# AI21 Labs argues enterprise GenAI is harder to scale than expected

[Canonical reading](https://escapemeridian.com/readings/ai21-labs-argues-enterprise-genai-is-harder-to-scale-than-970ffbf7)

AI21 Labs published this account on January 21, 2025, arguing that enterprise GenAI has proved harder to scale than early expectations suggested. It says implementation costs and timelines exceeded projections, while data preparation and team training also demanded more effort than anticipated. AI21 Labs presents a combination of Retrieval-Augmented Generation (RAG), systematic workflows, prompt engineering and long-context processing as approaches that can improve output, while noting their costs and limitations. It describes Jamba, its hybrid model combining Mamba layers and Transformer attention, as designed to balance efficiency and accuracy. AI21 Labs points to benchmark results and a financial term-sheet workflow as evidence for its long-context approach, and says it deployed a similar solution for an international bank. The account does not give the bank’s measured results or the comparison used to assess the workflow. It also argues that current agentic workflows remain limited by autonomy, error recovery, integration and cost. AI21 Labs’ conclusion is that GenAI can deliver value with thoughtful implementation, while important challenges remain; its vision of more capable agents is a future prospect, not a demonstrated outcome. What measured results and comparison, if any, support the bank term-sheet example?

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

AI21 Labs · Article

Original source date: Jan 21, 2025 (2025-01-21T13:13:17Z).

Date basis: Date from the publisher’s own records, through its API.

Added to Escape Meridian: 2026-11-16T00:43:25.476Z. First observed: 2026-10-25T20:02:31.784Z. Retrieved: 2026-10-25T20:05:39.834Z.

- [Is GenAI Living Up to Its Expectations? Insights from 3 Years in the Trenches](https://www.ai21.com/blog/is-genai-living-up-to-its-expectations-insights-from-3-years-in-the-trenches/)

## Limits

Text as retrieved on October 26, 2026; originally published January 21, 2025. AI21 Labs does not provide measured results or a comparison for the international bank term-sheet deployment described in the text.


## Open question

What measured results and comparison, if any, support the bank term-sheet example?


## Source-stated

AI21 Labs says enterprise implementations had higher infrastructure costs and longer timelines than projected.

> infrastructure costs running 40-60% higher than projected to integration timelines stretching 2-3x longer than planned.

- [Supporting source](https://www.ai21.com/blog/is-genai-living-up-to-its-expectations-insights-from-3-years-in-the-trenches/)

## Source-stated

AI21 Labs describes Jamba as combining Mamba state-space layers and Transformer attention, and says it can process up to 256k tokens of context.

> we’ve developed Jamba, a hybrid model combining the efficiency of Mamba’s state-space layers with Transformer attention’s precision. This architecture enables processing up to 256k tokens of context

- [Supporting source](https://www.ai21.com/blog/is-genai-living-up-to-its-expectations-insights-from-3-years-in-the-trenches/)

## Source-stated

AI21 Labs says a similar term-sheet generation solution was deployed for an international bank, but does not provide measured results or a comparison for that deployment.

> We’ve deployed a similar solution for a leading international bank, automating term sheet generation while maintaining strict accuracy and compliance standards.

- [Supporting source](https://www.ai21.com/blog/is-genai-living-up-to-its-expectations-insights-from-3-years-in-the-trenches/)

## Source-stated

AI21 Labs says current agentic systems face limitations in autonomy, error recovery, integration and cost.

> True autonomy remains a significant challenge
> Error recovery capabilities are still primitive
> Integration across tools and systems needs improvement
> Cost and resource requirements can be prohibitive

- [Supporting source](https://www.ai21.com/blog/is-genai-living-up-to-its-expectations-insights-from-3-years-in-the-trenches/)

## 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.


## Retrieval-augmented generation

A setup in which a model first looks up relevant documents and then answers using them.

- [Definition reference](https://developers.google.com/machine-learning/glossary#retrieval-augmented-generation-rag)

## Prompt engineering

Writing and refining the instructions given to a model to get more useful results.

