Confluent: streaming infrastructure, governance and real-time AI for financial systems

Streaming Agents and Agent Management Console became generally available in May 2026. Agents run on Flink, consume events and use configured models and tools. The console exposes prompts, models, inputs, outputs and runtime logs. Individual newer AI functions can have a separate early-access or preview status; the general availability of the agent framework does not make every function production-supported.[10]

The runtime supports function and MCP tools, state recovery and logged interactions. Replay allows teams to compare changed agents against historical events without repeating external side effects. Reflection workflows use a drafter and critic until approval or an iteration limit. This is a software control loop: a model critic’s approval does not establish human review or the correctness of an investment, fraud or compliance decision.[9]

Connector controls include configuration checks, logs, role permissions and single-message transforms for masking or filtering. Specific connectors support private-network routes, modern cloud authentication and field-level encryption. Those capabilities must be checked against the particular source or sink rather than assumed for the whole catalogue.[12]

For remote model inference, the customer registers a model endpoint and supplies authorised credentials. Current documentation lists Anthropic, AWS Bedrock and SageMaker, Azure OpenAI and Azure ML, Fireworks AI, Google AI, OpenAI and Vertex AI. A Flink compute pool and permission to create models are required. Model hosting, provider terms, credential lifetime and external processing remain dependencies of that integration.[11]

IBM’s March 2026 completion announcement identifies immediate integrations with watsonx.data, IBM MQ, webMethods Hybrid Integration and IBM Z. These connect streaming data to analytical, messaging, integration and mainframe workloads. The announcement supports these named links; it does not make every IBM product part of a Confluent subscription.[13]

Confluent combines published Cloud usage rates with negotiated commitments and separate software subscriptions. The following website rates are US dollars, vary by cloud region and exclude the other chargeable components. Storage figures are presented before replication; Freight’s figure is an effective comparison rate.[14]

Elastic eCKUs apply to Basic, Standard, Enterprise and Freight; Dedicated clusters use provisioned CKUs. The documentation retains a legacy billing model for Basic and Standard organisations created before 16 April 2024. Its GB means 2³⁰ bytes. Connector tasks and throughput, Flink CFU-minutes, governance, link traffic, audit-log reads, Tableflow processing/storage and support can add charges. Tableflow upserts also incur index-processing charges.[15]

Annual commitments can obtain negotiated volume discounts, while qualifying automatic tiers reduce unit rates as usage grows. These are different mechanisms. Private Cloud is priced per node, above Platform Platinum node pricing, and Unified Stream Manager may require additional licensing. A complete enterprise quote therefore depends on deployment and selected services.[14][3]

The published November 2023 Cloud Services Agreement bills usage monthly in arrears, payable within 30 days, unless the written order says otherwise. The subscription term and committed payments follow that order. The contract index provides separate agreements for certain contracting entities and products.[17][18]

Cloud onboarding creates an environment, selects a cloud and region, provisions a cluster and defines topics. Producers, consumers and connectors receive appropriate credentials. The documented quick start then produces sample data, checks messages and inspects the registered schema. Cloud-management API credentials and Kafka-cluster credentials serve different purposes; a new cluster requires its own corresponding access configuration.[20]

Professional services include activation, architecture design, migration accelerators, production hardening, configuration optimisation and health checks, with custom statements of work. Education combines self-paced study, instructor-led training and certification. These offerings support deployment and ongoing skills; the public material does not prescribe one implementation duration for every institution.[21]

Trial use has a narrower contractual scope than production. Under the November 2023 Cloud agreement, trials are for internal testing on development or non-production clusters, carry no service-level commitment and must not contain production or legally regulated data. A promotional credit does not, by itself, establish a production entitlement.[17]

Deleting Cloud resources stops their continuing metered usage, but it does not cancel incurred fees or an order’s remaining commitments. The Cloud agreement makes the customer responsible for exporting content before discontinuation. The DPA separately provides return or deletion of customer personal data at the customer’s election when the agreement ends.[15][17][26]

For EU/EEA-registered customers and services within its defined scope, the September 2025 Data Act Addendum sets a switching process: at least two months’ notice through the support portal, followed by reasonable transition assistance. The standard transition is 30 days after the two-month notice period ends. A notified technical extension can run up to three months after that period, and the customer can request an extension. Contractual fees remain payable.[19]

The October 2025 support policy defines business hours as 09:00–17:00, Monday–Friday, in the customer’s local time zone. Requests enter through the portal; support is in English, with subsequent email or video communication and scheduled telephone calls when needed. First-response targets measure initial engagement, not resolution. P1/P2 cases require customer personnel to supply diagnostics. Customer code, unsupported configuration and third-party systems have stated exclusions.[23]

Kafka Streams support excludes Developer, requires extra fees and applies only when used with Cloud, excluding self-managed brokers. Premier’s technical-architect entitlement excludes BYOC unless expressly included in the order.[23]

Business and Premier coverage for custom connectors is limited to the Connect infrastructure. Self-managed component support requires the appropriate licensing. Premier’s technical-architect service includes office hours and periodic technical reviews, subject to conditions. Cloud plan selection therefore does not automatically cover every locally operated component.[22]

