IT and Business Intelligence teams, which together comprise roughly 1,500 technology professionals.

The bank's AI strategy is organised across three levels.

The first level is what Lad describes as "citizen AI", focused on employee productivity. The initiative aims to equip employees with AI tools, such as Microsoft Copilot, to automate routine work, including document analysis, presentation creation, and spreadsheet tasks.

Axis Bank has identified approximately 3,000 middle managers for AI training. Every month, around 250 managers spend about 30% of their time learning productivity-focused AI tools to identify processes that can be automated independently.

The second level consists of enterprise AI use cases. These include knowledge management, software development, operations automation, speech-based applications, collections-related initiatives, sales enablement, and customer-facing AI deployments.

Among the most prominent deployments is ADI, Axis Bank’s internal knowledge platform, which serves as a search engine for employees seeking information on products, policies, and procedures. Lad says the platform has around 40,000 active users. The bank has also introduced AI-driven personalised training programmes that adapt learning modules to employee roles and performance.

The third level focuses on governance, optimisation, and model development. Here, the bank evaluates the effectiveness and economics of AI deployments, establishes governance and security controls, and works on model optimisation. A specialised 40–50-member team within the Enterprise AI CoE is also focused on developing smaller, fine-tuned language models tailored to Indian banking and to Axis Bank's operating environment.

"We are extremely focused on making sure that no use case that goes live goes live without having tested the efficiency and efficacy," Lad asserts.

The bank is now deploying AI in inbound customer interactions and collections-related speech applications. On outbound interactions, the lender remains in the POC phase as it assesses customer comfort with AI-led conversations.

Building AI-Ready Data Foundations

Axis Bank's AI ambitions rest on a multi-year effort to modernise and organise its data architecture. "We started this journey about three years ago," Lad recalls, noting that the work began before the emergence of generative AI.

The bank manages approximately 2.5 petabytes of internal data stored across about 35,000 tables. To make that information easier to consume and analyse, Axis Bank has developed a three-layer data architecture.

The first layer is the ingestion layer, which governs how data enters the organisation and includes controls designed to detect failures and data drift. The second is a curated layer comprising fact and dimension tables that structure information consistently across the organisation. The third is a semantic or glossary layer that provides business context and definitions.

According to Lad, the semantic layer became particularly important with the advent of AI because systems require significantly more context than human users do.

Over the past six to nine months, the bank has focused on strengthening this knowledge graph and semantic layer by comparing AI-generated responses against management information system outputs and refining the context accordingly. Lad claims around 90% of the bank's data estate is suitable for AI applications.

The Complete AI Stack

Underlying these initiatives is a three-layer AI stack that Axis Bank has developed to support enterprise deployments.

The first layer is the data layer, which includes structured and unstructured internal data, external information sources, internet-based data, and India Stack APIs. This layer is supported by the bank's semantic and knowledge graph capabilities.

The second layer is what Lad describes as the "harness" layer. Every AI use case runs through this layer, which selects the appropriate model, manages governance controls, optimises performance and costs, and determines whether large language models, small language models, or other models should be used.

"The harness layer is basically where the governance sits, the choice optimisation sits," Lad explains.

The third layer consists of the business applications and interfaces where AI is ultimately deployed, ranging from employee productivity tools and operations systems to customer-facing banking channels.

Production Deployments

Axis Bank has already moved several AI initiatives into production.

One of the most significant is a conversational management information system (MIS) platform that allows employees to retrieve information through natural-language queries rather than relying on static reports.

"We don't create any more new MIS," Lad reveals. "Everything now is people have to come ask the queries, get the thing, and move on."

The platform receives around 20 million hits each month, according to the bank.

Another production deployment is within the software development lifecycle, particularly in coding. Axis Bank has also used AI to accelerate the migration of legacy SaaS-based code to PySpark.

Partnerships Without Surrendering Control

Axis Bank is pursuing a combination of internal development and external partnerships.

The bank works with hyperscalers such as Microsoft for productivity-related applications while building some of its own capabilities for more complex use cases. Lad cites voice AI as an area where partnerships make practical sense.

"Voice AI, technically, we can build the multi-language model ourselves. But Sarvam has done it. The amount of effort it will require for us to do it is just not feasible to do it as a bank," he notes.

The lender is also engaging with startups in India and overseas and works with technology service providers on systems integration and deployment. However, he emphasises that the bank does not want to lose control over core AI decision-making.

"The choice of the models within the harness, the choice on the governance, etc, has to be owned by the bank," he reiterates.

He adds that Axis Bank seeks to avoid arrangements that force it to use a single model or require customer data to be controlled externally. "The core principle is you don't want to lose the IP."

Self-Funded AI Investments

Axis Bank has created a dedicated AI budget but is funding those investments through operational efficiencies rather than increasing overall technology spending.

The executive says AI initiatives are expected to begin contributing around 1–1.5% this year, with the impact rising to about 10% of P&L over the next two years. "We will not bloat the costs up to do this."

According to Lad, technology spending growth will remain broadly aligned with the bank's historical annual growth rate of 5–6%.

Over the next five years, Axis Bank aims to become a fully AI-driven institution.