- [Definition reference](https://developers.google.com/machine-learning/glossary#prompt-engineering)

## Benchmark

A defined set of tests used to compare systems. Results depend on the tasks and measures selected.

- [Definition reference](https://arxiv.org/abs/2211.09110)

## Agentic AI

AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.

- [Definition reference](https://developers.google.com/machine-learning/glossary#agentic)

## AI21

AI21 publishes technical research and product announcements about its language models and AI systems.

- [Explore this context](/entities/ai21)
- [Reference](https://www.ai21.com/blog/)

## Editorial provenance

Version 1. Publication ID: 8c8483bc-e48c-a728-1bc2-65dbe64aaecf.

Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T00:40:15.923Z. 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: 963361a0ee4f0b50834d9960403971c762479b764b27ce10bef3e8d2c471353f. Source text hash: e8a7e7a9e4171c56cbd8ce767021d64fd03b11dfaeae6fcd05444866222fa86f.


## Revision and source-check history

### Version 1 — 2026-11-16T00:43:25.476Z

AI21 Labs published this account on January 21, 2025, arguing that enterprise GenAI has proved harder to scale than early expectations suggested. It says implementation costs and timelines exceeded projections, while data preparation and team training also demanded more effort than anticipated. AI21 Labs presents a combination of Retrieval-Augmented Generation (RAG), systematic workflows, prompt engineering and long-context processing as approaches that can improve output, while noting their costs and limitations. It describes Jamba, its hybrid model combining Mamba layers and Transformer attention, as designed to balance efficiency and accuracy. AI21 Labs points to benchmark results and a financial term-sheet workflow as evidence for its long-context approach, and says it deployed a similar solution for an international bank. The account does not give the bank’s measured results or the comparison used to assess the workflow. It also argues that current agentic workflows remain limited by autonomy, error recovery, integration and cost. AI21 Labs’ conclusion is that GenAI can deliver value with thoughtful implementation, while important challenges remain; its vision of more capable agents is a future prospect, not a demonstrated outcome. What measured results and comparison, if any, support the bank term-sheet example?

## Original source

AI21 Labs · Article

Original source date: Jan 21, 2025 (2025-01-21T13:13:17Z).

Date basis: Date from the publisher’s own records, through its API.

Added to Escape Meridian: 2026-11-16T00:43:25.476Z. First observed: 2026-10-25T20:02:31.784Z. Retrieved: 2026-10-25T20:05:39.834Z.

- [Is GenAI Living Up to Its Expectations? Insights from 3 Years in the Trenches](https://www.ai21.com/blog/is-genai-living-up-to-its-expectations-insights-from-3-years-in-the-trenches/)

## Limits

Text as retrieved on October 26, 2026; originally published January 21, 2025. AI21 Labs does not provide measured results or a comparison for the international bank term-sheet deployment described in the text.


## Open question

What measured results and comparison, if any, support the bank term-sheet example?


## Source-stated

AI21 Labs says enterprise implementations had higher infrastructure costs and longer timelines than projected.

> infrastructure costs running 40-60% higher than projected to integration timelines stretching 2-3x longer than planned.

- [Supporting source](https://www.ai21.com/blog/is-genai-living-up-to-its-expectations-insights-from-3-years-in-the-trenches/)

## Source-stated

AI21 Labs describes Jamba as combining Mamba state-space layers and Transformer attention, and says it can process up to 256k tokens of context.

> we’ve developed Jamba, a hybrid model combining the efficiency of Mamba’s state-space layers with Transformer attention’s precision. This architecture enables processing up to 256k tokens of context

- [Supporting source](https://www.ai21.com/blog/is-genai-living-up-to-its-expectations-insights-from-3-years-in-the-trenches/)

## Source-stated

AI21 Labs says a similar term-sheet generation solution was deployed for an international bank, but does not provide measured results or a comparison for that deployment.

> We’ve deployed a similar solution for a leading international bank, automating term sheet generation while maintaining strict accuracy and compliance standards.

- [Supporting source](https://www.ai21.com/blog/is-genai-living-up-to-its-expectations-insights-from-3-years-in-the-trenches/)

## Source-stated

AI21 Labs says current agentic systems face limitations in autonomy, error recovery, integration and cost.

