2026.09.29 | TB | Software | Enterprise AI Expert: Wells Fargo Cloud

Client-led • 28 Jul 2026

Companies: MSFT US, IBM US

Expert Former Senior Executive, Cloud Transformation at Wells Fargo & Co

Agenda

• Around 30 years' experience across the financial services, healthcare and enterprise IT industries, focusing on platform engineering, cloud architecture, cybersecurity, modernisation and risk management
• In-depth understanding of AI-driven cybersecurity, including zero trust architecture, identity security, AI security products, vendor evaluation, pricing considerations and future enterprise security investment priorities
• Well-positioned to discuss enterprise technology procurement and vendor management, with experience overseeing sourcing, contract negotiations, legal reviews and strategic technology partners' onboarding
• Knowledgeable on enterprise AI governance and agentic AI deployment, including centralised AI decision-making processes, governance frameworks, use-case prioritisation and cross-functional stakeholder management

Introduction

Client [00:09:15]: I cover the global tech sector. I spent most of my time on AI, cloud and software. I want to understand what is going on inside large enterprises, especially in the financial services industry. Where is the AI money going? What's going on? What's working and what isn't?

Questions

00:11:18] For companies, is the AI decision more centralised?

Expert [00:11:32]: Yes. It's a centralised governance function. The decisioning from the different use cases, different models, it could come from various line of business, let's say, retail, consumer lending, wealth management. It's all funnelled through one centralised portal. the governance board look into each and every AI specifically use case, vet them and understand what is the business use case, what is the business rationale, what's the business outcome.

00:18:04] What's the percentage of the AI budget to the overall IT budget?

Expert [00:18:25]: In the banking, we have run the bank, change the bank. Change the bank is where the AI initiative falls into. Change the bank is not just all AI. I would say right now, for the total IT budget, it's probably less.

00:19:06] Let's talk about the growth of the AI budget over the last half a year. If the gen AI spend at the start of the year was 100, what is it now? I want to understand the growth of the gen AI spend.

Expert [00:19:52]: It's a steady growth.

00:22:54] There are two views on tokenising. One, for enterprises still, it's too early in AI deployment. They don't need tokenising now. Only the AI first movers have to control the token cost. Another view on that is that for the enterprise, they also need to discuss that. What's your view on that?

Expert [00:23:34]: Right now, there are multiple variables when we're talking about overall token, how we're consuming it. Basically, the token economics, where the data is attributing the capture today, where they are left. Is that what you're asking? What's the attribution, where is the driver?

00:24:08] Do you think it's a concern to control the token billings?

Expert [00:24:27]: Actually, it is. As I was mentioning, because the cost from the token is going up and up in the various channels, the way it has been consumed, the way it's been allocated. Should we CAPEX, should we OPEX? What can be tangible assets? If you're writing, using tokens to build or rewrite a code, that is actually proprietary to bank. That is a tangible asset. Definitely, that's a concern how it's changing the landscape and how the developer community and end-users consume and what's going in the prompt, and how the output is coming from the prompt.

Even today, if you look at all the different SaaS providers, they're also changing their model from how they went from perpetual to subscription. Now they're talking about using some token consumption model. The whole industry is changing in this space.

00:39:19] Can you can break it down into the LLM models' providers and the remaining data infra investment, or Palantir? How do you break it down?

Expert [00:40:01]: If I take for model where we're talking about all the token consumption and everything, I think it's 80/20 ratio, 20% is actually the building, supporting, but the 80% goes in the leveraging those model expense. That's how you should be looking into it. I'll give you an example. Let's say, I'm using OpenAI model, but we are using the platform to host it, the resources to manage and sustain it. Those probably is less than maybe 20%. The actual model, the token, the training, the inferencing, that is actually a significant portion, which is 80%. It is higher than actually 80%, but yes.

Client [00:40:48]: Is that the cloud providers or the deployment cost of building the AI agent?

