The Real Cost of Enterprise AI in APAC: What the Vendor Quotes Leave Out

The model API cost is the line item everyone focuses on. It is also usually the smallest one.

When enterprise teams in APAC build the business case for an AI workflow, they typically anchor to the model cost — the cost of API calls to OpenAI, Anthropic, or Google, or the infrastructure cost of a self-hosted model. This number is visible, quotable, and easy to include in a spreadsheet.

It is also, in our experience, between 5% and 25% of total cost of ownership.

Here is where the rest of the spend actually goes.

Data preparation: typically 30–50% of project cost

The most consistently underestimated cost in enterprise AI projects is the work required to get data into a state where a model can use it. This includes:

• Labelling and annotation: For supervised learning workflows, human labelling is the rate-limiting step. Labelling 10,000 document classifications for a routing model takes subject-matter expert time — often 3–6 months of work from operational staff who also have day jobs.
• Cleaning and normalisation: Enterprise data collected over years in systems that were not designed for ML consumption is messy. OCR errors, encoding inconsistencies, missing fields, duplicate records, and legacy schema decisions all require remediation work before training.
• Schema mapping: Getting data out of legacy systems — CRMs, ERPs, custom-built case management tools — and into a format usable for training or retrieval requires integration work that is often scoped as "minor" and turns out to be the longest phase.

Across APAC enterprises with 5–20 year old core systems, data preparation typically represents 30–50% of total project cost and 40–60% of project timeline. It is also the phase where the most project delays occur.

Integration and change management: typically 20–30% of project cost

An AI model that runs in isolation from the operational workflow delivers no value. The integration cost — connecting model outputs to the systems where the work actually happens — is consistently underscoped.

This includes:

• API development and maintenance to push model outputs back into operational systems (CRMs, case management tools, ERP)
• Change management for the staff whose workflow changes when the AI workflow goes live
• Training on the new process for operational staff
• Escalation path design (what happens when the AI is wrong, and who reviews it)

Change management is particularly underfunded in APAC deployments. The assumption is often that staff will adapt with minimal support. Our experience is the opposite: in markets with strong process compliance cultures (Japan, Korea, Singapore public sector), detailed change management documentation and explicit workflow sign-off is required for adoption to stick.

Governance and monitoring: 10–20% ongoing annual cost

A model deployed in production requires ongoing maintenance that vendors rarely scope into the initial contract. This includes:

• Drift monitoring: Monitoring model performance against a baseline, detecting when accuracy degrades as the distribution of inputs shifts.
• Retraining pipelines: The mechanisms to collect newly labelled data, retrain the model, validate performance, and push to production on a cadence.
• Audit trail maintenance: For regulated industries, maintaining logs of model inputs, outputs, and human overrides — with retention periods specified by local regulation (13 months for MAS-regulated firms, longer for some APPI-governed data categories).
• Model card updates: Keeping the model documentation current as the model evolves.

This ongoing governance cost is typically 10–20% of the initial project cost per year. It is often zero in the business case because it is "not in scope for the initial engagement." It becomes in scope the first time there is a regulatory audit or a significant model drift event.

Opportunity cost of internal staff time

Every AI project requires sustained involvement from the client side: subject-matter experts for data labelling, operational staff for pilot testing, IT for integration support, compliance staff for governance review, and leadership for steering.

This time is real cost. It is often not counted because it is not an invoice line item — but the hours are diverted from other productive work. For a 4-month engagement with a 200-person team, we estimate 15–25 internal person-days per month on average across the project lifecycle. At senior staff rates in HK/SG, this is a material number.

A representative total cost of ownership

For a mid-market APAC enterprise deploying a document classification and routing workflow (roughly 3–6 month engagement scope):

The external engagement fee — the number everyone focuses on — is typically 30–55% of the first-year total cost of ownership, and 20–35% over three years when governance costs are included.

What this means for business cases

Business cases built on model API cost alone will systematically underestimate TCO. Business cases that do not include internal staff time will underestimate even further.

The practical implication: the ROI threshold for an AI workflow should be calculated against total first-year cost, not against vendor engagement fee alone. If the business case only works at a 3x ROI on the engagement fee, it may not work at 1.5x ROI on total first-year cost including data prep and internal staff time.

This is not an argument against AI investment — the ROI on well-scoped workflows is typically real. It is an argument for accurate business cases, which serve everyone better than optimistic ones that collapse on contact with the actual cost picture.

The ROI Calculator allows you to model costs and returns with configurable inputs for your organisation's specific context.