Bain Telco AI Costs Warning Exposes an Opex Trap

Bain has issued a telco AI costs warning with a sharp reversal: automation spending can increase operating expenses even after jobs and tasks disappear. The firm estimates that AI agents, tokens, and related data services could eventually represent 20% to 30% of an operator’s cost base.

The warning challenges a common assumption behind telecom AI programs. Cheaper models and fewer manual tasks do not guarantee a smaller operating budget. If operators keep their existing systems, contracts, approvals, and human workflows, AI becomes another expense layered onto the old structure.

That conflict matters because operators are already applying AI across customer care, network operations, software engineering, and business support systems. AT&T is experimenting with model routing and open models, while Microsoft is promoting specialized network agents. Bain’s argument is that these tools create savings only when operators redesign the work around them.

Bain’s Telco AI Costs Warning Changes the Savings Equation

Bain’s central message is that telcos can replace part of their cost base with AI, or add AI spending without removing the costs it was meant to replace.

The firm published its telecom analysis on September 10, 2026. It describes a likely shift in telecom operating structures during the next three to five years. AI agents will perform or support more work, while people retain responsibility for critical decisions.

Bain presents a possible operating model in which traditional expenses account for 70% to 80% of total costs. AI agents, token consumption, and related services make up the remaining 20% to 30%.

That model is not itself a forecast of savings. It describes a change in what operators pay for and how those expenses behave.

Under the favorable scenario, AI costs replace a comparable share of legacy spending. Operators remove redundant work, simplify processes, reduce unnecessary software, and change their staffing requirements. Total costs might remain broadly stable while speed, capacity, or service quality improves.

The dangerous scenario begins when an operator adds the same AI expenses without removing much of the existing base. Employees continue following established processes. Software subscriptions remain active. Outsourcing arrangements continue. AI assistants handle isolated steps but do not eliminate handoffs or duplicate systems.

Total operating expenses then rise by as much as the new AI layer. Productivity, customer experience, or revenue must improve enough to justify that increase. Otherwise, the operator has funded a more complicated organization rather than a more efficient one.

The distinction is important because operating expenditure, commonly shortened to opex, recurs as the business runs. A failed capital project can leave a visible asset and a defined write-down. Token usage, cloud inference, monitoring, and human review can instead produce variable monthly costs distributed across departments.

AI agents also create spending through more than model access. They call software tools, retrieve stored information, evaluate outputs, maintain context, and sometimes repeat unsuccessful steps. Operators must pay for the infrastructure and controls surrounding those actions.

Bain therefore rejects cost per token as the main measure. A cheap token does not reveal whether an automated workflow solved a problem. The useful unit is the cost of resolving a customer issue, handling a network incident, producing a proposal, or completing a software release.

The original reporting captured the starkest implication. Neither of Bain’s illustrated paths automatically reduces total costs. One replaces traditional spending with AI spending, while the other increases the combined bill.

That is what changed. The telecom AI conversation is no longer only about whether models can automate work. It is becoming a financial test of whether operators can retire the processes, contracts, and organizational layers that automation makes unnecessary.

Cheaper Models Can Still Produce Bigger Bills

Falling model prices matter less when usage, workflow complexity, and demand grow faster than the cost per token declines.

Bain says model prices have fallen roughly tenfold each year. That sounds like a direct path to lower AI spending. However, lower unit costs often encourage broader use, larger prompts, more complex reasoning, and longer-running agents.

An employee who once used AI to summarize one document may later ask it to review an entire case history. A network team may move from generating incident notes to coordinating diagnosis, remediation, validation, and escalation. Each improvement expands the amount of context, reasoning, and tool use involved.

Agentic AI magnifies that effect. An AI agent is software that can plan and execute a sequence of actions, rather than answering one prompt. It may consult several systems, reconsider a failed result, and call another model before completing one task.

Those loops can consume tokens without producing an acceptable outcome. Repeated context can be sent with every model call. Multiple guardrails may inspect the same output. Two agents may perform overlapping checks because separate teams designed them independently.