The July 2025 Cloud SLA distinguishes service functions and deployment configurations. Kafka Basic has a 99.5% threshold; Standard and Enterprise move from 99.9% at one eCKU to 99.99% with a minimum of two eCKUs. Dedicated requires multiple zones for 99.99%. The connector SLA covers designated supported, fully managed connectors; it excludes customer-uploaded, custom, preview and Connect with Confluent connectors. These are conditional contractual availability measures, not measured performance reported by this profile.[24][16]

A credit claim must arrive within five calendar days after the affected month ends and include configuration and failure evidence. Accepted credits apply to future bills, expire after 12 months and are the SLA’s exclusive availability remedy. Exclusions include specified customer, network, third-party and underlying cloud-provider failures.[24]

The September 2026 Security Addendum says customers choose the region storing Kafka message content. It specifies encryption at rest using AES-256 or equivalent/better, TLS 1.2 or higher in transit, logical customer separation and restricted personnel access with multifactor authentication. Its PCI scope requires customer-side message-level encryption. The document describes SOC and ISO certifications and a route to request supporting evidence; this review did not inspect the underlying audit reports.[25]

The October 2025 DPA distinguishes customer personal data from account data. It limits customer-data processing to documented lawful instructions, supported service functions, troubleshooting and legal permissions. Account information has separate administrative, security and service-improvement purposes. Subprocessors are listed and changes can be followed through a subscription. Customers remain responsible for credentials, appropriate backups and the protection of sensitive information they send.[26]

Remote AI inference sends configured inputs to the model endpoint authorised by the customer. The documentation explains credentials and providers, but does not establish one universal no-training or retention promise covering every external model service. Those conditions follow the selected deployment and provider. An endpoint connection is not evidence that generated output is validated for a financial decision.[11]

For an institution, Confluent is data infrastructure. Its post-trade architecture joins trade and order information, applies validation, manages out-of-sequence events and distributes updates between processing stages. Brokerage permissions, market-data redistribution rights, trade execution and settlement responsibilities remain separate from the software’s ability to transport and process those records.[32]

IBM acquired all outstanding common shares for $31 each in cash, with an announced enterprise value of approximately $11 billion. The completed transaction is the current ownership position. This profile covers the Confluent platform and associated deployment family; IBM’s wider software portfolio remains outside that product scope.[13]

Operator

Confluent, Inc.

Service scope

Event-streaming infrastructure, stream processing, connectors, schema governance, table materialisation and AI context/agent workflows for financial and other enterprises.

Customer types

Banks; Financial intermediaries; Exchanges; Payment businesses; Digital-asset businesses; Enterprise data teams

Delivery

SaaS; On premises; Hybrid; API

Pricing publication

Public and quoted pricing

Headquarters

Mountain View, California, United States

Platforms

Confluent Cloud; Confluent Platform; Confluent Private Cloud; WarpStream; Apache Flink; Stream Governance; Tableflow; Confluent Intelligence; Streaming Agents; Real-Time Context Engine

Integrations

Apache Kafka APIs; PostgreSQL; Oracle; DynamoDB; Salesforce; Snowflake; Amazon Redshift; Databricks Unity Catalog; Snowflake Open Catalog; Microsoft OneLake; MCP; IBM MQ; IBM Z; watsonx.data; webMethods Hybrid Integration

Pricing scope

Public Cloud usage and support rates, negotiated annual commitments, and separate deployment subscriptions; region and component charges apply.

Implementation

Cloud or customer-managed deployment, topic and schema design, credentials and network configuration, connector mapping, application validation and staged migration; advisory services and training available.

Support

Cloud support tiers range from account-only assistance to 24/7 technical coverage; response targets, custom-connector scope and self-managed licensing differ by plan.

Permissions and regulatory scope

Software and data infrastructure. Product security controls and financial-industry implementations do not confer brokerage, custody, execution or market-data redistribution permissions.

Provider metadata checked 21 Oct 2026Source 1Source 13Source 27Source 2Source 3Source 4Source 5Source 6Source 7Source 8Source 9Source 12Source 14Source 22Source 23Source 20Source 21Source 32

Recent changes affect operational design. On 19 October 2026, managed MCP servers gained Confluent-account OAuth sign-in with the user’s existing permissions. Earlier October changes added custom-connector plugin/version management and inherited release priority. Automatic Failover for Cluster Linking, announced on 30 September, remained an Early Access Program feature: it pairs supported Enterprise or Dedicated clusters across regions behind one bootstrap address and requires access through the account or support team.[28]

Maintenance also includes customer-side libraries. The 1 October 2026 advisory CONFSA-2026-22 affects Python confluent-kafka versions through 2.14.2 only when client-side field-level encryption uses the HashiCorp Vault KMS integration over HTTPS. TLS certificate verification was disabled in that path. The advised fix is 2.15.0 or later, with the appropriate CA configuration for private certificates. The notice explicitly states that Confluent Cloud and Confluent Platform are not affected.[29]