Technology, Media & Telecommunications

Connect AI to Product and Revenue Strategy

That could mean:

• TechnologyEmbedding AI directly into software products, search, development environments, or customer platforms.
• MediaCreating better discovery, recommendations, localization, audience experiences, or advertising products.
• TelecommunicationsUsing AI to improve network services, customer propositions, enterprise offerings, or new data-driven services.

This distinction matters.

A project that saves internal time can create value.

But an AI capability that makes the company's product more useful, differentiated, or valuable to customers can have much greater strategic impact.

Deloitte's 2026 TMT outlook highlights this growing tension between major AI investment and the need to convert that investment into sustainable monetization and practical customer value.

Design for Real-Time Digital Scale

TMT systems often operate very differently from traditional back-office workflows.

An AI feature that works well for 100 internal users may behave very differently when exposed to millions of customers or constant network activity.

Before implementation, TMT organizations should consider:

Latency

How quickly must the AI respond?

Availability

What happens if the AI service is unavailable?

Volume

How many requests must it handle?

Location

Should processing happen in the cloud, at the edge, or on-device?

Cost

What happens to inference costs when usage grows?

Deloitte expects AI inference—the running of trained AI models—to represent a large share of AI compute demand in 2026, making infrastructure design and operating economics increasingly important.

For TMT, scalability is not something to solve after the pilot.

It should be considered from the beginning.

Protect Content, Data and Intellectual Property

This factor is especially important in Technology and Media.

AI systems may interact with:

Before AI is allowed to use this information, organizations need clear answers to questions such as:

• Can this content be used by the AI?
• Who owns it?
• Is the organization licensed to process it?
• Can outputs be reused commercially?
• Should this data ever leave the organization's environment?
• How can generated content be traced back to its source?
• What should never be exposed to public models?

This is not simply an IT issue.

For media companies in particular, AI raises fundamental questions around copyright, licensing, provenance, talent rights, and protection of valuable content libraries.

Deloitte notes that major studios remain cautious about generative AI in core creative production partly because of intellectual-property and rights concerns, while adoption is moving faster in areas such as localization, marketing, distribution, and archive management.

Protect Trust While Increasing Personalization

TMT companies often use AI directly in customer-facing experiences.

That makes trust especially important.

More personalization can improve the experience.

But poor AI decisions can also damage trust very quickly.

Organizations should therefore think about:

Accuracy

Is the information reliable?

Transparency

Should users know AI is involved?

Bias

Could certain users or content be treated unfairly?

Brand safety

Could generated or recommended content damage the brand?

Privacy

Is personalization using customer data appropriately?

Human escalation

Can sensitive cases reach a person quickly?

This is particularly important for media, telecommunications, and consumer technology companies because AI decisions can immediately affect large numbers of users.

Engineer for Reliability, Not Just Intelligence

The smartest model is not always the best production system.

That means organizations need to monitor more than output quality.

Telecommunications makes this especially clear.

Network environments already require extremely high reliability. AI-driven network operations therefore need to improve resilience rather than introduce another unpredictable layer.

McKinsey's 2026 work on AI-driven telecom networks emphasizes using AI to improve network economics, reliability, and operational performance as network complexity continues to increase.

For production TMT systems:

Orchestrate AI Across Systems — Don't Create Another AI Silo

One of the biggest opportunities in TMT comes when AI can work across multiple digital systems.

Consider a telecom service issue.

An AI agent might need to:

1. Understand the customer's request.
2. Retrieve account information.
3. Check network status.
4. Search previous incidents.
5. Identify a likely cause.
6. Create or update a service ticket.
7. Recommend the next action.
8. Notify the appropriate team.

Or consider media operations.

An AI workflow might:

1. Receive new content.
2. Analyze the content.
3. Generate metadata.
4. identify rights restrictions.
5. prepare localization.
6. classify the content.
7. route it into publishing systems.

The value comes from orchestration.

This is becoming particularly important with agentic AI.

Deloitte expects practical progress in 2026 to depend heavily on the successful orchestration of AI agents rather than simply deploying isolated AI assistants.

That requires clear:

• APIs
• permissions
• action boundaries
• monitoring
• escalation rules
• human approval points

Make the Economics Work at Scale

This is one of the biggest differences between TMT and many traditional automation projects.

AI may create value—but AI also consumes infrastructure.

can have a compute cost.

For a small pilot, that may barely matter.

For millions of users, it can completely change the business case.

Before scaling, TMT organizations should understand:

• Cost per interaction
• Expected usage volume
• Cloud and inference cost
• Model size and performance requirements
• Whether smaller models can perform the task
• Whether some processing should happen on-device or at the edge
• How infrastructure costs change as usage increases

Deloitte's 2026 outlook highlights the growing importance of inference economics, power consumption, AI infrastructure investment, and the continuing gap between AI spending and monetization.

For TMT companies, AI economics can become part of the product economics.

