Any Agent Can Run Fast. IBM Bob Has to Testify.

The ransomware arrived through a Windows server and spread. M.R. Williams, a family-owned distributor running mission-critical applications on IBM i, contained it, hardened the environment and then went looking at the decades of RPG code that had quietly kept the lights on.

That is where Jasmine Kaczmarek, the company’s vice president of technology, produced the number IBM cites in its collateral. A senior developer had spent six hours the previous day tracing a report’s field logic. Kaczmarek did it in 20 minutes.

Impressive. Also, for a chief data officer, nearly beside the point.

Speed is the easiest thing to demo and the hardest thing to govern. IBM Bob, generally available since late April 2026, lands in a market already crowded with tools that write code fast. The question for anyone who owns data policy is narrower and meaner: when the agent touches a governed asset, what happens next?

The gap where the value escaped

Sandeep Sivadas, who works on Bob at IBM, describes the prior state of the art like a product owner.

“A year ago the answer was kind of deflating,” he says. “AI could read your data catalog just fine, but it couldn’t do anything with what it found. You’d get a recommendation, and then a person still had to go make the actual change in the actual system.”

That gap is where the return on investment evaporates. “Every vendor had a version of the ‘knowing’ half,” he says. “Almost nobody had solved the ‘doing’ half.”

What IBM built instead is a loop. Bob calls watsonx.data through an MCP server while it works, rather than looking things up once before it starts. It pulls the governed definition of the asset it is about to touch, the lineage behind it, and any data-protection rule sitting on it.

Then the real useful part begins. “If that rule says this thing needs sign-off, Bob doesn’t just print that out for a developer to notice and act on later,” Sivadas says. “It stops itself and routes to an approval checkpoint automatically. If there’s no restriction, it just goes, makes the change, and logs exactly which definition and which policy check cleared it.”

Note what Sivadas doesn’t claim: Bob has no privileged view of the data. “Any assistant hooked into that same MCP server can see the same governed data,” he says. “The difference is what Bob does with it. That data doesn’t just sit there informing the agent. It’s actually steering it.”

Peel off the adjectives and the CDO pitch gets simple. Governance metadata stops being reference material and starts behaving like a control signal, one that changes what the agent does while it is doing it.

Where the math changes

Closing this matters in today’s enterprises, which already own a coding assistant and a data catalog. The main problem is that these two just don’t speak at the moment that matters. “The coding assistant knows about code. The catalog knows about data,” Sivadas says. “Neither knows about the relationship between them at the moment of execution.”

But understanding this relationship really matters to chief data officers. “For a CDO, that changes the math on one specific question: can I let this agent act, or do I have to keep a human in every loop?” Today the answer is almost always the human, because the agent’s actions can’t be traced back to governed definitions in a way that satisfies an auditor. Bob’s claim is that “the evidence trail to support it exists from the first action forward.”

“Any assistant hooked into that same MCP server can see the same governed data. The difference is what Bob does with it. That data doesn’t just sit there informing the agent. It’s actually steering it.” — Sandeep Sivadas @ IBM

Whether that survives a regulator’s questioning is an open experiment, but it is being run at scale. At BNP Paribas, a 20-person pilot expanded into full development squads as business lines onboarded themselves. IBM Consulting has more than 12,000 licenses deployed across retail banking and airline engagements. That’s production, not a lab.

The rest of the numbers make chief financial officers and chief information security officers look closer. EY Tax Americas reports a 60% cut in logic-implementation effort and four weeks from pilot to production. Blue Pearl compressed a Java modernization from 30-plus developer days to three. CrushBank, whose business is unlocking knowledge stranded in systems never built for AI, halved its delivery timelines running Bob alongside watsonx.data and watsonx Orchestrate, and now finishes proofs of concept in hours.

The black box problem

Bob’s multi-model routing is its most-discussed feature, and it exists partly to kill decision fatique. RedMonk’s senior industry analyst Kate Holterhoff told The Register that it is double-edged, since developers can be suspicious of black-box tools. However, she admitted, “This may be a path forward as vendors like GitHub are forced to change billing structures to absorb the high cost of running these models.”

Sivadas reframes it as an operating principle. “You don’t assign your best security architect to write boilerplate test cases.” Bob decomposes a task, routes each piece to the best-suited model, and logs which model handled what. “The governance layer is what makes it safe to let that happen without a human approving each routing decision.”

Futurum’s vice president and practice lead for software lifecycle engineering Mitch Ashley argued that buyers will increasingly evaluate these tools on evidence-layer strength, making auditability the unit of value. Jason Andersen, vice president and principal analyst at Moor Insights & Strategy made a cooler assessment in his comments to TechTarget, seeing Bob as a tool that fills “a crucial gap in IBM’s product portfolio and creates a new entry point into its underlying IT automation toolset for existing customers,” without outrunning its competitors on features.

Sivadas also outlines the ceiling, which vendors rarely do. Bob underperforms on what he calls “genuinely novel architecture decisions,” greenfield work where the right answer depends on strategic choices that were never in the codebase. “That's still human-led work, and it should be” — which is the main point.

For APAC banks sitting on COBOL and RPG, Sivadas makes a sharper point. The scarce asset isn’t the code. “It’s the business rules embedded in the code that nobody has written down.” A quick win transforms the code. A program extracts those rules into governed form first, building institutional memory while the people who hold it retire.

The metric to focus on

Asked what proves Bob changed anything a year out, Sivadas points to Bobalytics, IBM’s usage telemetry. It tracks how widely Bob has been adopted, then a Bob Commit Factor measuring how much committed production work Bob touched, then consumption.

“Third, we’ll look at Bob Coin consumption and usage patterns,” he says. “They help organizations understand not just adoption, but depth of usage.”

But it is his other marker that holds up much better, because no agent can game it. “The shift I look for is when the conversation moves from, ‘Did Bob work in the pilot?’ to, ‘Which teams and applications do we bring onto Bob next?’”

Everyone is deploying agents. Far fewer can defend one in front of an auditor. Kaczmarek’s 20 minutes make a good slide. The log sitting behind them is what a CDO will actually have to stand on.

Image credit: iStockphoto/iweta0077