> True autonomy remains a significant challenge
> Error recovery capabilities are still primitive
> Integration across tools and systems needs improvement
> Cost and resource requirements can be prohibitive

- [Supporting source](https://www.ai21.com/blog/is-genai-living-up-to-its-expectations-insights-from-3-years-in-the-trenches/)

## 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.


## Retrieval-augmented generation

A setup in which a model first looks up relevant documents and then answers using them.

- [Definition reference](https://developers.google.com/machine-learning/glossary#retrieval-augmented-generation-rag)

## Prompt engineering

Writing and refining the instructions given to a model to get more useful results.

- [Definition reference](https://developers.google.com/machine-learning/glossary#prompt-engineering)

## Benchmark

A defined set of tests used to compare systems. Results depend on the tasks and measures selected.

- [Definition reference](https://arxiv.org/abs/2211.09110)

## Agentic AI

AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.

- [Definition reference](https://developers.google.com/machine-learning/glossary#agentic)

## AI21

AI21 publishes technical research and product announcements about its language models and AI systems.

- [Explore this context](/entities/ai21)
- [Reference](https://www.ai21.com/blog/)

## Editorial provenance

Version 1. Publication ID: 8c8483bc-e48c-a728-1bc2-65dbe64aaecf.

Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T00:40:15.923Z. 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: 963361a0ee4f0b50834d9960403971c762479b764b27ce10bef3e8d2c471353f. Source text hash: e8a7e7a9e4171c56cbd8ce767021d64fd03b11dfaeae6fcd05444866222fa86f.


### uncertain — 2026-11-16T00:40:15.923Z — anthropic/claude-opus-5.5

Publication: 8c8483bc-e48c-a728-1bc2-65dbe64aaecf

attribution — pass: This is a single AI21 Labs blog post with its own argument, and the reading consistently attributes its claims, including Jamba's performance, benchmark leadership and the bank deployment, to AI21 Labs.

> At AI21 Labs, we’ve developed Jamba

summary — uncertain: The scaling-difficulty argument is clear. However, the Jamba passage relies on terms that are not explained in the reading or its definitions: Mamba layers, Transformer attention, long-context processing, and tokens (256k, in the claims). A general reader cannot tell why combining these matters. Smallest repair: add Transformer, Mamba/state-space model, context window and token to the term library. The ending is somewhat flattened: the source closes with 'GenAI isn't just living up to expectations—it's redefining them,' which is more forceful than 'can deliver value.'

> Transformers are memory-hungry and painfully slow with large input sizes.

date — uncertain: The January 21, 2025 date matches the UTC day of originalPublished. However, the limitation states the text was retrieved on October 26, 2026, while retrievedAt is 2026-10-25T20:05Z UTC, so that date is wrong. The page was also edited after publication (dateModified 2026-03-25). The text shows signs of later additions, including a duplicated, expanded 'Road Ahead' paragraph and appended limitation lists. Cited cost and agent-limitation passages may come from that edit.

support — uncertain: The claims match the source text, and the bank-results gap is accurately noted. However, editedAfterPublication is true, and several cited passages look like later insertions. These include the intro's 40-60% cost and 2-3x timeline figures, which are repeated in a 'What We Got Wrong' list, and the 'we must be realistic about current limitations' agent list, which sits beside a duplicated, revised conclusion. These passages are not plainly part of the January 2025 original, so claims 1 and 4 cannot be confirmed as original content.

> Now while the future of GenAI lies in agentic workflows, we must be realistic about current limitations:

media — pass: No media was selected, and no verified media candidates were supplied, so there is no media identity or relevance issue.

## Recorded references

Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter.

No public references recorded here yet.

[Reference data (JSON)](/readings/970ffbf71f382548e4b250ae6602decdcdd332737044cdd6b36a00d117a42437/references.json)

## Corrections and challenges

[Send a correction or an opposing source](https://escapemeridian.com/corrections/)

[JSON record](https://escapemeridian.com/readings/ai21-labs-argues-enterprise-genai-is-harder-to-scale-than-970ffbf7.json)
