Documented deployments — including the course corrections — teach more than any benchmark. Three global cases, the adoption numbers behind them, and the patterns that translate to this region.
Klarna: scale, then the correction
The fintech's customer-service assistant became the reference case for agent ROI: within its first weeks it was handling roughly two-thirds of support chats, doing work the company equated to about 700-850 full-time agents, with resolution times falling from minutes to seconds — and Klarna projected roughly $40-60M in annual profit improvement. The sequel matters just as much: by 2025 the company publicly walked back its AI-only stance, rehiring humans for complex and emotionally charged cases after quality complaints, and settling on a hybrid model. The lesson is not that the agent failed — volumes stayed automated — but that escalation paths and quality evals are part of the product, not an afterthought.
JPMorgan: a portfolio, not a pilot
The bank reports hundreds of AI use cases in production — spanning fraud detection, payment validation, developer tooling and advisor copilots — under a heavily governed platform approach: centralised model access, risk review per use case, and observability as a requirement rather than an option. The pattern for any enterprise: treat agents as a managed portfolio with shared infrastructure (gateway, tracing, evals), so the fiftieth agent costs a fraction of the first.
Salesforce: the focused legal agent
A narrower, instructive case: an internal agent for contract review and routing that the company credits with roughly $5M in annual savings. Nothing exotic — one well-bounded document workflow, tight integration with the systems lawyers already use, and human approval retained on judgment calls. Focused beats flashy: the highest-ROI agents in most organisations look like this, not like a general assistant.
What the adoption data says
171% 74% 3.7x
average ROI reported by US enterprises on agent initiatives Landbase, 2025 of deployed enterprises saw ROI within year one Google Cloud / KPMG average return per $1 invested in genAI IDC / Microsoft
Two more findings shape strategy. Sector-specific agents materially out-earn horizontal assistants — Google Cloud's survey work puts top-quartile, domain-focused deployments at multiples of generic ones, and analysts track domain agents as the fastest-growing segment. And the failure data is equally loud: most 2025 proofs-of-concept never reached production, and Gartner expects over 40% of agentic projects to be cancelled by 2027 — with unclear KPIs and weak data quality, not model capability, as the leading causes. Everything in Chapters 11 and 12 exists to keep you out of that statistic.
Patterns that fit this region
Across the Gulf, three agent shapes recur because they match local industry structure — tourism, logistics and trade, retail and real estate — and a bilingual, WhatsApp-first customer base:
Tourism — guest concierge bilingual EN/AR on WhatsApp; itineraries, bookings, visa & weather FAQs; hands off to staff for complaints Logistics — dispatch copilot reads orders & telematics; proposes routes and load plans; flags customs-document gaps before the truck moves Retail & real estate — lead & order desk qualifies enquiries, checks stock or listings, books viewings or confirms orders; CRM-native
Figure 13.1 — Three regional patterns. Each is an L2-L3, single-agent system with MCP tool seams.
All three share the same skeleton from Chapter 12's worked spec: modest autonomy, CRM or TMS as the system of record, Arabic-English evals, and residency-aware deployment — which is exactly why a portfolio of such agents can share one platform.