Own your learning loop

The core problem: The biggest infrastructure decision telco will make this decade is whether to own its AI learning loop or rent it. Telcos that put AI decisions inside the foundation model—instead of in an ontology they own—are extending decades of BSS vendor lock-in at machine speed. The architectural fix is to separate decisions from the model and put them in a layer the operator controls. The Totogi Ontology is a pre-built operational layer for telco that lets operators own decision precedent, exception logic, and validated rules.

In 2014, at TM Forum Live in Nice, T-Mobile announced it was moving off Amdocs, replacing them with Ericsson across billing, charging, order management, product catalog, and CRM. CTO Neville Ray put T-Mobile’s name on the exit as the most insurgent “uncarrier” in modern telco, saying it would walk away from the largest BSS vendor in the industry. 

More than a decade later, in February 2026, T-Mobile renewed its Amdocs contract for another multi-year term. This time with GenAI on top. 

Why can’t T-Mobile quit Amdocs? Because Amdocs owns T-Mobile’s learning loop, and it won’t let it go.

Well, there’s hope for T-Mobile, in the form of AI. A few weeks ago, Satya Nadella, CEO of Microsoft, published a great blog on what he sees as the key to enterprise AI. His thesis: The real opportunity with AI isn’t in choosing the right model, but in building a “learning loop” on top, which the enterprise should own. The organization’s workflows, domain knowledge, and accumulated judgment all should be encoded into a proprietary system that compounds with every use.

You don’t want your learning loop in the model, but many telcos are about to make this mistake. They are planning AI deployments where the AI makes decisions at inference time, inside the model. The pitch sounds clean: ask the model, get the answer, take the action. Satya reminds us that the models are commodities. They are interchangeable inputs. None of them holds anything that belongs to you.

The trouble with telcos’ approach is that a decision made inside the model is a decision made jointly—by you, and by whoever owns the model. Different models give different answers to the same question, and as a result you aren’t building your learning loop. You’re renting reasoning by the query.

The fix is architectural. Separate the decision from the model. Let the model do the inference work. Insist that the decision lives in a layer you control, where you can see it, change it, and grow it.

The question his thesis forces on every telco CIO: Do you own your learning loop?

Telcos outsourced their knowledge

Telco already outsourced its learning. Decades ago. The receipts are everywhere; you just stopped adding them up. Look at where your operational knowledge actually lives today:

1. In consultants’ heads: Amdocs (and every other long-term incumbent vendor) has been inside your company for twenty years. You know that senior consultant who knows why your provisioning state machine handles credit memos differently on the third Tuesday of the month? She’s on Amdocs’ payroll, billing you by the hour, and her knowledge goes home with her every night. You did not hire her. You cannot hire her into your own org. Every operational expert of every domain of your business has a vendor logo on their business card. That’s where the learning from thirty years of running your network has been built and where it resides today.
2. In code nobody can access: Your operational knowledge lives in five million lines of COBOL (Java if you’re lucky) inside the billing system, modernized but not understood, every business rule encoded as a function call somebody wrote in 2003 and stopped explaining in 2007. Same goes for the rate plan logic, the eligibility rules, the exception handling, and the reason a B2B customer with a past-due balance can still upgrade if they signed up in 2014 but not if they signed up in 2018. Somewhere in your stack, a routine does the right thing for the right reason and nobody around can tell you why. The code fires away but the documentation (and the knowledge) is lost.
3. In your own veterans’ heads: Thankfully, you do own this; the problem is, they can leave at any time. They are brilliant people, and walking single points of failure. Most of them will retire in the next five years, and the knowledge inside their head will leave the building with them. How do you replace that?

None of these places are queryable, executable, or transferable. None of them were written in plain language anywhere a new employee, a new vendor, or a new model can read. Forget about Satya’s test about owning your own learning loop; you’re not even at the point where it’s inside a model. It’s inside people and inaccessible code. And good luck getting help from your vendor to transition the knowledge. They have contract clauses for that.¹ (If you want proof, see the footnote below for a link to a snippet from an old Amdocs contract.)

