Backbase releases AI agents that cancel bank cards and check payments within each lender's limits

Suppose self-service containment means interactions that end without a person. Then the South African figures can be turned into a count of work that still reaches people. At 20% containment, 80 of every 100 interactions reached staff. At 70%, 30 do ^[13]. Human-handled work per hundred interactions fell 62.5% ^[14]. That is a little over three times the 20% top of the operating-cost range McKinsey estimates for agentic AI ^[7][15].

The two numbers measure different things. The 30 interactions in a hundred that still reach a person are the ones the agent hands over because they need investigation or a human decision ^[3]. I'd expect the cost of each escalated case to rise even as their number falls.

The containment result comes from a single deployment, and both outlets carried Backbase's figures unchanged ^[5]. Neither report names the South African lender, says over what period containment rose, puts a number on any customer's savings or gives a price. Meriwest CTO Gene Fichtenholz said, "We've seen tangible savings through increased automation and call containment while giving customers a faster and more convenient service experience." ^[9]

For a bank weighing a build, the hard part is the control layer. Under Backbase, each bank decides which data and tools its agents can reach, which actions they may take and where a human must sign off ^[6]. Customers must confirm any significant action, and the context, checks, decisions and results are logged for audit ^[6]. The control half of Backbase's pitch rests on those confirmations and logs. Backbase founder and CEO Jouk Pleiter said, "That future has now arrived early: one where agents do the work, and banks stay in control." ^[12]

A bank with its own engineers could take that list of controls as a specification and keep the permission layer in-house. Contained conversations could also come back as repeat contacts. Part of McKinsey's projected saving comes from cutting repeated requests for information ^[7]. A containment rate measured only on first contact would overstate the gain. The South African lender's mix of requests may also suit automation better than other banks' mixes do.

I think the buy case holds for banks that lack the staff to build and audit an action layer of their own. The counter-thesis is that permissions and logging are routine work on systems a bank already runs, and that the language handling is the only part a vendor adds. The view is wrong if a bank of similar size builds an equivalent agent in-house and reports comparable containment, or if Backbase's other deployments publish containment far below the 70% seen in South Africa ^[5].