The model invoice also captures only part of the expense. A production workflow may require orchestration software, storage, observability, evaluation, cybersecurity controls, and human supervision. Operators can optimize inference while overlooking the larger system surrounding it.

Demand creates another source of pressure. Once employees discover useful applications, adoption spreads across business units. Network operations and customer care are especially exposed because they contain high-volume workflows with large amounts of operational context.

Teams may also default to newer reasoning models for routine work. These models can provide better analysis, but their longer reasoning chains consume more tokens. Paying less for each token does not help when a task uses many more of them.

This pattern resembles the rebound effect seen in other technologies. Efficiency lowers the cost of an individual activity, which encourages people to perform more of it. Total consumption can rise even though every unit becomes cheaper.

Telecom operators face an added complication because their workloads are persistent. Customer requests, alarms, tickets, configuration changes, and fraud checks continue around the clock. A small increase in the cost of each automated event can become significant at network scale.

The correct comparison is therefore not one model against another. Operators must compare an AI-enabled workflow with the full cost and outcome of the process it replaces.

That calculation should include completion rates, retries, escalations, errors, supervision, infrastructure, and downstream corrections. It should also recognize value that does not appear as immediate labor savings, such as faster restoration or lower customer churn.

The result can favor AI even when direct spending rises. A network agent that shortens a major outage may produce substantial business value. A chatbot that creates longer interactions without resolving more requests may do the opposite.

This is why Bain telco AI costs analysis focuses on cost per completed task. It forces operators to connect spending with a measurable result, rather than celebrating lower token rates or higher adoption.

The Real Opex Trap Is the Dual Operating Model

AI becomes an opex trap when it performs new work around the edges while the legacy organization continues operating unchanged.

Telecom companies rarely start with clean systems. Many operate multivendor networks, customized business support platforms, long-term outsourcing agreements, and software accumulated through years of expansion and acquisition.

Adding an assistant to that environment can make one step faster. It does not automatically remove the remaining steps.

Consider a customer service agent using AI to summarize a call. The summary may save time, but the employee still handles authentication, searches several systems, requests approvals, updates records, and resolves the customer’s issue. The operator still pays for every underlying platform and organizational handoff.

A network operations assistant creates the same risk. It may classify an alarm or recommend a response, while engineers continue using the established monitoring, ticketing, and escalation process. The AI output becomes another artifact that employees must review.

Bain calls this a dual operating model. The human-led workflow remains intact, while AI performs fragments around it. Licenses, outsourcing costs, management layers, and manual controls survive because the underlying process never changes.

End-to-end redesign produces a different result. An operator begins with the outcome, maps every step required to reach it, and decides which steps should disappear. AI then participates in a simpler process rather than attaching itself to inherited complexity.

Bain cites Vivo’s closed-loop network operations work as an example. According to the consultancy, the operator redesigned a complete detect-to-resolve process. AI can detect anomalies, identify likely causes, perform corrective actions, validate results, and escalate exceptions.

That structure targets the whole incident rather than one analytical step. Human attention moves toward cases requiring judgment or intervention. The workflow can generate value even if AI spending remains substantial because fewer manual handoffs delay resolution.

However, operators must treat vendor contracts as part of the redesign. If agents take over work previously performed through software seats, external service providers, or outsourced teams, those commitments should be reconsidered. Otherwise, automation creates technical savings without financial savings.

Workforce planning requires equal care. Removing experienced employees before the new process works can weaken service quality and institutional knowledge. Keeping every role unchanged after automation matures can preserve duplicate costs.

Gartner added a broader warning on September 9. Its future of work forecast predicts that 30% of employees laid off because of AI replacement will need rehiring by 2029. The firm expects many returning hires to cost more because talent pipelines and institutional knowledge will have weakened.

That prediction covers industries beyond telecom, but it directly reinforces Bain’s pressure test. A rushed reduction can make the initial business case look attractive, then produce recruiting, training, service, and remediation costs later.

