The case studies most vendor pitches skip

If you read AI-vendor case studies in 2026, you'd think every deployment is a triumph. In reality, some are excellent, some are ordinary, and some are quietly getting killed at renewal.

This piece is different. We're pulling twelve real vertical agent deployments we've either worked on directly, or that are publicly documented in enough detail to say something honest about. For each: what the agent does, the deployment pattern, what makes it work, what would have made it fail. The goal is applicable pattern-recognition, not vendor marketing.

Names are used where they're already public. Details are lightly anonymized where we worked directly. Numbers are approximations rounded to preserve honesty without violating confidentiality.

1. Lyft's customer-support agent

What it does. Handles rider and driver support tickets end-to-end for a defined set of common cases (refunds within policy, ETA disputes, missing-item lookups). Escalates complex cases to human agents with a structured summary.

Deployment pattern. LLM-orchestrated with strict tool scopes. Refunds are auto-approved up to a limit; anything above requires human review. Every decision logged.

What makes it work. Aggressive scoping. The agent doesn't try to handle everything — it handles the top 40% of ticket types very well and escalates the rest. Deflection rate is real, not inflated.

What would have made it fail. Trying to do the hard cases (safety, disputes, driver deactivations) autonomously. They keep those human. Discipline > ambition.

2. Stripe's Kai (internal engineering agent)

What it does. Assists Stripe engineers with codebase questions, API documentation, and cross-team knowledge. Reportedly built on Anthropic's Deep Agents primitives.

Deployment pattern. Internal deployment first. Not customer-facing. Multiple months of internal use before any external productization.

What makes it work. The corpus (Stripe's internal knowledge) is highly structured and well-maintained. The user base (Stripe engineers) is technically sophisticated. Feedback loops are tight.

What would have made it fail. Shipping externally too early. Internal agents can be rougher; external ones face user expectations that internal deployments don't.

3. Monday.com's Sidekick

What it does. In-product AI assistant that helps users configure workflows, understand board structures, and automate repetitive project management tasks.

Deployment pattern. Embedded in the existing product, not a separate app. Contextual — knows what the user is looking at and what team they're on.

What makes it work. Deep product integration. Sidekick can actually do things (create tasks, modify workflows) not just answer questions. Users get value in the flow, not by switching contexts.

What would have made it fail. Answers-only chatbot. The competitive edge is action, not information.

4. Vodafone's customer-service voice agent

What it does. Handles inbound customer service calls in multiple European languages. Common queries (bill inquiries, plan changes, service issues) resolved fully by the agent; complex cases handed off.

Deployment pattern. Voice-first (see voice agents comparison). Regional data residency. Sophisticated escalation logic based on user frustration signals.

What makes it work. Language quality per region. A Portuguese-speaking customer gets a genuine Portuguese-speaking agent, not a translation artifact. Escalation to human is fast when needed.

What would have made it fail. Under-investment in language quality. Multilingual voice done poorly is worse than monolingual done well.

5. A Bengaluru fintech's AI collections agent (our client)

What it does. Contacts customers with overdue accounts via WhatsApp and calls, negotiates payment plans within policy, and completes payment collection flows. Handles conversations in Hindi, English, Kannada, Tamil, and Telugu.

Deployment pattern. Multi-channel (WhatsApp + voice). Multi-language (5 Indian languages). Deep compliance layer for RBI regulations. Human escalation for anything outside policy.

What makes it work. Cultural fit. The agent's tone in each language matches how collections calls actually happen in that region. Compliance is built in, not bolted on.

What would have made it fail. English-only. Or a "translated" agent that literally translates English scripts. Collections is deeply cultural.

6. LATAM Airlines' rebooking agent

What it does. Handles rebooking during flight disruptions. When a flight is delayed or cancelled, the agent reaches affected passengers, proposes options within airline policy, and completes rebooking.

Deployment pattern. Event-driven. When flight status changes, the agent activates. Proactive contact via WhatsApp and email.

What makes it work. Proactive beats reactive. Passengers who get a rebooking option pushed to them within minutes have a completely different experience than passengers stuck in a call queue.

What would have made it fail. Waiting for passengers to call in. The agent's value is timing.

7. A UAE public-sector document processing agent (our client)

What it does. Handles resident applications for various government services — reads uploaded documents, extracts data, validates against policy rules, and routes for human approval or issues.

Deployment pattern. On-prem deployment (data residency requirement). Arabic-first with English support. Deep integration with government systems.

What makes it work. Structured task, high volume, clear rules. Ideal use case for AI extraction and rule application.

What would have made it fail. Trying to handle unstructured or ambiguous cases the same way. The agent explicitly declines ambiguity and routes to a human.

8. GitHub's coding agent (post-hardening)

What it does. Autonomously handles issues in a repository — writes code, opens PRs, responds to reviewer comments, iterates. Explicit human approval before merge.

Deployment pattern. Sandboxed compute per task. Every action logged. No merge authority without human review. Post-2026 hardening: strict tool scoping following the prompt injection incidents earlier in the year.

What makes it work. Constrained agency. The agent has code-writing power but not merge power. Blast radius is real but contained.

What would have made it fail. Auto-merge. Some teams enabled this early. Some teams regretted it.

9. A KSA e-commerce platform's product-catalog agent (case study)

What it does. Automates product listing for merchants — takes a photo and rough description, generates SEO-optimized titles and descriptions in Arabic and English, tags with appropriate categories, sets up basic variants.