Expert [00:41:08]: Yes, exactly. Let me walk you through it. The way it works is you need to have an AI platform before you start hosting a model. I'll give an example. When you're working with Microsoft hyperscaler to build the Azure AI Foundry, it's a platform. There is a platform cost to manage and maintain. That cost we pay to Microsoft. Then there are all the engineers supporting and building, security goes to platform. Less than 10-20%. Once the platform is done, now you're actually going to subscribe to OpenAI, Codex, Claude Code, Cowork, everything through that on the platform which you built. Now you're going to start paying to those specific model, and that is actually 80%. That's what the whole 20% is the building, the landing zone, building the platform to host those model, paying to the hyperscalers to manage, provide security of that platform where you're planning on hosting it.

00:46:59] Can you walk me through a typical agentic AI project? What's the whole process? What are the key steps? How long does it take for the whole process from the proof of concept to production?

Expert [00:47:24]: It's a journey. As I mentioned, every use case, let's say, a specific line of business has a use case coming through. It goes through demand intake planning. Once it goes through intake, enterprise governing committee or governing board review that use case, there's all the different stakeholders, security part, the line of defence. Let's say that with the approval goes through, let's say, one month. Once the use case gets finalised, prioritised, then various things get assessed and goes in the actual build stage. The build stage is through the low environment of test and dev, which probably will take maybe one month or so, depending upon the use case, of course, how complex. Maybe we already have all those approved services. If you need any net new service, then it has to go through committee for an approval, that will add maybe one month or so. Once that is done, then piloting, post that test and then goes to the QA.

Overall, I would say it's 3-6 months, give or take, depending upon the level of involvement to building that use case. Some use cases, like agentic is a very new capability. It's just how the use case being complex and then some use case is just the gradual progression, like expanding the use case across various departments. Give or take, it's 3-6 months runway from ideation, submission to actually getting something in a pilot or MVP.

00:50:38] How long does it take from the limited production to the scale production?

Expert [00:50:47]: It's case by case. It's not just launching something. In production, make sure that we have a solid operationalisation. We have a support system built. If in case something goes wrong, then we can entertain the call, we can fix it. That's why actually it's case by case, but scale gradually from there on.

00:56:13] We know the engineer has an AI coding agent to use. What about departments beyond engineering? Do they have access to frontier models such as Claude Code, Codex or ChatGPT? Do they have access to that?

Expert [00:56:59]: Actually, yes. When we say, it's an approved model then it doesn't mean that it's available for everyone to use. They have to still go through intake and then understand what is the use case, why they need this and what they're trying to achieve with that model. Then once that is done, then team will make that model available for their consumption because there's a cost associated with that. When I say approval, it means it's approved for consumption, but also it's actually based on your specific need. That's how the whole intent is case by case, what's being available for them at disposal and what's actually for limited use but require approval.

Client [00:58:20]: Workers beyond engineering have access, but they need additional approval for that, so most of them don't use that in their daily work, right?

Expert [00:58:41]: Correct. Specific division, if they have data scientists, if they have the data modeller, if they have data analytics. Scientists today, they definitely will have access to those specific models.

01:04:08] Can you elaborate on the volume of logs going up?

Expert [01:04:25]: It's the way you're consuming the logs. That cost is doubling, quadrupling over time because we consume a lot of logs. The logging cost is actually directly proportional to what's your logs coming in from all these AI models, tooling and everything.

01:06:36] Within cybersecurity, which subsector do you think is the top priority? Which sector gained the most momentum?

Expert [01:06:47]: It's actually the modern security. All the models which are actually being built is actually driving the cost because we are inspecting, assessing that model and how it's been communicating, who's actually been consuming it, how the data is transferring. That's what's driving the logs. If you're communicating from XYZ agent to customer, back to the agent. That communication needs to be logged for discoverable, for forensic. You see all those communication has grown significantly, and that is being logged. These are the biggest driver, I would say.