Capabilities
A capability is something AI can do for the business, described by what it does rather than how it is built: “answer tariff questions with the source cited”, not “use RAG”. It is a means to a business outcome, not the outcome itself.
It is the top level of the taxonomy in 1.2, and it stays the same when the product underneath changes. The eight below are common examples for a telco, not a complete list; a programme will name its own.
Of these, four rest on generative AI, three on classic machine learning, one on either. That split matters more than any vendor’s list: it decides cost, explainability and how predictable the output is.
Generate & summarise
Draft, summarise and rewrite text, code and images.
- Call summaries for the care team
- Offer and campaign copy
Does not solve: Whether the content is right.
- Capability
- Generate & summarise
- Solution
- Call summaries and wrap-up
- Pattern
- AI workflow
- Standard
- OpenAPI, OAuth 2.0 / OpenID Connect
- Component
- Data connectors, Workflow orchestrator, Language model +4
Converse
Hold a dialogue in text or voice, and hand over.
- Billing chat and voice bots
- Agent assist for the next reply
Does not solve: Knowing tariffs or changing accounts (3 and 8).
- Capability
- Converse
- Solution
- Customer self-service assistant
- Pattern
- RAG (retrieval-augmented generation)
- Standard
- OpenAPI, OAuth 2.0 / OpenID Connect
- Component
- Document store, Embedding model, Vector store +4
Answer from knowledge
Answer from company knowledge, citing the source.
- Tariff and policy questions
- Engineers searching manuals
Does not solve: Out-of-date or contradictory sources.
- Capability
- Answer from knowledge
- Solution
- Care knowledge assistant
- Pattern
- RAG (retrieval-augmented generation)
- Standard
- OpenAPI, OAuth 2.0 / OpenID Connect
- Component
- Document store, Embedding model, Vector store +4
Understand text & documents
Extract, classify and tag documents, emails and calls.
- Orders read into fields
- Complaints classified by intent
Does not solve: A missing target data model.
- Capability
- Understand text & documents
- Solution
- Order and invoice document processing
- Pattern
- Document intelligence
- Standard
- OpenAPI, TM Forum TMF622 Product Ordering
- Component
- Data connectors, Document parser, Language model +5
See & hear
Analyse images, video and audio.
- Checking installation photos
- Transcribing calls for review
Does not solve: Bad photos and noisy audio.
- Capability
- See & hear
- Solution
- Call summaries and wrap-up
- Pattern
- Document intelligence
- Standard
- OpenAPI, TM Forum TMF622 Product Ordering
- Component
- Data connectors, Document parser, Language model +5
Predict & detect
Forecast outcomes and flag anomalies.
- Churn risk scores
- Fraud and fault detection
Does not solve: The action: a score needs a process to use it.
- Capability
- Predict & detect
- Solution
- Proactive network fault resolution
- Pattern
- Event-driven AI workflow
- Standard
- AsyncAPI, OpenAPI +1
- Component
- Event broker, Feature pipeline, Prediction model +4
Recommend & optimise
Choose the best action, offer or allocation.
- Next best action and offer
- Field technician scheduling
Does not solve: Conflicting business rules.
- Capability
- Recommend & optimise
- Solution
- Field technician scheduling
- Pattern
- Predictive analytics pipeline
- Standard
- OpenAPI, AsyncAPI
- Component
- Data connectors, Feature pipeline, Prediction model +2
Act on systems
Carry out steps on business systems, within set permissions.
- A billing dispute resolved end to end
- A fault investigated, ticket raised
Does not solve: Missing APIs or unclear process ownership.
- Capability
- Act on systems
- Solution
- Proactive network fault resolution
- Pattern
- AI agent with human approval
- Standard
- Model Context Protocol (MCP), OpenAPI +4
- Component
- Event broker, Agent runtime, Language model +5
Capability and business outcome
A business outcome is the measurable change the business is paying for: lower cost to serve, faster resolution, fewer repeat contacts, less revenue lost to fraud or churn. It has a business owner, a baseline measured before any pilot, and a target agreed with finance. A capability is one candidate way of moving it.
| Capability | Business outcome | |
|---|---|---|
| Example | Answer tariff questions from company knowledge, citing the source | Average handling time for tariff calls down from 7 to 5 minutes; fewer repeat contacts |
| Answers | What can AI do here? | What has to improve, by how much, and for whom? |
| Owned by | The AI or platform team that builds and runs it | The business owner who funds it and changes the process to bank the benefit |
| Proven by | Evaluation: accuracy, cited sources, latency, cost per request | Measurement against the baseline, over at least a full budget cycle after go-live |
The direction matters: the outcome is chosen first, then the capabilities that could move it. One outcome can need several capabilities (cutting dispute handling time may take understanding documents, answering from knowledge and acting on billing), one capability can serve several outcomes, and some outcomes are better served by a process or data fix with no AI at all.
A capability working does not mean the outcome is reached. An assistant can answer correctly and handling time stay flat, because agents do not trust it, the process around it is unchanged, or the minutes saved are absorbed rather than taken out of cost. That gap is the most common reason an AI pilot that passes its technical tests fails to win production funding.
Generative or classic: why it matters
| Generative AI | Classic machine learning | |
|---|---|---|
| Output | Text, code or images; can differ for the same input | A score, a class or a forecast; the same input gives the same output |
| Explaining a result | Hard; show the sources it used | Easier; the features behind a score can be documented |
| Cost shape | Per request; grows with use | Mostly training and hosting; cheap per prediction |
| Maturity in telcos | Recent; operating practice still forming | Long established in churn, fraud and forecasting |
| Data it needs | Documents and knowledge, kept current | Labelled history, in volume |