Most companies do not fail to see AI’s potential. They fail because they do the math wrong. The most common mistake is treating usage as value: people logged into ChatGPT, teams played with copilots, a few flashy demos impressed management—therefore AI is “working.” It usually isn’t. Being able to access a tool and being able to complete a business process end-to-end, repeatedly, with measurable economics are two very different things.
My view is straightforward: AI ROI should not begin with model intelligence. It should begin with whether the workflow was redesigned, whether unit economics improved, and whether risk costs were included. For SMEs and mid-sized firms across Hong Kong, Taiwan, Singapore, and mainland China, the winning question is rarely “Which model is best?” It is “Which use case can convert labor, speed, quality, and revenue impact into auditable numbers within 6 to 12 months?”
In actual deployments, we repeatedly see the same pattern: firms overestimate license costs, underestimate integration costs; overvalue demos, undervalue change management; overstate time savings, and ignore compliance, governance, and review overhead. That is why I use a simpler executive framework: the three-ledger model of AI ROI—the efficiency ledger, the decision ledger, and the risk ledger. If you only measure the first, you will often make the wrong investment decision.
Are you measuring tool adoption—or business outcomes?
Many AI programs report adoption rate, active users, prompt volume, or number of copilots deployed. Those are easy metrics to collect. They are not the metrics owners should lead with. What matters is output: Did average case handling time fall? Can each support agent manage more concurrent conversations? Did proposal turnaround time shrink? Did conversion improve?
Gartner repeatedly warned in 2024 that many generative AI initiatives risk remaining stuck at the proof-of-concept stage if they are not tied to clear business goals and operating changes. McKinsey’s 2023 report, The Economic Potential of Generative AI, argued that the largest gains are in functions such as customer operations, marketing and sales, and software engineering—but only when AI is embedded into the workflow, not isolated as a chat window.
That leads to a hard rule: if a metric cannot be connected to the P&L or a meaningful operating KPI, it should not be your primary ROI metric. A legal team may say summarization is 60% faster. Fine. But if the lawyer still rewrites everything from scratch, actual savings may be only 10%. A customer service bot may deflect 35% of inbound queries; if CSAT falls and escalations rise, ROI may still be negative.
The three ledgers: efficiency, decision, and risk
I recommend splitting AI ROI into three ledgers, each with a different logic.
1. The efficiency ledger. This is the easiest to quantify: labor hours saved, outsourcing costs reduced, throughput increased, response times shortened.
A simple formula:
Efficiency ROI = (value of labor saved + reduction in outsourcing/error costs + value of additional capacity) / total AI cost
Take a Hong Kong insurance intermediary processing 8,000 documents a month. Initial classification and summarization took 6 minutes per file; with AI, it drops to 2 minutes. That saves 32,000 minutes, or roughly 533 hours monthly. At a fully loaded labor cost of HK$180 per hour, that is about HK$95,940 per month, or over HK$1.15 million annually. These are the projects that tend to win approval quickly because the cash-flow logic is clear.
2. The decision ledger. This is where many companies leave money on the table. AI does not only save time; it can improve conversion, pricing, cross-sell, inventory turns, and forecast quality. If a sales team uses AI to segment prospects and draft proposals, and quote cycle time drops from 5 days to 2 while win rate rises from 18% to 21%, the economic value may exceed labor savings. IDC’s AI business value research has consistently pointed to revenue impact and process acceleration as more strategically important than pure cost takeout.
3. The risk ledger. This is the ledger boards should care about most, and many teams ignore it almost entirely. It includes data leakage, bad recommendations, copyright exposure, regulatory risk, model drift, and the cost of human review. In regulated sectors—financial services, healthcare, legal, cross-border trade—assuming risk cost is zero almost guarantees inflated ROI. Deloitte’s 2024 enterprise GenAI surveys made the point clearly: governance and risk controls will determine scaling speed as much as model capability does.
The real cost is not the subscription—it is lifecycle ownership
Owners often ask: “If it’s only US$20 or US$30 per user per month, how expensive can it be?” That is the wrong question. License price is not total cost of ownership. Many AI ROI calculations fail because they calculate subscription cost and ignore everything else.
At minimum, your TCO should include six buckets: 1) model or SaaS subscription, 2) API usage, 3) systems integration with CRM, ERP, document repositories, and knowledge bases, 4) data cleaning and access governance, 5) employee training and workflow redesign, and 6) human review plus risk control.