The goal is not to protect every existing activity. It is to avoid treating headcount reduction as proof that the redesigned system works.

Operators should first establish whether AI can complete the targeted workflow at acceptable quality and risk. They can then change roles, contracts, approvals, and systems around verified performance. That sequence connects organizational change to operating evidence.

The opex trap is therefore not simply expensive inference. It is the survival of yesterday’s cost base underneath tomorrow’s automation layer.

AT&T Shows What Token Economics Looks Like in Practice

AT&T’s response illustrates the alternative: match model capability to each task and design orchestration around total workflow cost.

Bain says AT&T examined how to process billions of tokens each day more efficiently. The operator redesigned its orchestration so large coordinating agents could delegate work to smaller, specialized models.

According to AT&T’s account cited by Bain, that approach cut costs by up to 90% while tripling throughput. The numbers represent a company claim, not an independently audited industry benchmark. They still show what a serious cost-control strategy tries to optimize.

AT&T calls its approach tokenomics. In this context, tokenomics means managing model consumption as an operating resource, with choices based on task complexity, performance, and cost.

The operator has explored open models because they provide more control over deployment and optimization. Its tokenomics strategy does not depend on sending every request to the largest available model.

A routine classification task may need a small model. A complicated network diagnosis may justify a more capable reasoning system. A coordinating agent can route work between them instead of using one expensive model for every step.

This strategy matters because model selection often happens invisibly. Employees choose a familiar tool, or developers set one default model for an application. The cost difference seems small during a pilot but expands as usage reaches production volume.

Routing creates its own costs and risks. The operator must evaluate which model can reliably handle each task. It needs fallback rules, performance monitoring, security controls, and clear boundaries for human review.

Still, those mechanisms support a better financial question. The operator can ask what combination of models, tools, and people resolves the task at the required service level.

Microsoft’s experience supplies a useful comparison. Its telco industry CTO said large anomaly-detection models added substantial complexity while providing limited improvement over traditional statistical methods. The company shifted attention from predicting every network failure toward responding quickly when one occurs.

That admission is valuable because it separates a technically interesting use case from a financially useful one. A more sophisticated model does not create value when simpler methods perform nearly as well.

Microsoft continues to see promise in network operations and business support systems. Its Network Operations Agent framework uses specialized agents for functions such as telemetry, ticketing, troubleshooting, and field coordination.

Far EasTone Telecom has used that framework in its network operations center. According to Microsoft’s deployment account, nearly 60% of the operator’s network operations center activities are AI-assisted.

Assisted does not mean autonomous. Microsoft maintains human supervision for changes to its own network, even where agents support advanced operational workflows. That distinction limits operational risk but also preserves part of the human cost.

This is the central tradeoff in Bain telco AI costs management. Operators want greater automation, yet reliability requirements demand evaluation, approvals, and accountability. Removing oversight too early can create outages or incorrect actions. Keeping excessive oversight can erase the economic benefit.

AT&T’s approach does not eliminate that tension. It makes the tension measurable. Model routing, dedicated budgets, and per-task accounting let leaders see where greater autonomy improves the outcome and where it merely increases consumption.

The practical lesson is not that every telco should copy one architecture. It is that AI spending needs a designed control system before adoption reaches billions of daily tokens.

What the Cost-Savings Narrative Still Does Not Prove

Bain offers a useful operating model, but its percentages are scenarios, not verified savings or losses across the telecom sector.

The 70-to-30 structure is an estimate of how costs might evolve. Bain notes that its illustration does not incorporate absolute changes in expenditure. It should not be interpreted as a prediction that every operator will spend the same share on AI.

Operators have different network footprints, labor structures, outsourcing arrangements, regulations, and cloud strategies. A model that fits one company may be inappropriate for another.

The cost of a completed task also depends on the outcome selected. A customer interaction can be measured by handling time, resolution, satisfaction, retention, or revenue. An automation program can look successful under one measure and weak under another.

