Most enterprise chatbots die quietly. A team ships one on the marketing site, it deflects a handful of password questions, and eighteen months later nobody can point to the person who owns it. Kore.ai built its business on the opposite premise: an assistant that handles money transfers, insurance claims and prescription refills needs the same governance, testing and analytics discipline as the core systems sitting behind it.
The company launched in 2014 and now sells into banks, insurers, telcos and hospital groups, with customers including AT&T, Cigna and Zurich Insurance. A $150 million round led by FTV Capital in early 2024 pushed the business past a billion-dollar valuation, which tells you how much enterprises are prepared to spend on automation that survives contact with real customers. For a wider view of how this fits into the sector, it’s worth reading about how Kore.ai is rewiring the way enterprises talk to customers.
What Kore.ai actually sells
Strip out the marketing language and there are two products. The first is a platform for building conversational assistants that run across chat, voice, email, SMS and messaging apps such as WhatsApp from a single design. The second is a set of packaged applications for contact centre automation, employee helpdesks and enterprise search.
The build layer spent years under the name XO. Kore.ai has since reorganised its stack around an agent platform, with modules for experience optimisation, contact centre AI, search AI and agent AI. Same engine underneath, different packaging on the box.
It is not an LLM wrapper
Kore.ai shipped its first products before large language models were commercially viable, and that history shows in the architecture. Dialogues are built as explicit flows with intents, entities and deterministic branches, then LLMs sit on top to handle flexible phrasing and long-tail questions. You can route through OpenAI, Anthropic, Google, Meta or Mistral models, or fall back to a smaller fine-tuned model for cost control, and pin high-stakes intents to fixed answers so a balance enquiry never gets improvised by a generative model.
A typical deployment includes:
- A visual flow builder with no-code and pro-code options for the same project
- Support for well over 100 languages, including voice-specific tuning
- Prebuilt connectors for Salesforce, ServiceNow, Zendesk, Genesys and similar systems
- Built-in testing, A/B experiments and analytics on containment, resolution and fallback rates
- Deployment choices spanning multi-tenant cloud, single-tenant VPC and on-premises
- Guardrails for PII redaction, policy checks and audit trails on every conversation turn
Where the platform earns its keep
The repeatable wins cluster around a handful of journeys. PIN and password resets sit at the top because volume is enormous and the process is boring enough to automate safely. Order and claim status, appointment booking, policy lookups and IT service desk requests fill out most of the rest.
Financial services is the most demanding customer base here, largely because regulators ask uncomfortable questions about how an AI system reached a decision. Kore.ai’s answer is logging and versioning: every conversation is traceable to a specific flow version and model configuration, which matters far more in an audit than a slick demo does.
Voice is the harder half
Chat deployments are forgiving. Voice is not. Speech recognition has to cope with accents and background noise, callers interrupt mid-sentence, and a response that takes two seconds feels broken. Voice projects routinely take two to three times longer than the equivalent chat build, and the failure modes are harder to spot in testing because real callers behave nothing like the scripted scenarios in a QA plan.
The internal helpdesk is the underrated use case
Customer-facing assistants get the headlines, but employee helpdesks often deliver cleaner returns. The user population is known, the knowledge base is finite, and the questions are repetitive. Teams running an internal assistant for IT and HR queries frequently report that more than half of tier-one tickets never reach a human, and the ones that escalate arrive with context already attached.
What it costs to run
Kore.ai does not publish list prices. Contracts are quoted against users, concurrent sessions or virtual assistant volume, with enterprise agreements typically landing in the six-figure range annually and larger multinational rollouts going well beyond that. Three cost lines get underestimated again and again:
- Integration work. Connecting to twenty-year-old backend systems is where the money goes. Budget two to three dollars of services for every dollar of licence in year one.
- Model tokens. If every conversation routes through a frontier LLM, inference costs scale with usage. Intent-based routing keeps a large share of traffic on cheaper paths.
- Ongoing content. Someone has to prune stale answers, retrain intents and review transcripts. It is a part-time job at minimum, and a full-time one at scale.
The limitations worth knowing before you sign
Integration sprawl is the big one. Assistants promise to do things, and doing things means APIs. Where an API does not exist, teams fall back on robotic process automation, which is effective until the underlying screen changes and the flow breaks silently.
Knowledge maintenance is the second trap. A bot that confidently returns a discontinued policy is worse than one that admits ignorance, and keeping retrieval sources current requires governance that most organisations do not have on day one. Then there is handoff design. Moving a conversation to a human agent with full context is straightforward in theory and fiddly in practice, particularly across multiple contact centre vendors.
How it stacks up against the alternatives
IBM watsonx Assistant appeals to organisations already deep in IBM infrastructure. Microsoft Copilot Studio wins on price and Teams integration but is thinner on regulated voice deployments. Google’s contact centre AI is strong on speech and telephony. Specialist vendors such as Yellow.ai, Ada and Sierra each carve out narrower ground, often with faster time-to-value on a single channel.
Kore.ai’s differentiator is breadth plus control: one platform for chat, voice, search and employee workflows, with deployment options that satisfy banks and healthcare providers who cannot send data to a shared cloud tenancy. If you need a bot live in three weeks on one channel, that breadth is overhead. If you need twelve channels and an audit trail, it is the point.
Running a pilot that actually tells you something
Pick one journey with measurable volume, not a sampler of six. Define success before launch: containment rate, escalation rate, customer satisfaction parity with human agents, and average handle time after handoff. Run for 30 to 60 days against a control group, log every fallback, and check repeat contact within seven days. Deflection looks great until the same customer calls back twice, and that number is where most business cases quietly fall apart.
Where the market is heading
The direction of travel is clear enough. Assistants that answer questions are becoming assistants that complete tasks: raising a dispute, rebooking a delivery, ordering a replacement card, chasing an internal approval across three systems. That shift pushes the hard problems from language understanding toward orchestration, permissions and observability, and it raises the stakes on getting guardrails right.
Regulation is catching up too, with transparency obligations landing on conversational systems in the EU from 2026 and similar rules queuing up elsewhere. The vendors that win the next five years will not be the ones with the best underlying model, since everyone rents the same handful. They will be the ones whose customers can explain, months later, exactly why an assistant said what it said.