Worse, your incumbent vendor has no plan to help you out of this mess. They are pitching the same outsourcing learning loop model wrapped in agentic AI. Call it aOS, call it whatever they brand it next quarter; they are agents bolted on top of the same managed services contracts, doing the same work at machine speed. The architecture is faster. The trap is identical. Agents handle your churn interventions, your retention offers, your order remediations—and the learning compounds in the vendor’s agentic layer, not yours. This is the 1990s contract optimized for velocity.

You have to break the cycle, and AI is the way.

Here’s your chance

For the first time, telcos have a way to encode operational learning into a layer the business actually owns. Remember the veteran who knew why the third-Tuesday credit memo rule existed? Her reasoning can be captured, made queryable, and turned into something a new employee or a new model can read. The 17 reasons your churn model misfires on prepaid-to-postpaid? They need to live in a system, not inside someone’s head.

Put the decision in a layer between the data and the model—which we at Totogi (where I’m CEO) call an ontology. Not a data dictionary, not a metadata catalog, not whatever the BI vendors are now repackaging as “semantic consistency.” It’s a living, executable, write-back operational layer encoding how your business actually works: every state transition, every eligibility rule, every exception precedent your operators carry in their heads. Every action flows through it. Every exception enriches it. Every rule it learns lives there, queryable in plain language, executable in production.

That is the architecture that lets you own your learning loop. How can you test that you really own it? Three ways:

Swap your foundation model. GPT-5 to Claude to whatever ships next year. Learning survives. The ontology lives in your stack, not in the model’s.

Swap your BSS vendor. Amdocs to Oracle, or any combination underneath. Learning survives. The adapters under the ontology abstract vendor schemas; your operational logic does not care which billing engine sits below it.

Bring managed services in-house and run it with your own team. Fire the consultants, hire your own ops people. Learning survives. The ontology is the brain. Decision precedent becomes queryable. Exception logic becomes visible. The operational knowledge that took thirty years to accumulate inside someone else’s runbook now lives in your system, where it learns from every new decision to compound in value for your organization, not your vendors’.

Three swaps, same answer: that’s what ownership looks like.

Totogi is building this

Building your own learning loop will not be easy. But it IS solvable. And we are making it happen for Tier 1 operators at Totogi.

When telcos deploy our ontology, they can finally see their decision structure. It’s not in a consultant’s head, or in code, or even in a document. It’s a visual representation of how the telco makes decisions, callable by code: Totogi’s, another vendor’s, or your own generated applications.

For one customer, we showed them that their enterprise sales team only follows the “happy path” process 20% of the time. Ten days wasted waiting for Legal. More time wasted waiting on VP approval—and the VP ALWAYS approves the exception.

The Totogi Ontology lets your SVP and domain expert control the actual decision-making, so the process models reality. Bottlenecks go away. AI decides instead of waiting on humans. The sales cycle goes from 66 days to 35. That’s revenue this quarter, not next. It’s what owning your learning loop looks like in real life.

The Totogi Ontology is a pre-built operational layer for telco. The base—semantic consistency, the knowledge graph, the actioning—is our IP, built on TM Forum SID and ODA. Open standards. Built to be connected to, accessed, and used.

What accumulates on top—your data, your decision precedent, your exception handling, your validated rules—is yours. For the first time, you control your operational expertise.

Transforming for AI is the biggest infrastructure decision telco will make this decade. It’s going to take work. The question is whether you do the work once, the right way, and own what you build. Or whether you do it twice—once for AI, and again ten years from now when you realize you signed up for the same trap at machine speed.

Own where your learning lives. You already know who owns it today. The next AI contract you sign is when you decide whether you’re going to break the cycle or extend it. I say, break it.

^{1. In 2012, Amdocs filed a Customer Care and Billing Services Agreement with the SEC. Section 6.5 is titled Transfer Assistance: ten pages governing what happens when a customer tries to leave. They call it Disentanglement. I guess that also implies that you are entangled. At least their lawyers got the word right.}

Recent Posts

Get my FREE insider newsletter, delivered every two weeks, with curated content to help telco execs across the globe move to the public cloud.