AT&T's autonomy gate for 500 workflows: a model for NZ automation

By DataForge team17 November 20264 min read

AT&T keeps a person on every new AI workflow until accuracy reaches the mid-to-high 90s. NZ firms can run the same gate at a much smaller scale.

What AT&T disclosed

Luke Corbin is AT&T Business's assistant vice president of digital transformation and AI. At an AWS event in Boston on 4 November, Corbin said the telco has close to 500 agentic workflows live and plans to triple that count in 2027 (TokenPost). SiliconANGLE published a case study on 14 November. It adds that AT&T's agentic platform handles more than 1 million transactions a month and processes about 45 billion AI tokens a day (SiliconANGLE).

The volume figures will get the headlines. For a mid-sized NZ operator, the more useful detail is how AT&T decides when a workflow can run without a person watching. Every production workflow starts with human oversight. An agent is allowed to work unsupervised only once its accuracy sits in the high 90s and the remaining risk is judged acceptable. Any workflow that touches customers or materially affects the business also has to pass security, privacy and compliance review, and continuous monitoring checks for drift after release (SiliconANGLE). Employees can still build personal productivity agents without going through that review. AT&T's controls therefore sit on the work that carries business risk and leave individual experimentation alone.

Process mapping comes before the agent

AT&T set up a Business Agent Foundry in early 2025. The Foundry maps each business process against industry frameworks and looks for unnecessary steps, handoffs between teams and dependencies on legacy applications. That mapping happens before anyone builds an agent (SiliconANGLE). AT&T runs a small set of agents against centrally managed business logic instead of a large population of specialised bots.

One example involves changes to customer fibre installation orders. Staff used to pull tickets from queues and update mainframe records by hand, and agents now handle parts of that work. In buildings that already have fibre, AT&T says it has halved the time needed to provision dedicated internet service.

Most NZ automation backlogs we see are built the other way round: staff suggest task ideas, and each one gets automated on its own. Task-by-task backlogs produce quick wins, and they also leave behind flows that nobody clearly owns. AT&T's top-down mapping takes longer to get going. A redesigned process does give the team an obvious place to put the human checkpoint, and it gives one named person the decision about when to remove that checkpoint. Our process and automation work starts from the same principle.

What AT&T hasn't published

AT&T's figures are self-reported. Corbin claims production AI investments return five times their cost within a year (SiliconANGLE). Rajesh Beri's analysis notes that AT&T quoted a two-times return as recently as January. Beri also points out that the company hasn't said how it scores accuracy or weighs its "acceptable risk profile" (The Daily Brief). Outside AT&T, nobody knows whether the accuracy figure is measured per step, per transaction or per model.

How accuracy is defined matters as much as the threshold itself. Take a five-step workflow where each step succeeds 97 percent of the time. The whole workflow will finish correctly only about 86 percent of the time. A 3 percent error rate across a million monthly transactions works out to roughly 30,000 wrong outcomes (The Daily Brief). Beri also notes that AT&T's AI gateway sends each request to the cheapest model likely to meet the quality bar. An accuracy test passed on one model may therefore stop holding when the gateway picks a different one.

Why the gate suits NZ and Australian firms

Very few NZ organisations run automation anywhere near AT&T's volume, but the gate works just as well at small scale. Picture an accounts payable team putting 2,000 invoices a month through an AI extraction workflow. The team could compare the workflow's output against human review for a few weeks, then write down a threshold before turning the review step off. Even at 97 percent accuracy, that team would get about 60 invoices wrong each month. The finance manager needs to decide whether an error budget of 60 invoices is acceptable before the workflow runs unsupervised, and needs a plan for how those errors will get caught downstream.

Firms with Australian customers also face a near deadline. From 10 December 2026, entities covered by the Australian Privacy Principles must say in their privacy policy when they've arranged for a computer program to make, or do something substantially and directly related to making, decisions that could significantly affect an individual's rights or interests. The disclosure has to cover the kinds of personal information used and the kinds of decisions made solely by the program (OAIC). A register showing each workflow's autonomy status is the natural source for that disclosure. NZ businesses with Australian operations should confirm with their advisers whether the obligation applies to them.

Setting thresholds and naming who owns the autonomy decision is a governance task as much as a technical one. Our AI consultancy team covers that kind of work.

Practical next steps

  • List every automation and AI workflow in production. Record the owner, the model or service the workflow depends on, and whether a person currently reviews its output.
  • Define accuracy for each workflow in business terms, such as an invoice coded to the correct account, and measure it end to end across all steps.
  • Before removing human review, write down an autonomy threshold and a monthly error budget based on real transaction volume. The business owner should sign off on both.
  • Re-run accuracy checks whenever the vendor changes the underlying model, your business rules change, or a router starts sending work to a different model.
  • Keep personal, self-serve automations apart from workflows that touch customers or financial records, and send the customer and finance workflows through security and privacy review.
  • If you serve Australian customers, identify which workflows make or substantially inform significant decisions about individuals, and update your privacy policy before 10 December.

Want to know what this means for your organisation? Our team in Auckland can show you your top three opportunities in a free 30 minute assessment.

Get my free assessment How we help
Get new articles by RSS