Why Gulf AI projects overrun their budgets: six hidden costs your spreadsheet misses

AI strategy & adoption

Roland Berger's April 2025 analysis of Middle East AI projects found costs that frequently exceed initial projections by 500–1,000%, with more than half of AI initiatives in the region terminated because of budget overruns. This is not a pricing article. It is a diagnostic of why your estimate is wrong before anyone writes a line of code. Six cost categories specific to the Gulf are almost never in the spreadsheet: Arabic NLP compute overhead, data fragmentation remediation, vendor switching costs, the expat talent premium, data sovereignty compliance, and the gap between global delivery benchmarks and local integration reality. Here is what each one costs, how to size it, and how to rebuild the budget line so your finance director and your CTO are looking at the same number.

Why the spreadsheet is already wrong

The pattern is consistent. A vendor or an internal team produces a budget for an AI project, the board approves it, and within two quarters the number has moved so far that the project is either re-scoped into something unrecognisable or quietly cancelled. Roland Berger's analysis of AI project costs in the Middle East, published in April 2025, puts figures to it: costs frequently exceed initial projections by 500–1,000%, more than half of AI initiatives in the region are terminated because of budget overruns, and fewer than 15% of organisations properly identify, quantify and measure the costs, risks and value of their AI projects.

The money is not the constraint. The Middle East AI market is projected to reach USD 320 billion by 2030. Cost discipline is.

Estimates fail for a structural reason. A typical spreadsheet covers the visible components Roland Berger lists: data acquisition and management, model development, application build and integration, and ongoing operational spend such as licences and inference. Those are real costs, and they are roughly the costs a global SaaS vendor prices for, which is why every competing bid includes them. The costs that sink projects in the Gulf sit underneath them.

Regional survey data shows the same gap from a different angle. The Deloitte and Mohamed bin Zayed University of Artificial Intelligence study of 155 companies across Saudi Arabia, the UAE and Qatar found that 26% of firms cite managing risks as a top barrier to adoption, that more than 80% feel pressured to adopt AI, and that almost half say they lack the talent and technology capabilities to scale it. A budget overrun is the most visible form of that risk. When the estimate is wrong, governance, ROI and vendor choice are wrong with it.

The six categories below are the ones we see missing when clients in the UAE and the wider GCC ask us to review an AI business case. Each one can be sized, and each belongs in the budget as its own line.

Hidden cost 1: Arabic NLP compute overhead

Roland Berger flags implementation complexity first: organisations deploying Arabic natural language processing "often encounter substantially higher computational requirements due to dialectical variations and linguistic complexities", and real-time analysis needs more compute than projected.

The mechanism is straightforward. Most frontier models were trained predominantly on Western datasets and tokenised for English. Arabic text produces more tokens per unit of meaning on common tokenisers, and a workspace assistant that must handle Emirati, Egyptian, Levantine and Gulf dialect input in one conversation cannot rely on a single register. Dialect variation also multiplies evaluation work: every prompt change needs re-testing across dialects, not once.

The sizing rule we apply is simple. Budget Arabic-first inference at 2–3 times the token volume of an equivalent English-only workload, then add retry overhead on top. Here is the arithmetic in code, using a bilingual workspace assistant as the example:

# Bilingual workspace assistant: monthly inference budget (USD)
seats, sessions_per_seat, tokens_per_session = 500, 40, 12_000
price_per_million = 8.00          # blended input/output price
retry_overhead = 0.08             # 5-10% of calls retry at least once
arabic_share, dialect_multiplier = 0.60, 2.5

base = seats * sessions_per_seat * tokens_per_session
english = base * (1 - arabic_share)
arabic = base * arabic_share * dialect_multiplier

for label, monthly in [("English-only estimate", base), ("With Arabic dialect load", english + arabic)]:
    usd = monthly * (1 + retry_overhead) / 1e6 * price_per_million
    print(f"{label}: USD {usd:,.0f}/month, USD {usd * 12:,.0f}/year")

Output:

English-only estimate: USD 2,074/month, USD 24,883/year
With Arabic dialect load: USD 3,940/month, USD 47,278/year

That is a 90% increase on the single line most people think of as fixed, before a single word of dialect evaluation is scheduled. If your budget shows one inference figure and most of your users are Arabic-speaking, that figure is fiction.

Hidden cost 2: Data fragmentation remediation

The MIT Sloan Management Review Middle East and Pure Storage readiness study (July 2025) gives the number that belongs on every board slide: only 33% of organisations report having more than 30% of their data accessible for AI use. Integration complexity (67%) and outdated infrastructure (56%) are the most commonly cited obstacles to scaling AI.

