AT&T says it has nearly 500 agentic AI workflows in production and plans to triple that number next year. The company’s approach offers a useful look at how AI agents are moving from workplace experiments into customer-facing business operations, with human oversight as a gate rather than a permanent fixture.
Watch Desk analysis
What happened
SiliconANGLE reports that AT&T processes about 45 billion AI tokens a day and that its agent platform handles more than one million transactions a month. The account draws on comments from Luke Corbin, AT&T Business’s assistant vice-president of digital transformation and AI.
The workflows include customer provisioning and sales prospecting. In one example, automation handles parts of the process for changing customer fibre-installation orders. AT&T also says it has halved the time needed to provision dedicated internet connections in buildings where fibre is already installed.
AT&T uses a small set of agents to carry out workflows with centrally managed business logic, rather than building a large collection of specialised agents. Employees can make personal productivity agents, but customer-impacting workflows go through the company’s Business Agent Foundry and security, privacy and compliance reviews.
New production workflows start with human oversight. According to Corbin, they can operate autonomously only after reaching accuracy in the mid-to-high 90 per cent range within an acceptable risk profile. AT&T says it continuously monitors their performance for drift.
Why it matters
The figures describe more than a chatbot trial: AT&T says agents are handling substantial volumes of transactions and performing work across existing business processes. The examples also show where automation could matter to customers, from order changes to faster internet provisioning.
The safeguards are part of the deployment model, not a footnote. Human review is the starting point, and autonomy is conditional on accuracy and risk. That still leaves practical questions about how accuracy is measured and what happens when an agent gets a decision wrong.
Our read
The interesting move is from automating isolated tasks to redesigning whole workflows, while keeping customer-facing changes behind review and performance gates. That is a more useful test of agentic AI than a flashy demo: what work is actually being done, and who is still accountable when it goes wrong?
AT&T’s examples and performance figures come from an account of its own deployment, not an independent audit. They are a concrete case study, though, and the human-oversight threshold gives readers something specific to scrutinise rather than another promise that agents will transform everything by Tuesday.
What to watch
- Whether AT&T reaches its goal of tripling the number of production workflows next year.
- How it measures accuracy and defines an acceptable risk profile before allowing autonomy.
- Whether the reported time savings and returns hold up across more workflows and customer-facing services.
Discussion spark: Should customer-facing AI agents be allowed to operate autonomously once they meet a stated accuracy threshold, or should a person approve every consequential action?
Sources and evidence
- AT&T uses agentic AI to reinvent business processes from the ground up – SiliconANGLE (14 November 2026, 19:37 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.