Key Takeaways
- Gartner's October 2026 CIO survey found AI spending up 18-fold in two years, but only 13% of CIOs report significant or transformational value.
- Split the 2027 AI budget into four buckets: Run, Scale, Prove and Foundations, each with an owner, one metric and a review date.
- Fund no pilot without a written stopping rule; Gartner found only 31% of CIOs have one.
- Budget usage-based AI as a range with hard caps, and track cost per completed task.
- Present the budget to the board as a portfolio of bets with exit points, not a technology shopping list.
Here is the number that should shape every 2027 AI budget conversation in Hong Kong. According to Gartner's annual CIO survey of more than 2,000 technology executives, AI spending has risen 18-fold over two years, yet only 13% of CIOs say their organisation has realised significant or transformational value from it.
That gap is now landing on the CFO's desk. Budget season is here, and the question has shifted from whether to invest in AI to which AI spend to keep, which to grow and which to stop. This guide gives you a portfolio framework that answers all three before your numbers go to the board.
Why is the 2027 AI budget harder to approve than last year's?
Finance teams now have two years of AI invoices to compare against results. Spending has grown faster than proven value, usage-based pricing makes costs harder to forecast, and agentic AI adds new risk. CFOs no longer fund experiments on trust; they want evidence, ownership and an exit plan for every line.
Deloitte's State of AI in the Enterprise survey, published on 29 October 2026 and covering 3,500 leaders in 25 countries, calls this moment an ROI reckoning. Most respondents (81%) report faster decision-making and execution, and the share reporting revenue gains from AI rose to 28% from 20% in January. Yet only 25% say they are using AI to deeply transform the business. For most organisations, AI is speeding up existing workflows rather than changing the work itself.
Three pressures that make this year different
- Value scrutiny. Pilots funded in 2025 and 2026 now need to show measurable outcomes or make way for something better.
- Cost volatility. Seat licences are giving way to credits, tokens and per-task pricing, so last year's spend is a weak predictor of next year's.
- Risk catch-up. Gartner analyst Kristin Moyer noted that only 21% of CIOs have agentic AI governance today, while 71% expect to deploy agentic AI within 12 months, as reported by InformationWeek.
What should a 2027 AI budget include?
A credible 2027 AI budget has four buckets: Run for AI already in daily use, Scale for pilots with proven value, Prove for new pilots with stopping rules, and Foundations for data, governance and skills. Each line needs a business owner, one outcome metric and a review date.
The four-bucket AI portfolio
| Bucket | What it funds | Approval test | Review rhythm |
|---|---|---|---|
| Run | Licences, AI credits and inference for tools already in daily use | Usage and adoption data from 2026 | Quarterly, by cost per task |
| Scale | Pilots that hit their target and are ready for more users or processes | Measured outcome against a baseline | Quarterly, against the business case |
| Prove | New pilots in priority processes | A written hypothesis, budget cap and stopping rule | At each stage gate, typically 8 to 12 weeks |
| Foundations | Data quality, integration, governance, security and training | Linked to at least one Scale or Prove line | Half-yearly |
The split between buckets will differ by organisation, and there is no universal benchmark worth copying. What matters is that the split is explicit. Many enterprises discover, the first time they sort their AI spend this way, that most of it sits in Run with no clear owner and almost nothing sits in Foundations.
One rule keeps the portfolio honest: money moves from Prove to Scale only when a pilot clears its target, and it leaves the portfolio when a pilot hits its stopping rule. That single discipline does more for AI return on investment than any vendor negotiation.
How do you set stopping rules before funding an AI pilot?
Write them before the pilot starts. A stopping rule names the metric, the threshold, the date and the person who decides. For example: if straight-through processing stays below 60% ten weeks after go-live, the pilot ends. Agreeing this in advance takes politics out of the decision and protects budget for better bets.
Gartner's October 2026 survey found that only 31% of CIOs have rules for stopping unsuccessful AI pilots, and just 17% have criteria for abandoning low-value AI already in production. The CIOs Gartner calls vanguards, those reporting the strongest outcomes, are more likely to set these rules.
A practical stopping rule has four parts:
- Metric: one business outcome, such as cost per claim processed or first-response time, not model accuracy alone.
- Threshold: the minimum result that justifies continued spend, agreed with finance before launch.
- Date: a fixed review point, usually 8 to 12 weeks after go-live.
- Decision owner: a named business leader, not the vendor and not the project team.
Apply the same logic to the Run bucket. If an AI tool licensed in 2026 still shows low active use after onboarding and training, it deserves the same review as a pilot.
How should you budget for usage-based AI costs?
Budget a range, not a single number. Model a base case from 2026 usage, a high case for adoption growth and a ceiling you will not exceed, then set hard spend caps and alerts per tool. Track cost per completed task, because total spend alone hides whether AI is becoming more or less efficient.
Deloitte's October survey shows how immature cost tracking still is: 71% of companies track total AI inference and API spend, but only 51% track token volume and 44% track per-unit inference cost. Without the unit view, a rising bill could mean healthy adoption or wasteful use, and finance cannot tell which.