Translate that into project work. If your customer records live in a 12-year-old core banking or ERP system, your contracts sit in shared drives with no owner, and half your scanned documents are Arabic PDFs with no machine-readable text, then "data preparation" is not a two-week task at the start of the project. It is months of de-duplication, lineage mapping, ownership negotiation across departments, OCR and annotation, plus the governance to keep it clean afterwards. Roland Berger makes the same point: preprocessing and standardisation requirements "frequently exceed the initial resource allocations".

Planning guidance: where the readiness survey says a third of data is accessible, assume remediation work proportional to the gap, and give it its own budget line and its own owner. Never let it hide inside "discovery". A useful sanity check: if your estimate does not name the systems of record, the document formats and the person who owns each data domain, the data line is a guess.

Hidden cost 3: Vendor lock-in and the switching bill

Roland Berger's hidden cost list includes vendor dependency: organisations reliant on specific AI vendors "encounter substantial costs when trying to switch to alternative solutions", and these are "frequently underestimated during initial project planning".

This bites harder in the Gulf than elsewhere. Vendor selection is often driven by the least risky option and by speed to a pilot, and by the time the model is embedded, three things have happened: your prompts are tuned to one provider's behaviour, your evaluation suite encodes that provider's failure modes, and your data sits in a region or a format another provider cannot take without re-architecture. When the provider later changes pricing, drops a model version, or cannot satisfy a data residency requirement, the switching cost arrives as an unbudgeted emergency.

The fix is architectural, and cheap relative to the risk: put an OpenAI-compatible gateway in front of every provider so routing, budgets and access control live in one place, and keep prompts, evaluations and embeddings in infrastructure you control. A gateway policy can be as small as this:

# Illustrative gateway policy: one interface, providers swappable
providers:
  primary:  { base_url: "https://llm.internal.example/v1", model: "self-hosted-32b" }
  fallback: { base_url: "https://provider-b.example/v1",   model: "hosted-small" }
routing:
  strategy: cheapest-first
  budgets:
    workspace-assistant:
      monthly_usd: 4000
      on_exceed: degrade-to-cheap

Budget for the switch from day one. A realistic one-off line for re-embedding a corpus, rewriting prompts and re-running evaluations against a new provider is three to six weeks of engineering time. Note that Thread Transfer's LLM cost breakdown puts the crossover point where custom routing infrastructure pays for itself at roughly USD 100,000 per month of API spend; below that, a lightweight gateway is still worth it for the switching option alone. Our own FastLLM Proxy is an OpenAI-compatible gateway in front of your own LLM servers and 80 hosted providers, with routing, budgets and access control in one place, which is exactly this abstraction.

Hidden cost 4: The expat talent premium

Financial management is a major factor in executive dissatisfaction with AI projects in the region, second only to insufficient access to qualified personnel. Roland Berger's human capital finding is blunt: organisations in the UAE predominantly rely on expatriate talent, "resulting in elevated compensation requirements and associated onboarding expenditure".

The Deloitte and MBZUAI study puts numbers on the demand side: more than 80% of regional organisations feel pressured to adopt AI, yet almost half say they lack the talent and technology capabilities to scale successfully. On the supply side, Thread Transfer's model for LLM operations puts running a production AI system at 0.5–2 fully loaded engineers, and prices a fully loaded engineer at USD 150,000 per year, or USD 75,000–300,000 annually for that function alone.

Three costs hide inside that range. First, the salary premium and the relocation, visa and housing packages attached to it. Second, onboarding: a senior hire needs three months before they understand your data estate, which delays everything else. Third, retention and upskilling, which Roland Berger lists separately because training programmes "frequently require expansion beyond the initial projections". A project that budgets 0.5 FTE of AI engineering and runs at 1.5 FTE has created a recurring six-figure dirham gap that no amount of inference optimisation will close.

Hidden cost 5: Data sovereignty and compliance overhead

UAE compliance is not a one-off legal review at the end of the project. Federal Decree-Law No. 45 of 2021 concerning the Protection of Personal Data (in force since 2 January 2022) creates standing operational obligations that cost money every quarter:

• A record of processing activities, covering categories of data, authorised persons, processing times and cross-border movements, which must be produced to the UAE Data Office on request (Article 7).
• A Data Protection Officer where processing uses new technologies at high risk, involves systematic profiling, or handles large volumes of sensitive personal data (Article 10).
• A data protection impact assessment before processing with modern technologies that pose a high risk to privacy, including systematic automated evaluation of individuals (Article 21).
• Breach notification to the Data Office and to affected data subjects (Article 9).
• Cross-border transfer conditions under Articles 22 and 23, which is precisely the review that turns "we will use the global API" into a project task.