Define the Outcome Before Choosing the AI

Start with the result you want to change.

Instead:

The more clearly the outcome is defined, the easier it becomes to decide what AI capability is actually required.

For telecom specifically, TM Forum's latest AI-native guidance recommends starting from an end-to-end business outcome, then identifying the capabilities, data, processes, and KPIs required to deliver it.

Decide How Much Authority AI Should Have

Not every AI system needs to act autonomously.

Before designing the architecture, define its level of responsibility.

A useful progression is:

For example, a telecom network agent might initially identify the likely cause of a fault and recommend remediation.

Only after the system has been validated extensively might it be allowed to execute certain low-risk changes automatically.

The GSMA's 2026 Agentic AI Testbed reflects this need for practical evaluation before putting autonomous agents into live telecom environments.

The goal is not maximum autonomy.

It is the right level of autonomy for the risk involved.

Choose the Right Model and Runtime Architecture

A proof of concept can often use a single cloud AI API.

A production TMT system may require something much more deliberate.

You need to decide:

1. Which model is appropriate?

   A large general model may not always be necessary.

   A smaller or domain-specific model may offer lower cost, faster responses, or better control.

2. Where should inference run?

   Depending on the application:

   • Cloud may provide flexibility.
   • Private infrastructure may provide greater control.
   • Edge processing may be necessary for low-latency applications.
   • On-device AI may make sense for consumer products.

3. How quickly must the system respond?

   • A media metadata workflow may tolerate several seconds.
   • A network-control system may not.

Deloitte's 2026 infrastructure research emphasizes that production AI is increasingly forcing companies to choose infrastructure according to latency, resilience, sovereignty, intellectual-property protection, and inference economics rather than defaulting to cloud for every workload.

Architecture decisions should therefore be based on the workload—not the popularity of a particular model.

Connect AI to the Systems That Run the Business

AI becomes operationally useful when it can work within existing TMT systems.

• For a Technology company, this might include:Git repositories, development pipelines, ticketing systems, cloud environments, product analytics, documentation, and support platforms.
• For Media:CMS, DAM, archive systems, rights databases, publishing platforms, advertising systems, and audience analytics.
• For Telecommunications:OSS/BSS, network-management systems, billing platforms, CRM, service assurance, telemetry, and customer-care systems.

Consider a telecom fault-management workflow.

A useful system could:

This is the difference between AI as a tool and AI as part of the operating system of the business.

TM Forum's AI-Native roadmap similarly emphasizes cross-domain workflows connecting data, decisions, actions, and control points rather than isolated AI applications.

Build the Trust Layer Before Production

TMT systems can contain highly sensitive assets.

• Technology companies may expose:source code, intellectual property, customer information, and production systems.
• Media companies may expose:licensed content, unreleased assets, talent rights, archives, and copyrighted material.
• Telecommunications providers may expose:customer data, billing information, network infrastructure, and service information.

Before production, define:

• who or what the AI is
• what it can access
• which actions it can take
• which actions require approval
• how activity is logged
• how access can be revoked

For media workflows, content provenance is becoming increasingly important as well.

Adobe began rolling out C2PA metadata support across AI-generated and AI-edited content workflows in 2026, allowing supported assets to carry machine-readable information about how they were created or modified.

Trust needs to be part of the architecture—not something added after deployment.

Test in a Realistic Environment

A successful demo is not enough.

TMT AI should be tested under the conditions it will actually face.

• For a software system, test unusual code, incomplete context, permissions, production failures, and conflicting information.
• For media, test different languages, rights restrictions, ambiguous content, sensitive material, and unusual formats.
• For telecom, test network failures, alarm storms, conflicting telemetry, missing data, and conditions the model has not previously seen.

Where possible, use a sandbox, digital twin, testbed, or shadow mode before giving AI production authority.

The GSMA's Agentic AI Testbed was created specifically because generic benchmarks are not enough for telecom. It allows operators to compare models and agents against real telecommunications scenarios, including network planning, capacity management, and fault troubleshooting.

A useful validation process should evaluate more than accuracy.

It should also evaluate:

latency, reliability, cost, security, escalation behavior, and the consequences of being wrong.

Operate AI Like a Production System

Once AI enters production, the project does not end.

It becomes an operational system.

This is especially important for AI agents.

Traditional application monitoring may tell you whether an API is available.

Agent monitoring needs to tell you whether the system is behaving correctly.

TM Forum now describes AI Operations and governance as core capabilities for moving telecom AI from experimentation into production-grade environments.

Scale the Capability, Not Just the Pilot

Once one AI implementation works, avoid building every new use case from zero.

Instead, identify reusable capabilities.

This turns AI from a collection of isolated projects into a platform capability.

It also reduces duplicated integrations, security controls, and model infrastructure.

TM Forum's 2026 AI-Native work specifically focuses on common data architecture, security, agent interaction, model operations, and reusable architecture so telecom organizations can move away from fragmented pilots and toward scalable AI operations.