Network incidents create a similar problem. Faster diagnosis helps, but the result matters only if the recommendation is accurate and the repair succeeds. An agent that generates quick but unreliable answers can increase review and correction work.

Bain’s argument also assumes operators can identify and remove legacy costs. That is often a contractual and organizational challenge rather than a technical one. A software agreement may run for years. A regulated control may remain mandatory. An outsourced function may cover activities that AI handles only partially.

Demand uncertainty further complicates the forecast. Token consumption can accelerate as employees find more uses. It can also stabilize when governance improves, smaller models mature, or conventional automation replaces unnecessary AI calls.

Benefits may appear outside the operating expense line. Better network reliability can reduce churn. Faster proposal generation can improve sales capacity. Automated rural kiosks can serve markets that would not support a staffed store.

Bain points to a Telia pilot involving AI-enabled self-service kiosks in rural communities. The example changes the economic question from reducing existing work to serving demand that was previously uneconomical.

That is a growth case, not a simple cost-cutting case. It can justify new AI expenditure if the service creates sufficient revenue or customer value. Treating every project as an opex reduction program would miss that distinction.

There is also a risk that companies use the warning to slow worthwhile deployments. AI expenses are real, but so are delays, outages, repetitive work, and customer frustration. Avoiding every variable cost would preserve the inefficiencies operators want to address.

The better response is disciplined experimentation tied to business outcomes. A pilot should establish quality, completion cost, adoption, and operational risk. It should also identify exactly which old expense can disappear if the pilot scales.

This is where proofs of concept often fail. A demonstration can show that a model performs a task. It rarely proves that an organization can remove a workflow, renegotiate a contract, or change accountability around that capability.

The skeptical reading of Bain’s warning is therefore balanced. Its analysis does not establish that AI will increase every operator’s total costs. It shows why lower model prices and impressive pilots are insufficient evidence of savings.

The burden now shifts to operators. They must publish or internally validate task-level economics, not simply adoption figures, token volumes, or staffing announcements.

Three Signals Will Show Whether Telcos Escape the Trap

The next test is whether operators connect AI consumption to completed outcomes, retire duplicate costs, and expand proven workflows without weakening service.

The first signal is task-level reporting. Operators should track the full cost of resolved customer cases, network incidents, software releases, and sales proposals. Those figures need to include models, infrastructure, evaluations, retries, tools, and human review.

If task costs fall while completion quality holds steady, Bain’s favorable path gains support. If token rates decline but the total cost per outcome rises, the opex trap is taking shape.

The second signal is visible retirement of legacy expenses. Operators should show that successful automation leads to fewer redundant applications, shorter process chains, revised outsourcing contracts, or redesigned roles.

Staff reductions alone are not enough. Gartner’s rehiring forecast highlights the risk of removing people before automation, training, and governance are ready. Sustainable savings require the process itself to change.

Evidence that operators are reducing duplicate software and manual handoffs would strengthen Bain’s thesis about workflow redesign. Continued growth in both AI spending and legacy contracts would support its cost-creep scenario.

The third signal is controlled expansion from pilots into high-value production workflows. Bain recommends selecting three costly, difficult processes for end-to-end redesign. Customer care, network operations, and software delivery are obvious candidates.

A successful expansion should produce more than a demonstration. It should establish ownership, acceptable error rates, human escalation rules, and financial accountability. Every agent should have a named owner and measurable outcome.

Watch how operators discuss these programs in financial updates and operational briefings. Claims about the number of agents or tokens consumed reveal activity, not value. Metrics tied to resolution, restoration, retention, throughput, and total cost reveal whether the operating model changed.

The next several months will not settle the long-term economics of telecom AI. They can reveal whether executives have accepted the immediate accounting challenge.

Bain telco AI costs research leaves operators with a concrete choice. They can keep adding assistants to old workflows, or rebuild selected workflows around measurable outcomes. Buyers and employees should ask which costs disappear when each new agent arrives. If leaders cannot answer that question, the promised efficiency remains an expense forecast rather than a business result.