Labs · Feb 20, 2025
AI21 Labs argues enterprise AI needs systems, not just models
Markdown · JSON · Original source
Labs’ page was published on February 20, 2025. It argues that enterprise AI must move beyond models and experiments toward systems designed to fit business workflows. The company says that even strong performance on tests does not ensure dependable results in complex, high-stakes work: errors can compound across multiple steps, and existing methods such as , and do not eliminate the underlying problem. AI21 Labs’ proposed direction is orchestration: a system that can retrieve current information, use specialized tools, and route complex requests for human oversight. Its example is a customer query that draws on product updates and CRM data before reaching the appropriate team. The company urges businesses to judge systems by whether outputs are controllable and reliable, whether the systems integrate with existing tools, and whether decisions can be traced. It concludes that enterprises are ready to move from experimentation to execution, and says AI21 Labs focuses on systems because businesses need them for that transition. The page does not give evidence for its estimate of how many projects reach production, or report measured outcomes from the proposed approach.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
Labs · Article
Source details
Original source date: Feb 20, 2025 (2025-02-20T12:56:13Z).
Date basis: Date from the publisher’s own records, through its API.
Added to Escape Meridian: 2026-11-15T06:18:30.045Z. First observed: 2026-10-30T07:16:21.788Z. Retrieved: 2026-10-30T07:17:49.717Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
Text as retrieved on October 30, 2026; originally published February 20, 2025. The page gives no underlying details for the Dataiku and Deloitte estimate, such as the projects counted or how production was defined. The page does not report measured results from the described orchestration approach; its customer-query scenario is an example, not a reported evaluation. The page is published by Labs and presents the company’s argument for enterprise AI systems.
Open question
What evidence would show that an orchestrated enterprise AI system produces more dependable results in real workflows than the methods Labs describes?
Source-stated
AI21 Labs says Dataiku and Deloitte report that only 20–30% of GenAI projects make it to production.
Dataiku and Deloitte report that only 20-30% of GenAI projects ever make it to production.
Source-stated
AI21 Labs argues that LLMs alone cannot guarantee desired outputs, and that errors can compound in complex tasks.
These errors compound when an LLM handles complex, multi-step tasks, and the hallucinations may multiply through layers of decision-making
Source-stated
Ori Goshen, AI21 Labs’ Co-CEO, says leading models struggle with nontrivial workflows and that real-world enterprise use cases need AI systems, not just LLMs.
There are no shortcuts. AI systems—not just LLMs—are needed to tackle real-world enterprise use cases.
Source-stated
AI21 Labs describes an enterprise AI system as using orchestration and planning to retrieve data, use specialized tools, and involve human oversight.
Dynamic orchestration and planning are needed to decide when to retrieve factual data, when to offload calculations and tasks to specialized tools, and when to rely on human oversight.
Source-stated
AI21 Labs concludes that enterprises are ready to move from experimentation to integrated systems, and says its focus is systems because businesses need them to move from experimentation to execution.
We focus on systems, not just models — because that’s what businesses actually need to move from experimentation to execution.
Source and contributor context
AI21 Labs · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
Retrieval-augmented generation
A setup in which a model first looks up relevant documents and then answers using them.
Prompt engineering
Writing and refining the instructions given to a model to get more useful results.
AI21
AI21 publishes technical research and product announcements about its language models and AI systems.
Editorial provenance
Version 1. Publication ID: 5b342d2d-8f98-426e-3cb5-fd7cc981133f.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-15T06:13:56.103Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: 939857afc189074b5dc8b9e5349740746f205f3f1983911f72ed2c8d22f87c50. Source text hash: a700502a4cc19e8356dcd33132e7df89923c5a368df5f4d079c8b9d8ed859b7b.
Revision and source-check history
Version 1 · 2026-11-15T06:18:30.045Z
AI21 Labs argues enterprise AI needs systems, not just models
Labs’ page was published on February 20, 2025. It argues that enterprise AI must move beyond models and experiments toward systems designed to fit business workflows. The company says that even strong performance on tests does not ensure dependable results in complex, high-stakes work: errors can compound across multiple steps, and existing methods such as , and do not eliminate the underlying problem. AI21 Labs’ proposed direction is orchestration: a system that can retrieve current information, use specialized tools, and route complex requests for human oversight. Its example is a customer query that draws on product updates and CRM data before reaching the appropriate team. The company urges businesses to judge systems by whether outputs are controllable and reliable, whether the systems integrate with existing tools, and whether decisions can be traced. It concludes that enterprises are ready to move from experimentation to execution, and says AI21 Labs focuses on systems because businesses need them for that transition. The page does not give evidence for its estimate of how many projects reach production, or report measured outcomes from the proposed approach.
Original source
Labs · Article
Source details
Original source date: Feb 20, 2025 (2025-02-20T12:56:13Z).
Date basis: Date from the publisher’s own records, through its API.
Added to Escape Meridian: 2026-11-15T06:18:30.045Z. First observed: 2026-10-30T07:16:21.788Z. Retrieved: 2026-10-30T07:17:49.717Z.
Limits
Text as retrieved on October 30, 2026; originally published February 20, 2025. The page gives no underlying details for the Dataiku and Deloitte estimate, such as the projects counted or how production was defined. The page does not report measured results from the described orchestration approach; its customer-query scenario is an example, not a reported evaluation. The page is published by Labs and presents the company’s argument for enterprise AI systems.