The table below is not a buyer’s guide. It is a reminder that price is only the visible layer; integration model and fit-for-purpose positioning usually determine ROI.
| Product / approach | Reference pricing | Primary positioning | Strengths | Risks / constraints |
|---|---|---|---|---|
| OpenAI ChatGPT Team | US$25–30/user/month | General knowledge-work copilot | Fast adoption, strong drafting/summarization | Requires governance; limited process integration out of the box |
| Microsoft Copilot for Microsoft 365 | ~US$30/user/month | Deep office productivity | Strong integration with Outlook, Word, Excel, Teams | ROI is overstated if core work happens outside M365 |
| Google Workspace with Gemini for Business | ~US$20–30/user/month | Google-native collaboration | Natural fit for Gmail, Docs, Meet workflows | Value depends on existing Workspace penetration |
| Salesforce Einstein / embedded AI | Module- and edition-based pricing | CRM-native sales and service AI | Deep linkage to customer data and workflows | Higher cost; requires mature CRM discipline |
| Self-built RAG + open-source model stack | Varies by cloud, talent, and support | Proprietary knowledge, higher-control use cases | More control; can keep data on private cloud/on-prem | Higher build and maintenance cost; technical complexity |
In much of Asia-Pacific, a 50- to 300-person business with a limited IT team should often start with general copilots. But in finance, healthcare, legal services, manufacturing supply chains, or cross-border operations, the bigger value usually comes from embedded AI in high-frequency processes: document review, customer query handling, knowledge retrieval, quote configuration, and claims support. The first approach looks cheaper. The second is more likely to create durable ROI.
No baseline, no ROI: the six metrics every owner should track
AI ROI does not start after launch. It starts before launch, with a baseline. Without a baseline, every “improvement” is mostly storytelling.
At minimum, track six metrics:
- Unit task cost: the full cost per report, policy, service interaction, invoice, or claim.
- Cycle time: how long it takes from request to completion.
- First-pass accuracy: the share of outputs usable the first time, not just total outputs generated.
- Human review rate: how much still requires substantial rework.
- Revenue impact: conversion rate, average order value, renewal rate, churn, upsell.
- Risk event rate: hallucinations, compliance breaches, leakage, complaints.
Forrester’s Total Economic Impact work on automation and AI repeatedly shows that the strongest programs are not defined by the most advanced model. They are defined by a clear baseline, a comparison method, and post-deployment measurement discipline. That matters even more for SMEs, because they do not have endless budget for experimentation.
A practical formula is:
AI ROI = [(cost savings + incremental gross profit + value of risk avoided) - total AI cost] / total AI cost
Be conservative. Measure over 3, 6, and 12 months. Apply an adoption discount—say only 60% of intended users actually use the tool consistently. Apply a quality discount—say only 70% of outputs are directly usable. The resulting number will usually be less exciting than the board slide, but much closer to reality.
Not every AI project deserves funding
This does not mean innovation projects are worthless. It means first-wave investment should go to workflows that are frequent, repeatable, rule-bounded, and reviewable. In Hong Kong and Taiwan especially, the fastest-payback AI deployments are rarely the most glamorous multi-agent systems. They are customer support knowledge response, document classification and summarization, sales proposal drafting, internal knowledge search, meeting recap, and invoice or purchase-order processing.
By contrast, if your first move is fully automated decision-making across departments in a heavily regulated data environment, the project often gets trapped in legal review, security concerns, permission design, and data remediation. Six months later it is still in pilot. Stanford’s AI Index Report 2024 makes the broader point: enterprise AI adoption is rising, but scaled value still depends heavily on organizational capability, data readiness, and governance maturity—not just benchmark scores.
So yes, starting small is not timid. Starting where value is measurable is capital discipline.
Don’t convert all time saved into cash savings
Here is the cold-water point many ROI decks avoid. A project may claim it saves thousands of hours per year. That does not mean the company immediately saves the equivalent salary cost in cash. If the saved time simply lets employees do other work, without reducing overtime, outsourcing, headcount growth, or improving revenue, then what you achieved is capacity release, not cash recovery.
That is not useless. For a growing Singapore startup, a Hong Kong professional services firm, or a Taiwan exporter, capacity release may mean delaying three hires, responding faster to customers, or taking on more business. That is real value. But be honest about the distinction: hard savings hit the P&L directly; soft savings improve flexibility and throughput. Boards can accept soft savings. They should not let teams pass them off as cash.
What to do next: three actions before you scale AI
If you run a business, keep it practical.
First, pick a workflow before you pick a model. Find a task that happens hundreds or thousands of times a month, has some standardization, and is already constrained by labor.
Second, build the case using the three ledgers. Efficiency ledger: what costs fall? Decision ledger: what revenue or speed improves? Risk ledger: what review, compliance, and error costs remain?
Third, set a 90-day proof threshold. Not “people liked it.” Instead: at least two or three measurable improvements, such as 30% lower cycle time, 15% higher first-pass accuracy, 20% lower review rate, or reduced outsourcing cost.
The most mature AI adopters are not the ones that buy the most tools. They are the ones that establish an ROI discipline that is auditable, repeatable, and scalable. AI is not magic. It amplifies what is already true in the business: good processes get stronger, bad processes get exposed.
Self-check questions
- Which of your current AI metrics reflect excitement and usage—and which actually tie back to the P&L?
- Does your ROI model include integration, governance, review overhead, and risk cost, or only subscription fees?
- If after 90 days people say the tool is “helpful” but cycle time, cost, and revenue have not moved, will you stop the project?
