From AI Experiments to Operational Adoption in Telecom

At a large enterprise technology company where I sold telco network solutions to the service providers of North America, the final days of any given quarter often came with the same standoff.

Months of drawn-out deal development, negotiation, scope wrangling and price haggling, finally a signed SOW and purchase order in hand. And just then, with hours to go before the books closed, the shared services team blocks the booking; the customer name on the PO differs from the name on the SOW. A missing LLC suffix. Sometimes a stray comma. “Please seek SVP approval for the legal entity name mismatch,” the email advises. At 9:40pm. On Black Friday. What fun.

The rule began as a workaround. The collections system couldn’t invoice against a PO name it couldn’t match, so a senior executive supplied the judgment the software lacked: a kind of Kafkaesque, pre-AI human-in-the-loop governance if you will.

The obvious objection to our complaints at the time was that this is a data problem: “Fix your data and you fix the problem!” someone might offer, unhelpfully. But the purchase order was not our data. Procurement issues whatever the PO says: a rebrand, an abbreviation, a new billing address, no shipping address at all. The incoming PO wasn’t subject to our internal governance, our data hygiene preferences, so deterministic fixed-rule processing meant escalations.

Bureaucracy is not exclusive to telecoms, most large companies harbor processes like this: a temporary fix for a specific failure that, a few reorgs later, is simply how we’ve always done it. In my customer work, today, these dark and dusty workflows are where agentification does its best work, not least because they have been forgotten by all but those who must endure them.

How to agentify

In an illustrative scenario, the AI agent is set up to scan the submitted booking documents for mismatches, then determines that the “Spectrum Gulf Coast, LLC” on the PO is a Charter subsidiary, formerly Time Warner Cable, and not the Charter entity named on the SOW. Then comes governance: the rules and policies that determine what you want to do with that reasoned information. Maybe you decide a same-entity match above 90% confidence clears itself, and anything the model reports as a subsidiary goes to a person for review.

Carriers have versions of this externally sourced data problem business-wide. Subscriber identities move between CRM, ordering, credit, billing and collections platforms. Product and address data travel through order management, inventory, provisioning and assurance systems, and each platform may represent the customer, service or network resource differently.

One system stores the ZIP+4 extension, the next accepts only five digits, and the address stops matching. Some of that is an internal schema problem, fixable in principle. But service addresses come from the customer, formatted however the customer likes. It’s not just an input problem, order fallout happens at any link in the chain, there is a permit delay or the modem is unexpectedly out of stock and the service request falls into a manual queue. One carrier told me its cost was $1 million per percentage point of affected orders.

How much of these operations should be automated? One way to look at this is to rank agentic outcomes by consequence. Retrieving records and explaining a discrepancy carry relatively little risk. Changing a subscriber master record, restricting service, or escalating a debt can have significant impact. Those actions should require more human input, segregation of duties, with an auditable record of the evidence, policy, and authorization.

As automation is dialed up, high-confidence, low-risk exceptions can move to approved agentic workflows. Ambiguous or high-stakes cases should stay with accountable people.

Today I recommend that practical adoption begins with measurement. Before deployment, exception volume, handling time, escalation rate, error rate and financial delay should be established. After deployment, evidence accuracy, reviewer acceptance, resolution time and any inappropriate actions should be measured. Every override should be recorded and fed back into policy, testing, and workflow design. That gives the carrier a controlled learning loop without asking the model to learn indiscriminately from every human decision.

And in the end, who knows, maybe Sales and Shared Services will get to enjoy their Thanksgiving this year.