For the 2027 plan, ask each budget owner for three numbers: expected monthly task volume, expected cost per task, and the cap that triggers a review. This turns a volatile line item into something a CFO can forecast and audit.
What does this framework look like in a Hong Kong enterprise?
Most Hong Kong enterprises find their AI spend scattered across IT, marketing and operations budgets with no single view. Sorting it into the four buckets usually reveals duplicated licences, one or two pilots worth scaling, and gaps in data and governance funding that hold everything else back.
Scenario 1: a regional logistics group
A logistics group with 400 staff ran three AI pilots in 2026: shipment-exception emails, customs document checks and a driver-facing chatbot. Only the document-check pilot hit its target. Under the portfolio framework, it moves to Scale with budget to extend to two more trade lanes, the chatbot is stopped, and the email pilot gets one more stage gate with a tighter threshold. The released budget goes to Foundations: cleaning the shipment master data that all three pilots struggled with.
Scenario 2: a professional services firm
A 150-person accounting firm finds that most of its AI spend is individual chatbot subscriptions bought on corporate cards. Moving these into Run under one owner, consolidating to an enterprise plan with proper data controls and setting a usage review becomes its first budget decision. The firm then funds a single Prove line, drafting first-pass client queries in the tax team, with a stopping rule tied to partner review time.
In both cases, the Personal Data (Privacy) Ordinance shapes what can be funded. The Privacy Commissioner's Model Personal Data Protection Framework for AI expects risk assessment and human oversight proportionate to the risk. That work belongs in the Foundations bucket, not in a later clean-up.
What mistakes derail enterprise AI budget planning?
The most common mistakes are budgeting by vendor rather than by outcome, rolling every 2026 pilot forward by default, treating governance as a cost to minimise, and leaving adoption unfunded. Each one makes the budget easier to write but much harder to defend when results are reviewed mid-year.
- Budgeting by vendor. A line called chatbot or copilot says nothing about the outcome. Name lines by process and result.
- Automatic rollover. A pilot that has not met its target should not inherit next year's budget simply because it exists.
- Underfunding foundations. Data quality and integration problems sink more pilots than model quality does. If no Prove line can succeed without clean data, fund the data first.
- Ignoring adoption. Licences without training and workflow redesign produce low usage and weak results. Budget for the people side explicitly.
- Agents before governance. The gap between 21% with agentic governance and 71% planning agentic deployment is a warning. Approve agent pilots only together with access controls, logging and a named owner.
How do you present the 2027 AI budget to your CFO and board?
Present it as a portfolio of bets with exit points. Show last year's results by bucket, what you are stopping, what you are scaling and why, the cost range for usage-based tools, and the controls that come with any new autonomy. Boards approve budgets they can monitor, not budgets that promise transformation.
A one-page board summary can follow this order:
- What 2026 delivered: outcomes by bucket, including pilots stopped and the budget they released.
- What 2027 funds: each Scale and Prove line with its owner, metric and review date.
- Cost envelope: base, high and ceiling cases for usage-based spend, with the caps in place.
- Risk and controls: the governance, privacy and security measures funded in Foundations.
- Decision points: when the board next sees results and what would trigger reallocation.
Gartner's advice to CIOs at its October 2026 symposium was to demonstrate measurable business value before increasing investment in any AI initiative. A budget built this way does exactly that. It also gives you a credible story when results come in mixed, because stopping a weak pilot reads as discipline rather than failure.
Fund fewer AI bets, and manage them harder
The organisations that pull ahead in 2027 will not necessarily spend the most on AI. They will be the ones that know what every AI dollar is for, who owns it and when it stops. The four-bucket portfolio turns AI from a list of hopeful experiments into an investment programme your CFO can track.
Start this month by sorting your current AI spend into Run, Scale, Prove and Foundations. The gaps you find will show you where the 2027 budget should go.
Empowering AI. Centered on you.
Turn your 2027 AI budget into one measurable first role
If your portfolio review points to a process ready to scale or prove, start with one AI role and one metric. Book a free AI employee configuration assessment and we'll walk you through every step, from choosing the process to setting the stopping rule and reporting results upward.
Frequently Asked Questions
How much should a mid-sized Hong Kong company spend on AI in 2027?
There is no reliable universal benchmark. Size the budget from value cases you can measure: the processes you plan to scale or prove, their expected outcomes and the foundations they need, then add a capped range for usage-based tools.
Should the AI budget sit with IT or with business units?
Usually both. IT typically owns Foundations and shared platforms, while business units own Scale and Prove lines because they own the outcome. One portfolio view across both prevents duplicated licences and orphaned pilots.
What if our 2027 budget is already locked?
Apply the framework inside the approved total. Sort existing lines into the four buckets, set stopping rules on current pilots and agree reallocation points with finance, so savings from stopped work can fund pilots that clear their targets.
Should AI insurance be part of the budget?
Treat it as a supplement, not a control. Gartner reported in October 2026 that only 1% of CIOs have bought AI insurance and 34% are considering it; fund governance and monitoring first.
How often should the AI portfolio be reviewed?
Quarterly for Run and Scale lines, at each stage gate for pilots, and half-yearly for Foundations. A quarterly portfolio summary to the executive committee keeps reallocation decisions timely.