Deployment pattern. Merchant-facing. Optional (merchants can bypass). Feedback loop from published listings back to the agent's tuning.

What makes it work. Arabic quality specifically. Machine-translated Arabic e-commerce copy is bad; native-quality generation is a real differentiator in the market.

What would have made it fail. English-first with Arabic as a translation afterthought. Doesn't compete in the KSA and UAE markets.

10. Salesforce Einstein GPT in production

What it does. Broad set of capabilities inside Salesforce — email drafting, next-best-action recommendations, meeting prep, forecast assistance.

Deployment pattern. Deeply integrated into existing Salesforce workflows. Users encounter it in the flow of their existing work, not as a separate destination.

What makes it work. Enterprise scale, mature product surface, trust already established with buyers. Vendors sell easier here than startups.

What limits it. Customization ceiling. If your workflow doesn't match Salesforce's model, the agent's value drops significantly. This is where custom builds beat vendor offerings.

11. A Bengaluru insurance company's claims-processing agent (our client)

What it does. Reads submitted claims documents (photos, PDFs, hospital bills), extracts data, cross-references against policy terms, flags exceptions, and routes for adjudication.

Deployment pattern. Hybrid — high-volume standard cases fully automated, complex cases augmented (agent prepares the case, human decides).

What makes it work. The volume of standard cases (dental, routine outpatient) is enormous, and the pattern is stable. Automating them frees adjudicators to focus on complex cases.

What would have made it fail. Trying to auto-adjudicate complex cases. The client explicitly keeps human decision-making for anything with structural policy interpretation.

12. Anthropic's own Claude Code

What it does. Assists Anthropic's own engineers with the majority of their code work. Reportedly writes ~80% of the code that gets merged internally.

Deployment pattern. Deep integration with Anthropic's development workflows. Continuously refined by dogfooding.

What makes it work. Dogfooding at scale. Anthropic uses it more aggressively than any external customer, so bugs, gaps, and improvements are caught fastest.

What would have made it fail. Not being their own most demanding customer. Vendors that don't use their own product hard often ship products that don't hold up under real use.

Patterns across the twelve

Looking at all twelve, the successful deployments share consistent patterns.

Pattern A — Aggressive scoping

Every successful agent does a bounded set of things very well and refuses everything else. The failures we see are agents that try to be general when they should be specific.

Pattern B — Human-in-the-loop for consequential decisions

Refunds, merges, medical decisions, financial approvals — the successful agents keep humans in the loop for anything with real consequences. The failures we see are agents that acted autonomously when they shouldn't have.

Pattern C — Deep integration over standalone

Agents embedded in existing workflows outperform agents that require users to switch contexts. Sidekick inside Monday, Einstein GPT inside Salesforce, coding agents inside the IDE — all inline.

Pattern D — Cultural and language fit

Agents that speak the user's language natively (not translated) win. Especially in markets like the Middle East and India where language quality is a competitive dimension.

Pattern E — Proactive over reactive when possible

The LATAM rebooking agent, the collections agent — value comes from initiating contact at the right moment. Passive agents miss opportunities.

Pattern F — Instrumented from day one

Every successful deployment has observability. Cost per invocation, accuracy, escalation rate, user satisfaction — all measured. Unmeasured agents drift into failure states nobody catches.

The failure modes we see

Not covered above, but visible across the industry:

• Over-scoped agents. "AI does everything in your finance department." Never works.
• Under-instrumented agents. Nobody knows if they're working. Renewal fights inevitable.
• Autonomous where they shouldn't be. Bad tools, bad outcomes.
• Language-poor agents deployed in language-diverse markets. Loses to competitors.
• Cost-uncontrolled agents. Bill grows, ROI story fails.

Frequently asked questions

Which of these patterns is easiest to copy? Ticket triage / support deflection. Highest volume, clearest ROI, most vendor and framework support, most examples to learn from.

What's the highest-value pattern for a fintech? Collections and claims processing. Both are high-volume, rules-heavy, and multi-lingual — perfect fits for agent automation. See our regional Gulf fintech piece for adjacent context.

Do vertical agents rely on custom models? Rarely. Most use frontier models via APIs, small models for specific steps. See small language models. The vertical value is in the workflow, tools, and data, not usually in a custom-trained model.

How long does it take to ship one of these? Simple, well-scoped agents: 6-10 weeks. Complex ones with heavy integration: 3-6 months. Very complex regulated deployments: 6-12 months.

Are these vendors or custom? Mix. Roughly half of the twelve are vendor products (Salesforce Einstein, Monday Sidekick, GitHub coding agent) and half are custom builds. Both paths work when the choice matches the context (see our build vs buy framework).

The strategic read

The interesting question in 2026 is no longer "can AI agents work?" It's "which vertical, which pattern, which scope?" The twelve deployments above are a starting library of what works and why. Match one of these patterns to your business, execute with discipline, and you're likely to end up in the "shipped value" category rather than the "shelved after nine months" category.

At Xenolve we build vertical AI agents for enterprise clients across the Middle East and India — including several of the deployment patterns above. If a specific pattern here maps to a business problem you're wrestling with, get in touch. We're happy to compare notes even if we don't end up doing the work together.

Twelve real deployments. The patterns are learnable. The pitfalls are avoidable. Ship accordingly.