Almost every AI system shipping into production in 2026 is applied AI — a bounded model serving a bounded task with a bounded data distribution. General AI, often called AGI, is something else entirely: a research and policy conversation about whether any single system could one day handle the open-ended range of tasks a person does. The distinction is not academic. It sets what an engineering team is allowed to claim, what evaluation it needs to build, and what failure mode it has to plan for.

We see this confusion show up early in scoping conversations. A stakeholder describes the system they want as “smart” or “able to figure things out”, and the team has to translate that back into a defined task, a defined input distribution, and a defined acceptance criterion. The applied-versus-general framing is the cleanest way to surface that translation before it costs anyone time.

What is applied AI in practice?

Applied AI — sometimes called narrow AI — is a system trained to solve one specific task with a measurable success metric. It does that task well, and outside the task it does nothing useful. The bounding is the feature, not the limitation.

The list of working applied AI systems in 2026 is long and concrete:

• Recommendation ranking on Netflix, Spotify, TikTok, YouTube
• Search and ad ranking at Google and Meta
• Driver-assistance perception stacks at Tesla, Mobileye, Waymo
• Medical-image triage for mammography and chest X-ray
• Code assistants like GitHub Copilot and Cursor
• Voice transcription at Whisper-class quality
• Document AI for invoices, IDs, and forms
• Industrial defect detection on production lines
• Conversational assistants scoped to a product domain

Each of these has a defined input distribution (resumes, mammograms, Python source files, customer support transcripts), a defined output schema, and a defined evaluation. When the input drifts outside that distribution, the system is expected to degrade gracefully or hand off to a human. That handoff is the engineering work that separates a deployed system from a demo.

The model architecture underneath varies — transformers for language and increasingly for vision, convolutional networks for some imaging tasks, gradient-boosted trees for tabular ranking, diffusion models for image generation. The architecture is a choice; the applied framing is what makes the system shippable.

How does general AI differ from what we actually build?

General AI describes a hypothetical system that solves a wide, open-ended range of tasks at human-level competence and transfers between domains the way people do. Most operational definitions published a decade ago — by DeepMind, OpenAI, and academic labs — required a single system to perform across language, perception, planning, and embodied action without task-specific retraining.

No current system meets those definitions. Frontier large language and multi-modal models have genuinely widened the range of tasks one system can attempt, which has made the boundary fuzzier than it was five years ago. Some researchers call current frontier systems “proto-AGI”; others argue we are missing core capabilities such as robust reasoning, embodied learning, and reliable long-horizon planning. The disagreement is honest and unresolved.

What is uncontested is that no production team in 2026 ships a system on AGI claims. The systems shipping are applied AI built on increasingly capable foundation models — which is a different statement from “we have built general intelligence.”

Why the distinction matters for engineering scope

The applied-versus-general distinction is the single cleanest gate for catching underspecified projects. A team that frames its work as applied AI inherits a checklist:

Scope element	Applied AI requirement	Failure if skipped
Input distribution	Named, sampled, version-controlled	Silent drift; evaluation goes stale
Success metric	One primary, ≤2 guardrails	Stakeholders argue about quality post-launch
Out-of-distribution behaviour	Detection + fallback path	System answers confidently on inputs it should refuse
Evaluation set	Held-out, refreshed on a cadence	Overfitting to last quarter’s data
Failure budget	Quantified error rate the business accepts	Indefinite “needs more training” loop

A team that quietly slides toward general-AI framing — “the model should just handle whatever the user throws at it” — skips the entire left column. The result is the recurring pattern we encounter on audits: a model that demos well, ships, and then fails in production on inputs nobody scoped. The team responds by training a larger model, which extends the demo radius but does not close the underlying gap. The gap is a scoping artifact, not a model-capacity artifact.

Foundation models complicate this only at the margins. A general-purpose LLM used inside an applied system is still applied — the bounded task is “answer customer questions about this specific product catalogue,” not “be intelligent.” The evaluation is still bounded, the fallback is still required, and the failure budget still has to be named.

Where this sits in the broader taxonomy

Applied versus general is one axis. The orthogonal axis is the family of techniques in play — symbolic reasoning, traditional supervised learning, deep learning, generative models, retrieval-augmented systems. A working taxonomy needs both axes, because an applied system can be built from any family and a general-AI claim can be made (incorrectly) from any of them too. The companion piece, Symbolic vs Generative vs Traditional ML: A Working Taxonomy for Practitioners, walks the family axis.

For programme-level context across our engagements, our Generative & Agentic AI R&D practice frames scoping conversations around applied delivery first, with foundation-model capability treated as a building block rather than a destination.

The practical takeaway is unromantic. Every successful 2026 deployment we have worked on is openly framed as applied AI with explicit scope. The teams that struggle are the ones still negotiating, midway through implementation, whether they are building a tool or building intelligence. That negotiation belongs at the start, not at the demo.

FAQ

What is the difference between applied AI and general AI?

Applied (or “narrow”) AI solves one specific task — spam classification, route planning, image segmentation, transcription. General AI (often “AGI”) would solve a wide, open-ended range of tasks at human-level competence and transfer across domains the way people do. Everything in production in 2026 is applied AI; AGI remains a research and policy debate rather than a deployable category.

Are we close to general AI in 2026?

Frontier large language and multi-modal models have widened the range of tasks any one system can do, which has made the boundary fuzzier than it was five years ago. Researchers disagree publicly: some call current frontier systems “proto-AGI,” others argue we are missing core capabilities (robust reasoning, embodied learning, reliable long-horizon planning). What is uncontested: no current system meets the operational definitions most labs published a decade ago.

What are concrete examples of applied AI today?

Recommendation systems (Netflix, Spotify, TikTok), search and ad ranking (Google, Meta), driver-assistance perception (Tesla, Mobileye), medical-image triage (mammography, chest X-ray), code assistants (Copilot, Cursor), voice transcription (Whisper-class), document AI (invoice, ID, forms), industrial defect detection, and conversational assistants. Each one solves a bounded problem with a bounded data distribution.

Why does the applied vs general distinction matter for engineering teams?

Because it sets realistic acceptance criteria. Applied AI ships with a defined data distribution, a measurable success metric, and a fallback for out-of-distribution inputs. Pretending you are building “general” AI usually means underspecified evaluation, missing guardrails, and shipped systems that fail in production on inputs nobody scoped. Almost every successful 2026 deployment is openly framed as applied AI with explicit scope.