Fost Singapore 2026 - The AI Iceberg: Navigating Hidden Costs to Reach Scalable ROI By Shi Mei Chin.

This talk explores the hidden costs associated with AI projects, highlighting the discrepancies between expected and actual expenses. The speaker discusses how projects often exceed budgets due to factors like inadequate data infrastructure, integration challenges, and talent acquisition costs. With a background as a CFO and auditor, he emphasizes the need for organizations to adopt a total cost of ownership framework before committing to AI investments. The speaker outlines five layers of costs beneath the surface, including data infrastructure, integration and migration, talent upskilling, governance and compliance, and ongoing maintenance. He advocates for a structured approach that involves mapping costs, stress testing business cases, and continuous budget monitoring to enhance financial accountability and project transparency.

Hello. Hi. Good morning, everyone. How are you doing today? Can I have a quick show of hands on whether any of you are up are building your own AI agent? Are you using Cloud, Cohere, Gemini, OpenAI? Well, imagine this. It's 12:00 a.m. in the morning. You are staring at your screen in the dark. You're jabbing into your keyboard, furiously typing your next prompt. You have a big

deployment tomorrow. You are in a constant state of flow, and you know that you can get this done. And then suddenly, this really annoying pop-up notification pop-up pops up that says, "Limit reached. Resets at 3:00 a.m." Frustrating, isn't it? Well, but at that moment, you you didn't think. You you just pay because you have to finish your project, and how much is a few more dollars to

get something important out of the way? That is exactly how AI costs increase without you even knowing it. Well, there's a kind of energy in the boardroom when AI gets put on the table. Words like transformation, efficiency gain, cost-cutting are bounced around. There's a lot of excitement um with this new toy that everyone has. Everyone is looking at the same glass, convinced it's half full. But, the

question is, what actually is in the glass? This is the AI iceberg. Above the waterline is what is visible to the naked eye. You see cost savings. You see revenue growth. Margins are improving due to better productivity and efficiency that AI brings on the table. However, the reality is, below the waterline are the real war stories that are untold. Late nights, extra costs, tied morale to get

things done. Those are the hidden costs behind AI. And we'll run through each and every one of these costs today. Some quick numbers. A study by Gartner says that 80% of AI projects fail to deliver their projected ROI. 20% of which are outright failures. We are expecting to spend about 2.5 trillion this year by end of 2026. That's a 44% jump from last year. Another research tells

us that budgets that were originally planned for AI business case end up two to three times more expensive. Two to three times. And the question is, have these been budgeted in? So, why I see this, or why am I so passionate about it? Well, firstly, I'm an accountant and an auditor. I used to work in the Big Four, in PwC. I live and breathe operations, workflows that

lead to numbers. I used to be a CFO of a fintech company and scaled the company from series A to pre-IPO. And I lived cost overruns firsthand. Lastly, and now, I am running my own CFO advisory firm, and I'm an ecosystem builder. I used to be in the Singapore Fintech Association, and I'm a judge, speaker, and a connector. So, I've not just been on the sidelines. heard,

seen a lot of stories out there. The real reality checks. Why this matters now. AI budgets are accelerating faster, way faster than finance discipline can keep up. Capital is being committed without a total cost of ownership framework. And as of now, while you're watching this video, someone within your organization is probably building up a business case that will be approved at the next board meeting. There's this

story of a series A fintech company. They had a compelling business case. Six months implementation, ROI within the first year, minimal additional headcount, drop-in integration. But, the reality is, it's been 18 months and counting. The ROI is projected to be negative by end of the year two. They realized that they need specialists, or people who have more experience in the space, and ended up with hiring three

new hires and two consultants. There is also a requirement for a full data pipeline rebuilt, as the originally tech stack could not support the And this hidden cost actually surfaced during the project implementation. There are five layers below the waterline. Data infrastructure, integration and migration, talent upskilling, governance and compliance, ongoing maintenance. We'll look at each one later today. Layer one, data infrastructure. I call the fundamentals, the

foundations. There is this saying, "Garbage in, garbage out." And that's how AI models live and breathe. They are only as good as their inputs. So, before you start an AI project, ask yourself, is your data model ready? Is it consistent, accurate, valid, relevant? Or are there any missing data? Because prevention is way better than the cost of correction. Then you have your pipeline engineering. Your ETL ELT

pipelines, your data lakes, your real-time streaming infrastructure. Are your data Is your data architecture segregated in such a way that it is in its most optimal and cost-efficient state? These are the soft questions that you should be asking your tech. Followed by storage and compute scaling. So, there is this saying, as a finance professional, we love to ask about variable costs. So, what do you mean by