Note also the carve-outs in Article 2: personal health data and personal banking and credit data have their own legislation, and free zones with their own data protection regimes sit outside the decree. If your AI touches those domains, the compliance perimeter is wider than the PDPL, and so is the budget.

The regional pattern confirms the cost. Deloitte found that 53% of surveyed organisations use outside vendors to audit and test their AI tools, and misuse of customer data is among the top concerns cited. Budget a first-year compliance programme (DPO, records, impact assessment, transfer review, breach runbook, external audit) as its own line, plus an annual run rate. The finding that 26% of firms cite managing risks as a barrier is not about fear. It is about the cost of doing this properly.

Hidden cost 6: Global benchmark timelines vs local integration reality

Global vendors quote delivery timelines from their standard implementation playbooks. Roland Berger's fourth recommendation is to take into account "extended implementation timelines compared to global benchmarks", and the readiness survey explains why: 67% cite integration complexity and 56% outdated infrastructure as obstacles, and only 46% of regional organisations have moved beyond pilots into broader implementation.

The damage from timeline slip is not the calendar. It is burn rate. Every extra month keeps the data engineers, the ML engineers, the security reviewer and the project manager on the payroll while nothing ships. A four-month pilot that becomes seven months does not cost 75% more. It costs the full team burn for those three months, on top of everything else, with no revenue or saving attached.

Budget for integration explicitly: name each legacy system, the interface it exposes, the owner who must approve access, and a realistic months figure per system. If the plan assumes a greenfield integration because the vendor's reference architecture says so, that assumption is worth zero in a Gulf enterprise running decade-old core systems.

Worked example: a bilingual workspace assistant for a Dubai venture studio

Consider a realistic scenario, not a client case. A Dubai-based venture studio is building an AI-powered startup workspace: a bilingual assistant over 40,000 internal documents, 500 seats, Arabic and English, retrieval-augmented, hosted in part on cloud infrastructure. The team wants the board to approve a first-year budget. Here is how the evaluation runs.

Step 1: the visible budget.

Step 2: apply the six hidden categories.

Step 3: the verdict. Realistic year-one cost is about AED 2.46 million against an AED 892,000 spreadsheet: roughly 2.8 times the original number. For a task-specific project like this one, that is a normal outcome. Roland Berger's 500–1,000% band applies to transformative strategic initiatives that hit the full spectrum of hidden costs at once. This workspace is worth funding, but only as a stage-gated range of AED 2.2–2.7 million, with the Arabic inference sizing, the data remediation scope and the compliance programme approved as separate lines before build starts.

Step 4: choose the estimating method.

What to do next

If you are building the AI budget line for the next financial year, do these five things before the next steering committee:

1. Split the budget into visible and hidden lines. If a cost category from this article is missing from the spreadsheet, it has not disappeared. It is simply unmanaged.
2. Size inference for the language your users speak. Run the token arithmetic on real session data, with the dialect multiplier and retry overhead included.
3. Commission the data readiness check first. Know what share of your data is genuinely AI-accessible, which systems of record are involved, and who owns each domain. The 33% figure from the regional survey is an average, not a floor.
4. Put switching costs in the plan. Price the one-off cost of moving providers, and buy the abstraction (a gateway, portable prompts, owned evaluations) while it is still cheap.
5. Approve ranges with stage gates, not point estimates. Present AED 2.2–2.7 million with gates, not AED 892,000 with confidence. Roland Berger's finding that fewer than 15% of organisations quantify costs properly is an advantage available to anyone willing to do the arithmetic.

At Azrty we do this work from Dubai: the readiness assessment, the cost model and the architecture that keeps switching costs low. If you want a second pair of eyes on an AI business case before it goes to your board, our AI strategy and readiness engagement is the right place to start. The projects that survive are not the ones with the best models. They are the ones whose budgets told the truth on day one.

https://www.azrty.com/blog/why-gulf-ai-projects-overrun-their-budgets-six-hidden-costs-your-spreadsheet-mis

<a href="https://www.azrty.com/blog/why-gulf-ai-projects-overrun-their-budgets-six-hidden-costs-your-spreadsheet-mis">Why Gulf AI projects overrun their budgets: six hidden costs your spreadsheet misses</a> (Azrty)