What you’ll learn: AI in finance comes down to three functions: classification, prediction, and generation. This post walks through each one with concrete examples your team already touches.
The problem with the word “AI”
Ask five vendors what AI means and you get five answers. The word covers everything from a spreadsheet macro to a self-driving car. For a Hong Kong finance team, most of that noise is irrelevant.
Strip it down. AI does three things:
- Classify: sort data into buckets
- Predict: estimate a number or probability
- Generate: produce text, summaries, or drafts
Every finance AI application is a combination of those three operations. Nothing more. Let’s apply them.
Classification: sorting transactions, flagging exceptions
Your compliance team already classifies things manually. They look at a transaction and decide: is this suspicious? Is this a corporate action? Does this match the standing instructions?
AI classification does the same thing at scale. It reads transaction descriptions, counterparty names, and amounts, then assigns a label. The label could be “high-risk jurisdiction,” “duplicate invoice,” or “needs manager review.”
A wealth manager in Central receives 200+ trade confirmations daily across client accounts. An AI classification layer tags each one by asset class, currency, and exception status before any human opens it. The analyst reviews only the flagged exceptions, roughly 15% of the daily volume.
The SFC’s guidelines on AI usage require that classification decisions remain explainable. Senior management (MIC-IT) must understand why a transaction was flagged. This rules out black-box models. The classification logic needs to be auditable.
Prediction: estimating NAV variance and default risk
Prediction is where AI gets the most attention and the most skepticism. In a Hong Kong finance context, the useful predictions are narrow in scope.
A multi-custodian fund administrator runs NAV calculations daily. Some positions settle late. Some FX rates shift. The prediction model estimates: given the current data set, what is the expected NAV variance at close? If the variance exceeds a 2% threshold, the system alerts the team before the end of day, not after.
Globally, financial institutions will spend USD 132 billion on AI by 2030, according to Quinlan & Associates. Yet 71% of them struggle to capture clear ROI. The reason is almost always the same: they bought the prediction model before they fixed the data pipeline. A prediction model trained on incomplete or inconsistent data produces unreliable outputs. The ingestion layer matters more than the model.
Generation: drafting client notes and briefing summaries
Generation is the newest capability and the one most finance teams approach with caution. For good reason. You should not let an AI generate a trade instruction or a compliance filing without human review.
But generation has a safe, high-ROI entry point: internal briefing notes. An AI reads the day’s portfolio movements, market news relevant to each holding, and the previous client meeting notes, then drafts a one-page summary. The relationship manager reviews, edits, and sends it. The RM’s job shifts from writing to verifying.
Generation saves the RM roughly 45 minutes per client update. For a team managing 20 client relationships, that is 15 hours per week recovered for higher-value work, such as structuring deals, deepening relationships, or actually analysing the portfolio rather than describing it.
The common thread: data readiness
All three capabilities depend on one thing: clean, structured data that the AI can read. If your trade confirmations arrive as scanned PDFs with inconsistent layouts, if your custodian statements use different column names every month, if your compliance filings live in a shared drive with no naming convention, classification, prediction, and generation will all fail.
This is the actual bottleneck. Not the AI. The data.
Workflow engines like n8n solve the data readiness problem by connecting the systems that hold your data, such as bank portals, custodian APIs, and internal databases, and feeding the results into a consistent format that AI models can process. The AI is the thin layer on top. The workflow engine is the foundation.
AI is not a product you buy
The most common mistake we see in Hong Kong finance is treating AI as a procurement decision. A team gets budget approval, evaluates three vendors, picks one, and waits for it to transform operations.
It does not work that way. AI is a capability you integrate into existing workflows. The firms that get ROI from AI are the ones that started by documenting their current process, identifying the data bottleneck, and building the ingestion layer first. The model comes last.
Frequently Asked Questions
The SFC does not require registration of AI usage itself. It requires that senior management (MIC-IT) understand and can explain any AI-driven decision process. The key regulatory principle is accountability: a human must own the outcome.
No. A data scientist is valuable for building custom models from scratch, but most finance AI applications use pre-trained models adapted to your data. The scarce skill is not data science; it's someone who understands both the finance workflow and how to structure data for AI consumption.
A properly scoped first workflow typically takes 4-8 weeks from discovery to production. The first 2-3 weeks are spent understanding the data, not the AI. If the data is accessible and reasonably structured, the AI layer can be added in days.
