Is this AI system creating value proportional to its cost, or consuming resources without return?
A project starts with a clear business case: automate invoice processing, reduce churn through predictive modelling, improve lead scoring. Budget is allocated. Initial results are promising. Then it stalls. Accuracy degrades but no one is responsible for retraining. API costs accumulate but no one tracks them against revenue impact. The engineering team moves on but the infrastructure keeps running.
This is the Zombie Project. It is not failed in any formal sense — no one has declared it a failure or conducted a post-mortem. It simply persists, drawing resources like a slow leak no one has been tasked with finding. In a well-funded enterprise, Zombie Projects are an inefficiency. In an SME, they are an existential drain — shortening runway and diverting capital from the initiative that might actually work.
Generative AI has introduced a cost structure many SMEs do not yet understand: token-based pricing. An SME running a GPT-4 class model for customer support can face monthly API bills that exceed the salary of the human agent the AI was supposed to replace. A successful deployment — one that customers actually use — costs more than an unsuccessful one. Without monitoring, popularity becomes the thing that makes it financially unsustainable.
DIFC Regulation 10, Section 10.5 requires organisations to demonstrate "necessity and proportionality" for each AI processing activity. If a system costs more than the value it creates, it fails the proportionality test. This is not abstract legal argument — it is a practical standard a regulator can apply to any system in the register.
Board directors in most jurisdictions have a fiduciary duty to prevent waste of company assets. An AI project consuming significant budget without measurable returns is, at minimum, a governance question for the board. At maximum, it is a breach of fiduciary duty that exposes directors personally.
Beyond operational costs, model collapse — where performance degrades through data drift, feedback loops, or upstream model changes — can mean writing off months of engineering investment entirely. Monitoring for model collapse is not optional. It is a prerequisite for protecting the investment the organisation has already made.
A startup burning USD 5,000 per month on an AI tool that produces no measurable return is not wasting money. It is shortening its runway.
Requires a register of processing activities demonstrating "necessity and proportionality." If an AI system costs more than the value it creates, it fails the proportionality test — making it not merely wasteful but non-compliant.
Directors have a legal obligation to prevent waste of company assets. AI projects consuming budget without measurable returns are a governance question that can expose directors to personal liability.
Mandates ongoing monitoring of high-risk AI systems post-deployment. Combined with proportionality principles, this creates an implicit requirement to track whether systems continue to deliver value relative to their cost and risk.
No AI system moves to deployment without a documented ROI hypothesis: projected usage, cost per interaction, break-even analysis, and the threshold at which deployment becomes uneconomical.
Continuous monitoring of actual AI costs — API spend, compute, storage, engineering hours — tracked against the approved budget and original business case.
Periodic review mapping each AI system's total cost against demonstrated value, producing evidence that meets DIFC Regulation 10 Section 10.5 proportionality requirements.
Defined accuracy thresholds that trigger review when breached, with escalation paths for systems showing signs of drift, degradation, or collapse.
When cost exceeds budget or performance drops below threshold, the system forces a decision to someone with authority: retrain, restructure, or retire.
Book a free consultation to benchmark your AI governance posture against 140+ global regulations.
Book a Call