Dubai case study: a logistics startup cuts operating costs 25% with 3 AI agents
An e-commerce logistics startup in Dubai deployed 3 AI agents (Make + Claude + Notion + WhatsApp API) to automate coordination, support and billing. Result: 25% lower operating costs in 90 days.
Dubai has become in a few years one of the most active e-commerce hubs in the Gulf. Volumes are exploding but margins remain tight. For an 18-person logistics startup serving mainly DTC brands selling across MENA, every hour wasted on manual coordination, every billing error, every mishandled complaint hits profitability directly.
It is in this context that Karim, the startup's founder, wanted to explore what AI could really change in his operation. Here are the 90 days that followed.
The starting context
The company handled a weekly volume of 800 to 1,100 parcels for a dozen client brands. Three functions consumed the most time:
- Driver coordination: 12 in-house drivers, 8 external contractors, complex daily planning to organise from WhatsApp and Excel.
- Customer support: 60 to 90 requests per day, mainly on delivery status, returns and address errors.
- Client billing and reporting: a weekly cycle to invoice each brand client with detail on parcels delivered, returned and refused.
Before the project:
- 4 back-office people fully dedicated to these 3 functions.
- 3.2% billing disputes per month on average, fixed manually.
- Average support response time: 3h40 on weekdays, 8h on weekends.
Karim wanted a pragmatic solution, deployable in less than 90 days, that would not require layoffs but would make the team scalable.
The IMPACT methodology over 90 days
Phase 1: Diagnostic-Scoping (weeks 1 to 2)
4-day field audit at the Business Bay offices and the Al Quoz warehouse. Mapping of the 3 critical functions, analysis of historical WhatsApp conversations (1,200 sampled messages), direct observation of back-office work.
Output: three priority agents to build.
- Coordination Agent: handles daily delivery planning, driver assignment and incident management.
- Support Agent: replies to 70% of simple requests (delivery status, standard return, address change).
- Billing Agent: produces every Friday a weekly report per brand client with prepared invoicing.
Phase 2: Implementation (weeks 3 to 8)
Simultaneous build of the three agents on the same architecture:
- Make: central orchestrator, triggers agents based on events (new parcel, new client message, Friday 6pm).
- Claude API: reasoning engine for each agent. Each agent has its own system prompt and knowledge base.
- Notion: shared source of truth. Business procedures, driver records, client contracts, pricing.
- WhatsApp Business API: single communication channel with drivers and customers.
Coordination Agent
Every morning at 6:45am, the agent fetches the day's deliveries from the client system (Shopify, WooCommerce, or CSV import for clients without a platform). It groups them by zone, optimises routes integrating constraints (preferred slots, available drivers, each driver's capacity), then sends the day's route to each driver on WhatsApp.
Throughout the day, the agent tracks delivery confirmations and alerts the supervisor on anomalies (more than 30 minutes late on a delivery, customer refusal, access failure).
Support Agent
Every customer message coming into WhatsApp is analysed by the agent. It identifies intent (status, return, complaint), pulls the order from the CRM, and crafts a 2 to 4-line reply in Arabic or English depending on detected language.
For simple requests (90% of volume), the reply goes out directly with the supervised operator's signature. For complaints or atypical cases, the agent creates a ticket in Notion and notifies a human.
Billing Agent
Every Friday at 5pm, the agent compiles the week's deliveries for each brand client. It applies contractual pricing, flags anomalies (parcels billed twice, mismatch between address and zone pricing), generates a PDF report per client and a credit note if needed.
The report is sent to a supervisor for validation before transmission to the client. Validation takes 15 minutes versus 4 to 5 hours previously.
Phase 3: Pilot and autonomy (weeks 9 to 12)
4 weeks of close supervision. Agents run in parallel with the human team, who validates each sensitive action before publication. After 4 weeks, autonomy thresholds are progressively raised.
Full documentation handed to the operations lead, who becomes the contact for future tuning.
Results measured at month M+3
On coordination
- Time spent on daily planning: dropped from 2h30 to 25 minutes per day.
- Deliveries within the promised slot: rose from 78% to 91%.
- Undetected incidents (late deliveries without alert): dropped from 14 per week to 2.
On customer support
- Average response time: dropped from 3h40 to 18 minutes on weekdays, and from 8h to 35 minutes on weekends.
- Volume handled without human intervention: 71% of requests.
- Customer satisfaction (NPS): rose from 38 to 56 in 90 days.
On billing
- Weekly billing cycle: dropped from 8 hours of back-office work to 1 hour (generation and validation included).
- Billing disputes: dropped from 3.2% to 0.7%.
- Client payment delay: shortened by 7 days on average thanks to invoices sent earlier.
On costs
- Total monthly operating cost (back-office + tools + errors): cut by 25%.
- Monthly AI tool stack: 1,200 dirhams.
- Net monthly savings: 18,500 dirhams after support amortisation.
Across 12 months, ROI is positive from the third month. The 2 redeployed people became sales reps and contributed to signing 3 new client brands in 6 months.
What worked
The initial field audit. Concretely understanding how the team worked before proposing a solution avoided the main mistake of AI projects: automating a poorly designed process.
The independent agents approach. Each agent handles one function and no more. When one drifts, you intervene on it without breaking the other two. When you want to evolve business logic, you modify a single prompt.
Active supervision over 4 weeks. Adjustments made during this period absorbed 95% of atypical cases. Without this phase, autonomy would have been risky.
What could have gone smoother
The WhatsApp Business API integration took 5 days longer than planned because of a Meta verification process longer than announced. For similar projects in MENA, plan a 3-week buffer on this point.
Automatic dialect detection (Emirati Arabic vs Egyptian vs Moroccan) required 3 iterations. Fix: a dictionary of local expressions added to the system prompt and an immediate human fallback in case of doubt.
What this case study teaches
Three well-defined agents are worth more than one fuzzy super-agent. Modularity was the key factor of robustness in production.
Well-deployed AI frees human capital for growth, not the other way around. Karim was able to accelerate sales while operations became smoother.
In the IMPACT methodology, this project combines the Implementation phase and the Autonomy phase. The Autonomy phase is non-negotiable: without internal transfer, the agent becomes a dependency on the consultant, which is never the goal.
If you operate in MENA or run a high-volume operational business, the TransformAudit identifies the most profitable AI automations in 2 days with a 90-day deployment plan.
Let's take action
Want to accelerate your AI transformation?
Free 30-minute diagnostic to identify your priorities and AI roadmap.
Book my free diagnostic →