AI should finish the job.

Most enterprise AI stalls after the pilot

Wonderful starts from a pattern its chief executive, Bar Winkler, says he saw across nearly 50 deployments in 18 countries: companies can get capable AI, and the trouble starts when it has to work inside the business. The model is rarely what stops it. What stops it is the call-routing rules of a telecom, the compliance steps of a bank, the scheduling system of a hospital, and the language and habits of each country. A McKinsey survey quoted when the two firms announced an alliance in April 2026 put the gap in numbers: 79 percent of organizations were experimenting with generative AI, and fewer than 10 percent had scaled AI agents.

companies don’t have a technology problem; they have an implementation problem.

Bar Winkler · Co-founder and CEO, Wonderful · Company essay, The Execution Gap No Model Can Close · 2026-01-07

The bet: delivery is part of the product

The belief that sets Wonderful apart is that delivery is the product, not an extra sold as professional services. Winkler writes that real AI transformation needs sustained investment on the ground in delivery "as a first-class capability, not an afterthought." In practice the company places engineers and strategists inside each customer's environment, in each country it sells in, and pairs them with its own platform. Insight Partners, which led the Series B and the Series C, wrote that one of its strongest convictions was that enterprise AI adoption would be won locally. From the start Wonderful aimed at markets where English is not the first language, tuning its agents for language, culture and regulation.

Why now is partly technology. Winkler says the story began when GPT-4 arrived and voice agents became fast enough for a phone call; as he put it in a May 2026 interview, a caller cannot wait three or four seconds for an answer. He and his co-founder, Roey Lalazar, started work at the end of 2024. By Winkler's account the company was established on 5 January 2025, had a product ready for customers at the end of February and signed its first deal in April. The other part is fatigue with pilots. In a February 2026 interview he said that for large companies the problem is not whether the technology exists, but whether it can be delivered so that it actually works in their systems.

One request, followed through the team

The path below is illustrative and follows the documented Telefónica Colombia deployment. A telecom wants an AI agent to answer billing questions. A Deployment Strategist, the role Wonderful sums up as making the work matter, leads discovery with the customer's operations owners: which questions arrive most, where they fail, and which result will count as success. A Forward Deployed Engineer, the role that makes it work, connects the agent to the systems that hold the answers, such as billing and payment records, telephony and identity checks. The agent then compares billing periods and explains a price change by voice or on WhatsApp.

People keep the decisions that carry risk. The customer agrees the scope and the success measures with the Deployment Strategist, and operational tools let a person step in when needed. After launch the pair watch live conversations, fix the cases that only appear in production, and look for the next workflow. Wonderful's own writing says this work after go-live, tuning responses, handling edge cases and updating the agent when policies change, is often as large as the initial build and never fully stops.

The bet is easy to state and hard to scale, because every new customer needs people who can sit inside it and ship. Watch whether hiring keeps pace: Wonderful had 350 employees in March 2026, planned about 900 by the end of the year, and reported 650 in September 2026.

For someone joining in Seoul, this chapter describes the job itself. Wonderful says each market gets a full local team of forward-deployed engineers, deployment strategists and go-to-market staff, and the Korea office was set up to supply exactly that team.

One platform, sold with the team that installs it

Wonderful sells a platform, which it now calls the AI OS, together with the people who put it into production. The company describes the AI OS as one operating layer for employees, agents and apps. Its Series C release lists four kinds of products on that layer: managed workflows that automate complex processes end to end, productivity agents that support employees, AI-native applications that complement or replace legacy software, and conversational agents that serve customers. Permissions, integrations and audit setup are built once and reused, so that, in the company's words, each new use case launches faster than the last.

Wonderful Agents is where the company began, and where most of its named results come from. An agent takes work from any channel, voice, chat, email or documents, and completes it end to end, reading data, triggering actions and updating records in the systems the work runs on. Teams assemble agents in a workflow editor that chains triggers, functions, MCP tools and model calls. In January 2026 the company launched Agent Builder, an agent powered by Anthropic's Claude that builds, tests and refines other agents. Wonderful says it cuts time to production by up to 50%.

Two other lines take the pitch beyond the service desk. Under the Systems name, Wonderful maps how a business function really runs, builds AI-native software around it, migrates the data from the legacy tool and then runs the new software for the customer. The AI Gateway picks a model for each task and enforces the customer's cost, security and performance policies. Its page puts the commercial argument in one line, "Model companies earn when you consume more," and says the gateway is there so customers ship more per dollar.

Customers, before and after

Telefónica Colombia, which trades as Movistar, is the clearest before and after. It had evaluated over 20 providers and already ran an AI solution, but that system served land lines, left broadband customers out and was slow to extend. The first ask was modest: match what the old system did. Wonderful's agent now explains bills and price increases by voice and WhatsApp for mobile, fixed-line and convergent customers. The case reports 91.5% containment on eligible interactions, average handling time 50% lower than with the previous AI solution, and call volume up 2.5 times in two months while NPS held steady. Its headline says the agent resolves 77% of issues; containment and resolution are different measures.

Integrating a fully generative AI agent into OTE's environment is not a trivial task: it requires deep investment in mapping and bespoke integrations, many different APIs, legacy systems, and strict operational requirements.

Yannis Vitzilaios · CIO, OTE Group · Customer quote in Wonderful's OTE case study · 2026-05-21

Later cases repeat the pattern in other industries. At OTE Group, Greece's largest telecom operator and part of Deutsche Telekom, the Cosmote TV agent covered 10 skill areas; deflection reached 50%, nearly triple the baseline, and handling time fell 30%. Banco Caja Social went from zero to a live collections agent in nineteen days, and in the first six weeks after launch, 65% of the agent's direct contacts ended with a promise to pay, against a 45% human baseline. Maccabi, an Israeli health provider, has its agent handle 45,000 calls a day on its appointments and service-information line. At Bank Hapoalim, a bank data engineer new to agents built a production agent in three weeks, reusing 15 tools that Wonderful's engineers had built for earlier deployments.

How to read these numbers

Every result above comes from Wonderful's own case studies, and each uses its own measure (containment, deflection, resolution, promise to pay) against its own baseline. Read them as evidence that agents reached production at named companies, and compare a figure only with the same measure.

The detail that repeats is time to production: nineteen days at Banco Caja Social, three weeks at Bank Hapoalim, four weeks at PPC Energie, ten weeks for three agents at Mercado Libre. A Korea hire would be asked to produce the same kind of first result at a Korean enterprise.