Financial Services

AI-assisted fraud narrative generation

A retail bank's fraud team reviews hundreds of flagged transactions daily, each requiring manual cross-referencing before writing a case narrative. Raw VSAM transaction records contain packed decimal amounts, internal offsets, and system codes irrelevant to the analyst. A REXX script flattens the VSAM structure and converts packed decimal values to readable form; DFSORT filters the output to the 12 fraud-relevant fields and caps record volume. Connect/AI passes each prepared record with a structured prompt to a cloud model, writing a plain-language case summary back to dataset for analyst review each morning.

Outcomes

• ● Case prep time reduced from 25 mins to under 5
• ● VSAM flattening handled by existing REXX skills
• ● DFSORT controls field selection and payload size
• ● Audit trail maintained entirely on z/OS

Insurance

Automated policy exception commentary

A life insurer runs nightly batch processing across millions of policies. When a policy hits an exception — lapse risk, beneficiary conflict, underwriting anomaly — it is written to a VSAM KSDS with a deeply nested, repeating-group structure. Compliance rules prohibit policyholder data leaving the on-premises environment. IDCAMS REPRO extracts the relevant exception records; a COBOL program restructures the variable-length records into a consistent flat layout, normalises date and currency fields, and strips fields not relevant to the exception type. Connect/AI targets a locally-hosted inference model, generating a plain-English explanation and suggested resolution path for each exception.

Outcomes

• ● Policyholder data never leaves on-premises
• ● COBOL normalisation reuses existing program skills
• ● Exception queue resolution time cut by 40%
• ● Compliance posture unchanged

Government

SMF-based capacity insight reporting

A government agency’s infrastructure team produces weekly capacity reports from SMF record types 30, 70, and 72 — covering CPU utilisation, DASD activity, and WLM metrics. Raw SMF data arrives as undecoded binary records, previously handled by a mix of SAS macros and manual effort. Standard SMF extraction tools decode the binary records to a workable sequential format; a REXX script then normalises field representations and aggregates readings to hourly intervals to reduce inference payload size, flagging records that breach capacity thresholds. Connect/AI passes the prepared dataset to a z/OS-hosted model requesting a trend summary and risk narrative, delivered by email to infrastructure management before the weekly review.

Outcomes

• ● SMF decode automated using standard extraction tools
• ● REXX aggregation replaces manual SAS macro work
• ● Capacity risks surfaced before the review meeting
• ● All processing sovereign — no data leaves z/OS

Mainframe Operations

Direct JES report inference for job failure triage

A mainframe operations team runs hundreds of batch jobs daily via JES2. When jobs fail or produce anomalies, operations staff manually scan spool output — JESMSGLG, JESJCL, SYSOUT — to identify root causes. With Connect/AI, the raw JES report is passed directly to the inference model, which understands fixed-width columnar formats, JCL syntax, and standard IBM message codes natively. No ETL, field extraction, or data preparation pipeline is required. The model reads the spool output as-is, identifies the failing step, interprets the relevant system messages, and drafts a triage summary routed to the operations team via Connect/Hybrid.

Outcomes

• ● No ETL or field-mapping pipeline required
• ● JES report format understood without pre-processing
• ● IBM system message codes interpreted in-context
• ● Faster mean time to triage on batch job failures
• ● No changes to existing JES2 job definitions

Logistics

Intelligent exception routing for supply chain events

A global logistics operator processes millions of shipment events daily through a z/OS core system, with exception records stored in DB2. Each record contains the full shipment history — hundreds of columns including rate tables, routing history, and carrier codes — far exceeding what an inference model needs. A DB2 extraction tool selects only current-day exceptions; DFSORT filters to the 15 impact-relevant fields and translates internal carrier and status codes to human-readable equivalents via a lookup. A two-step inference job follows: a lightweight local model classifies severity; for high-severity events only, a cloud model drafts a customer communication, routed via Connect/Hybrid into the CRM for agent review.

Outcomes

• ● DB2 extraction uses existing toolchain — no new skills
• ● DFSORT lookup handles code translation in-pipeline
• ● Cloud API costs limited to high-severity events only
• ● No mainframe application changes required

Performance Engineering

AI-driven Omegamon performance analysis and reporting

A z/OS performance engineering team monitors system health using IBM Omegamon — navigating multiple 3270 monitor screens daily to assess CPU, memory, I/O, and address space activity. Producing the weekly capacity and incident trend report requires manually visiting Omegamon screens, transcribing key metrics, and composing a narrative summary — typically a half-day task. A Python script automates the entire collection process through Connect/3270; DFSORT or REXX step merges and sizes the multi-screen data; Connect/AI then passes the prepared dataset to a cloud model requesting anomaly detection, trend commentary, and a recommended actions summary. The finished report is emailed to the performance team and uploaded via FTP to the infrastructure document store.

Outcomes

• ● Weekly report prep reduced from half a day to an overnight scheduled job
• ● Screen navigation fully automated — no engineer involvement in data collection
• ● AI identifies anomalies across all screens simultaneously
• ● Omegamon untouched — no configuration changes required
• ● Output delivered by email and FTP — no new tooling for consumers

Developer Productivity

Interactive prompt development and data exploration

A developer building a new fraud-narrative batch job needs to tune the prompt template before scheduling it. Rather than submitting a test job and waiting for each result, they open the Connect/AI workspace, load a sample of prepared records, and iterate on the prompt in real time — adjusting wording, comparing model outputs side by side, and confirming the output format. Once satisfied, the same prompt drops straight into the production JCL step. The interactive workspace and scheduled batch share one engine, so what works live works in production.

Outcomes

• ● Prompt tuning cycle cut from hours to minutes
• ● No test batch submissions required to iterate
• ● Same prompt promotes directly into production JCL
• ● Sample data never leaves z/OS

Mainframe Operations

Agentic incident triage across live mainframe data

When a production batch stream fails overnight, an on-call operator opens the workspace and asks Connect/AI to investigate. Acting agentically, it retrieves the relevant JES spool, reads the failing step and system messages, checks related datasets, and proposes a root cause and remediation — carrying context across each step and surfacing its reasoning as it goes. The operator confirms the fix without hand-scanning spool output across multiple jobs. The same agentic engine can then be scheduled to triage failures automatically and route summaries to the on-call queue.

Outcomes

• ● Multi-step investigation runs autonomously
• ● Operator reviews reasoning, not raw spool
• ● Faster mean-time-to-resolution on job failures
• ● All access under SAF, entirely on z/OS