Not yet judged↑
Frontier language models
General-purpose language models trained at scale.
- Maturity
- Not yet judged
- Attractiveness
- 44
- Evidence
- 61 (3 in last 4 wks) · Higher
Research · AI Radar
Which AI technologies are moving, and which are ready.
Five layers, from models to infrastructure. Momentum and attractiveness are computed from the evidence; maturity is a written judgement.
Models
Not yet judged↑
Multimodal & video models
Models that read, hear or produce images, video and audio as well as text.
- Maturity
- Not yet judged
- Attractiveness
- 15
- Evidence
- 1 (1 in last 4 wks) · Low
Not yet judged·
Reasoning models
Models that spend extra compute thinking through a problem before answering.
- Maturity
- Not yet judged
- Attractiveness
- –
- Evidence
- None yet
Not yet judged·
Open-weight & small models
Models whose weights can be run and tuned by the buyer, including small and on-device ones.
- Maturity
- Not yet judged
- Attractiveness
- –
- Evidence
- None yet
Agents & orchestration
Not yet judged↑
Agent frameworks & orchestration
Software that lets models plan, call tools and hand work between steps.
- Maturity
- Not yet judged
- Attractiveness
- 70
- Evidence
- 21 (8 in last 4 wks) · Higher
- Capital
- $128M
Not yet judged↑
Multi-agent systems
Several agents with different roles working on one task.
- Maturity
- Not yet judged
- Attractiveness
- 64
- Evidence
- 2 (2 in last 4 wks) · Low
- Capital
- $333M
Use now↑
Voice agents
Agents that hold spoken conversations, such as in contact centres.
- Maturity
- Scaling
- Attractiveness
- 62
- Evidence
- 5 (1 in last 4 wks) · Higher
- Capital
- $350M
Named deployments at volume: Viettel callbot ~33M calls (2024) and DBS digibot resolving 9 in 10 queries (H1 2026), both company-reported; no independent savings audit. Proposed by research run 2026-10-14; a person confirms.
Use now↑
Coding agents
Agents that write, test and ship code with limited supervision.
- Maturity
- Scaling
- Attractiveness
- 27
- Evidence
- 8 (4 in last 4 wks) · Higher
- Capital
- $1.0M
Measured gains reported in production (Sony Bank/Fujitsu: development period -30%, man-hours -40%; Techcombank: agent build time months to under a week; both company-reported) and strong lab benchmark scores, but an RCT found experienced developers 19% slower and benchmark contamination is documented. Proposed by research run 2026-10-14; a person confirms.
Pilot→
Computer-use & browser agents
Agents that operate screens, browsers and desktop software like a person would.
- Maturity
- Early commercial
- Attractiveness
- 10
- Evidence
- 1 (0 in last 4 wks) · Low
OSWorld-Verified leaders now exceed the ~72% human baseline (lab and leaderboard figures), but independent work finds agents complete very little real paid work end to end (Remote Labor Index: best agent automated 2.5% of projects). Proposed by research run 2026-10-14; a person confirms.
Not yet judged·
Agent protocols & interoperability
Shared standards for how agents call tools and talk to each other.
- Maturity
- Not yet judged
- Attractiveness
- –
- Evidence
- None yet
Data & knowledge
Not yet judged→
Retrieval & enterprise search
Giving models access to a company's own documents and data.
- Maturity
- Not yet judged
- Attractiveness
- 11
- Evidence
- 2 (0 in last 4 wks) · Low
Not yet judged·
Vector & AI-native databases
Data stores built to find items by meaning rather than keyword.
- Maturity
- Not yet judged
- Attractiveness
- –
- Evidence
- None yet
Not yet judged·
Agent memory & context
How agents keep and reuse what they have learned across tasks.
- Maturity
- Not yet judged
- Attractiveness
- –
- Evidence
- None yet
Not yet judged·
Data labelling & synthetic data
Producing training and evaluation data, by people or by models.
- Maturity
- Not yet judged
- Attractiveness
- –
- Evidence
- None yet
Applications
Not yet judged↑
Vertical copilots
Assistants built for one profession or industry's daily work.
- Maturity
- Not yet judged
- Attractiveness
- 71
- Evidence
- 17 (1 in last 4 wks) · Higher
- Capital
- $917M
Not yet judged↑
AI-native software
Products designed around AI doing the work, rather than humans using a tool.
- Maturity
- Not yet judged
- Attractiveness
- 49
- Evidence
- 3 (1 in last 4 wks) · Medium
- Capital
- $555M
Not yet judged·
Document & workflow automation
Reading, extracting from and routing documents and forms.
- Maturity
- Not yet judged
- Attractiveness
- –
- Evidence
- None yet
Not yet judged·
AI search & answer engines
Search that returns synthesised answers rather than links.
- Maturity
- Not yet judged
- Attractiveness
- –
- Evidence
- None yet
Infrastructure & trust
Not yet judged↑
AI security & governance
Protecting AI systems and proving they follow policy.
- Maturity
- Not yet judged
- Attractiveness
- 66
- Evidence
- 4 (3 in last 4 wks) · Medium
- Capital
- $281M
Not yet judged↑
Evals, observability & guardrails
Testing, monitoring and constraining AI systems in production.
- Maturity
- Not yet judged
- Attractiveness
- 49
- Evidence
- 2 (1 in last 4 wks) · Low
- Capital
- $34.0M
Not yet judged↑
AI chips & accelerators
Processors designed for AI workloads.
- Maturity
- Not yet judged
- Attractiveness
- 25
- Evidence
- 1 (1 in last 4 wks) · Low
- Capital
- $5.0M
Not yet judged·
Inference & serving
Running trained models quickly and cheaply for real users.
- Maturity
- Not yet judged
- Attractiveness
- –
- Evidence
- None yet
Not yet judged·
GPU cloud & compute
Rented or built compute capacity for training and running models.
- Maturity
- Not yet judged
- Attractiveness
- –
- Evidence
- None yet
Calls: Use now (maturity 4–5), Pilot (3), Watch (1–2). A technology shows "Not yet judged" until someone writes its maturity and a reason. See Method.