variable costs? When transaction volumes go up, when businesses increase, operations increase, does your storage your cloud storage increase as well? Will your vendor be charging you more based on the volume of data that flows through your infrastructure? So, these aren't fixed costs. And the question is have these been accounted for throughout the project life cycle? Or are you just assuming that it's a one-time? The second layer

is the integration and migration. Legacy system runs on many, many applications, and AI needs to talk to them. Are your API versionings up to date? Do you have guardrails in place? Do you have your rate limits, your traffic rate limits? uh You have token limits. I have a story for you. Imagine this. If you have a token usage that is not limited you deploy uh a project

and the system is running. You have no control. the number ballooned up to $24,000. Is that something that you have budgeted for? Next is the migration risk. So, there is a few ways to do migration. Are you doing a straight cut or are you doing a pilot parallel run? If you're doing a parallel run, your cost is double because you're running on two systems. And the question

is how long will this pilot be for? Is your project ready before you kick off the pilot? Because if your parallel run ends up being more 3 months firstly, you will incur more costs. Secondly, your employees will be very, very burnt out because they'll be running two data sets at the same time and probably will affect the team morale. Next is uh talent and upskilling. So, you

have the people get, and this is the most overlooked get that people see or people miss. If you're bringing in a vendor, do you have anyone who's owning the project that has the additional capacity to work with the vendor to roll out the project? If not do you need to hire more? Or do you need to hire specialist for that certain project? And HR cost is the

cost that is mostly overlooked at. You have recruitment costs. Get a recruiter in and you have to pay them commission. You have to pay CPF additional 17% above from what you originally planned for. You have insurance costs. You have bonuses to pay. You have training costs. once you onboard these new employees, how long does it take for them to assimilate themselves within the organization before they can

work at maximum capacity? And lastly, change management. Do you have the buy-in of the team working on a AI project or an AI automation project that can possibly put themselves or put their jobs at risk? That is something to think about. Next is the governance and compliance. Regulations are always trying to keep up technology. So, have that considered into your budget. The cost of hiring a compliance

or a legal person to ensure that you have sufficient budget to address these workflows. Could be things like uh compliance reporting, reporting to MAS or having the right control framework so that you are compliant. sector-specific regulations as well like uh PDPA TRM framework or OSPA. And then lastly, being audit-ready. ensuring that you have the right workflows and documentation processes to your your your log is auditable. Because

if you work backwards and that's not available, that is also going to be incurring more man-hours, which eventually lead to more costs. You have ongoing maintenance usually for tech projects is about maybe 15 to 30% of your implementation cost has that been Another thing to think about is also whether is there continuity upon implementation. configurations would definitely be required as and when the technology evolves and you

have the right talent or the right vendor who can help you or stay around to improve to to continuously maintain that. And also, your backup, you have version How How much backup do you have? And how many versions of of models do you have that's taking up space in your cloud? Well, this is definitely a finance problem disguised as a technology problem. The CTO builds it. Architecture,

models, tech stack, infrastructure, or deployment. And the CFO funds it. You have to look after the budget, the ROI, the total cost of optimization, risk, and board reporting. Today, I want to show you the AI cost reality framework. Four steps to an honest AI business Number one, mapping. Mapping the iceberg, identifying all the five cost layers before you commit the capital. Number two, stress testing the business

When I say stress stress test, you have the best-case scenario, the the average scenario, and the worst-case scenario. Thirdly, continuous monitoring of your budgets. Because in reality most planning don't go according to plan and it is up to us to be agile and adaptable to update inform, and react when the time comes. Lastly, measuring what matters. Not doing it for the sake of doing it, but using

it as a guide to help you achieve what you have planned out for. So, the difference a framework makes in a nutshell, it builds credibility and it builds trust. There's no surprises in place. The risk have been priced in. And that And if there are any variances or differences, they are explainable. And that becomes your biggest asset. Having a cost framework to build credibility and trust. So,

the real question isn't whether to invest in AI but it's whether you know the nuances, the details that goes into the project. Knowing what actually is in the glass. So, I have a full framework a reports if you are interested. You can reach out to me. Um I think by the time you watch this video, this barcode uh is probably expired by then. But feel free to

reach out to me to get a copy of the AI cost reality check practical guide. It covers whatever we talked about today. Um it also includes a self-assessment checklist for your next AI business And I've also included a zero case study for you to have a look what happens when got reals have been taken off and the potential financial impact that it can have onto a company,

not just from the operations perspective, but from a stakeholder, shareholder, market valuation perspective. Well, thank you and feel free to drop me a hello on LinkedIn. You can ignore the feedback form as this was uh the form for the API Day conference. But on the right, you can see a barcode that leads you to my LinkedIn. So do connect with me and drop me a hello. Would

love to connect with you and stay in touch. Thank you.