Build or Buy Agentic AI? The Strongest Case for Each, and the Layer That Decides It

Practitioners call this the context layer: the ontology of entities, relationships and business definitions that tells an agent what a customer, a device, a fault and a resolution mean inside a particular communications service provider (CSP). This essay argues that the decisive build-versus-buy line runs between the machinery of agentic AI, which CSPs can buy, and its meaning, which they have strong reasons to own.

The case for buying: time is the scarcest input

The empirical record favors purchase, with MIT NANDA's 2025 study of more than 300 public AI deployments finding that purchased tools and partnerships succeeded roughly 67 percent of the time, about twice the rate of internal builds. Critics question the study's sampling, yet its direction matches what many technology leaders observe in their own portfolios.

The other argument concerns time: Gartner found that AI projects needed an average of eight months to move from prototype to production, and only 48 percent completed the journey. An internal program adds recruiting to that clock, in a market where CFOs rank acquiring AI talent as their top near-term challenge, "not easy and it's expensive" (Gartner, March 2026). But in the meantime, the market keeps moving.  PwC projects global mobile ARPU to slip from $6.32 in 2024 to $6.20 by 2029. Cable MVNOs took 39 percent of U.S. smartphone net additions in the third quarter of 2025, according to MoffettNathanson, and Deloitte expects more than 1,000 direct-to-device satellites in orbit during 2026. Under those conditions, a slower, internal program costs market position as well as resources that take away from quick, competitive maneuvering.

Model economics reinforce the point. Epoch AI estimates that the price of a fixed level of model performance falls between 9x and 900x per year, depending on the task, so a model a CSP trains today depreciates faster than most programs can deploy it. Bought capability, by contrast, delivers quickly: Bell reports 25 percent fewer complaints with Gemini-driven agents, and Deutsche Telekom a 95 percent cut in network event handling time with Google's RAN Guardian.

The case for internal build: compatible agents can still disagree

The opposing camp starts from where projects can fail. Gartner finds that 63 percent of organizations lack, or are unsure they have, the data management practices AI requires, and predicts that organizations will abandon 60 percent of AI projects unsupported by AI-ready data through 2026. Readiness here means agreed definitions, lineage and business rules. A vendor can host them, but it cannot author a CSP's definition of a customer at risk, because that definition encodes the CSP's commercial strategy.

Buying at scale also carries a cost the purchase price conceals. Half of the 2,000 CEOs in IBM's 2025 study say rapid investment has left them with "disconnected, piecemeal technology," and only 16% of their AI initiatives have scaled enterprise-wide. Gartner expects the average Fortune 500 enterprise to run more than 150,000 agents by 2028, while only 13 percent of organizations consider their agent governance adequate, and it predicts the cancellation of more than 40 percent of agentic AI projects by 2027. Protocols such as MCP and A2A, which Deloitte identifies as the connective tissue of multi-agent systems (Deloitte Tech Trends 2026), make vendors technically compatible. Coherence remains unsolved, because each vendor's agent arrives with its own implicit definitions of a customer, an issue and a resolution. Without a central layer of meaning, every added vendor raises integration cost and lowers the odds that agents agree.

The deeper argument is economic. When one layer of a stack commoditizes, value migrates to its complements; as model capability becomes cheap, the scarce asset becomes the context that tells a model what the business means. That asset creates an unusual lock-in. Migrating data is an engineering project, whereas migrating definitions changes how every agent built on them behaves. Telstra's chief architect has argued that level four or five network autonomy "is going to require a standardized, ontology-driven approach" (IEEE ComSoc, April 2026). Two CSPs running the same bought model over the same bought context layer will converge on the same agents.

The line runs between machinery and meaning

Read together, the two cases conflict less than their advocates suggest, because they concern different layers. The buyers make the stronger case for machinery: models, agent runtimes, compute and even data hosting, where hyperscaler scale and the clock both favor purchase. The owners make the stronger case for meaning: the ontology, its definitions and the governance that keeps them consistent across every vendor's agents. A CSP can run its context layer on hyperscaler infrastructure and still own it, provided it authors the schema, controls changes and can carry it elsewhere.

Owning means also costs far less than owning machinery. The industry already has a shared information model in TM Forum's Information Framework (SID), which gives CSPs a starting vocabulary. Governing an ontology takes a small, senior team with cross-functional authority, a modest commitment beside a platform program. Data residency rules may still constrain where the layer runs.

MCE Systems: The Middle Ground for What You Choose

Wherever an organization draws the line, the decision still leaves a practical question of execution, and MCE Systems works on both sides of it.

For those that opt to build internally keeps full ownership of its ontology and its agents. MCE Systems supports that path by supplying additional signals from the customer's device, including battery health, storage, app performance and network experience as the handset records it. MCE Systems delivers this as structured data that feeds the CSP's existing data lakes, so internal teams can enrich the customer models they already run and draw stronger insights from them. The CSP's teams keep control of the meaning layer they chose to own.

Organizations that prefer to build with a partner can bring MCE Systems into specific parts of the program rather than the whole. TELUS's retail brand Mobile Klinik took this route. MCE Systems supplied the mobile device data and set up Eva, the AI agent that engages Mobile Klinik customers in the app. With live device-health context, Eva produced three times higher take rate on care-to-commerce offers and four times more customer engagement with the app. Braden St. John, VP of Products, Marketing and Logistics at Mobile Klinik, calls agentic AI in the app "the future of service."

The build-versus-buy debate resolves once a CSP decides which layer carries its meaning. Machinery will keep getting cheaper to buy, and meaning will keep compounding for whoever owns it. Either path gains from a partner that can take on part of the work without claiming the whole.

Explore how structured device intelligence fits into a CSP's context layer: talk to MCE.