Open question
What evidence would show that an orchestrated enterprise AI system produces more dependable results in real workflows than the methods Labs describes?
Source-stated
AI21 Labs says Dataiku and Deloitte report that only 20–30% of GenAI projects make it to production.
Dataiku and Deloitte report that only 20-30% of GenAI projects ever make it to production.
Source-stated
AI21 Labs argues that LLMs alone cannot guarantee desired outputs, and that errors can compound in complex tasks.
These errors compound when an LLM handles complex, multi-step tasks, and the hallucinations may multiply through layers of decision-making
Source-stated
Ori Goshen, AI21 Labs’ Co-CEO, says leading models struggle with nontrivial workflows and that real-world enterprise use cases need AI systems, not just LLMs.
There are no shortcuts. AI systems—not just LLMs—are needed to tackle real-world enterprise use cases.
Source-stated
AI21 Labs describes an enterprise AI system as using orchestration and planning to retrieve data, use specialized tools, and involve human oversight.
Dynamic orchestration and planning are needed to decide when to retrieve factual data, when to offload calculations and tasks to specialized tools, and when to rely on human oversight.
Source-stated
AI21 Labs concludes that enterprises are ready to move from experimentation to integrated systems, and says its focus is systems because businesses need them to move from experimentation to execution.
We focus on systems, not just models — because that’s what businesses actually need to move from experimentation to execution.
Source and contributor context
AI21 Labs · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
Retrieval-augmented generation
A setup in which a model first looks up relevant documents and then answers using them.
Prompt engineering
Writing and refining the instructions given to a model to get more useful results.
AI21
AI21 publishes technical research and product announcements about its language models and AI systems.
Editorial provenance
Version 1. Publication ID: 5b342d2d-8f98-426e-3cb5-fd7cc981133f.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-15T06:13:56.103Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: 939857afc189074b5dc8b9e5349740746f205f3f1983911f72ed2c8d22f87c50. Source text hash: a700502a4cc19e8356dcd33132e7df89923c5a368df5f4d079c8b9d8ed859b7b.
passed · 2026-11-15T06:13:56.103Z · anthropic/claude-opus-5.5
Publication: 5b342d2d-8f98-426e-3cb5-fd7cc981133f
attribution — pass: This is a single blog post by AI21 Labs, not a listing. The reading attributes its arguments to the company and the quote to Co-CEO Ori Goshen. It also frames the Dataiku and Deloitte figure as something AI21 Labs reports. Its limitations note that the page is AI21 Labs' own argument.
As Ori Goshen, Co-CEO of AI21 Labs, puts it:
summary — pass: The reading leads with the company's thesis that enterprises need systems, not just models, and explains why. It then describes orchestration in plain terms and gives the customer-query example. Fine-tuning, RAG and prompt engineering are defined beside the reading. The four evaluation criteria and the company's stake are stated without adopting its promotional framing. The question follows from the source's lack of measured outcomes. CRM is unexplained, but it is not central.
Dynamic orchestration and planning are needed to decide when to retrieve factual data, when to offload calculations and tasks to specialized tools, and when to rely on human oversight.
date — pass: The reading gives February 20, 2025, which is the UTC day of originalPublished. The page was edited in March 2026. However, the cited passages fit early 2025: DeepSeek R1 is compared to o1, and the post says "In 2025." Nothing is presented with hindsight. The capture date is disclosed. This is a publisher page, not an archive copy.
Neither DeepSeek’s R1 nor other leading models address the core challenges of controlling output precision and optimizing resource efficiency.
support — pass: Each claim matches its excerpt and stays attributed. The statement that tested methods do not eliminate the core issue is in the source. The routing example and the human-oversight point are slightly merged, but both appear in the source. The limitations correctly note there are no details behind the 20–30% figure and no measured results. The cited text reads as part of the original February 2025 piece, despite the later modification date.
Yet even these methods do not eliminate the core issue: LLMs on their own cannot guarantee a desired output
media — pass: No media was selected, and no verified media candidates were supplied, so this is appropriate.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
AI21 publishes technical research and product announcements about its language models and AI systems.
Explore AI21AI21 publishes technical research and product announcements about its language models and AI systems.
Explore AI21Further training of an existing model to adapt it to a particular task or use.
Definition referenceA setup in which a model first looks up relevant documents and then answers using them.
Definition referenceWriting and refining the instructions given to a model to get more useful results.
Definition referenceAI21 publishes technical research and product announcements about its language models and AI systems.
Explore AI21AI21 publishes technical research and product announcements about its language models and AI systems.
Explore AI21AI21 publishes technical research and product announcements about its language models and AI systems.
Explore AI21AI21 publishes technical research and product announcements about its language models and AI systems.
Explore AI21Further training of an existing model to adapt it to a particular task or use.
Definition referenceA setup in which a model first looks up relevant documents and then answers using them.
Definition referenceWriting and refining the instructions given to a model to get more useful results.
Definition referenceAI21 publishes technical research and product announcements about its language models and AI systems.
Explore AI21AI21 publishes technical research and product announcements about its language models and AI systems.
Explore AI21AI21 publishes technical research and product announcements about its language models and AI systems.
Explore AI21