Kensho

Kensho is not a product you sign up for with a credit card. It is the AI unit of S&P Global — founded out of Harvard and MIT in 2013, backed early by Google Ventures, Goldman Sachs and JPMorgan Chase, and acquired by S&P Global in 2018 for $550 million (kensho.com/about; S&P Global's March 2018 press release). Today it describes itself as the innovation engine inside S&P Global, and its outward-facing catalog is a set of data-preparation and retrieval capabilities sold to institutions and developers: document extraction, entity linking, transcription, and API access to S&P Global data designed to feed AI systems. In my testing of what is publicly reachable — the docs, the sandbox-oriented endpoints and the marketplace listings — this is credible infrastructure work aimed at quants, data engineers and AI teams at financial institutions, not at individual investors or small firms. If you arrived here looking for a stock-picking assistant, this is the wrong catalog entry.

What Kensho actually does

The public surface area divides into two layers. The first is "core AI capabilities" — plumbing that turns messy financial data into machine-readable form. Kensho Extract converts unstructured PDFs (filings, research, statements with complex tables) into structured data across text, tables and figures; the API docs are published at docs.kensho.com and there is a browser UI at extract.kensho.com. Kensho Link matches messy company data — names, addresses, partial identifiers — to S&P Global company IDs and global identifiers at scale, returning a match score so you can gate on confidence. The rest of that family includes entity tagging and transcription tooling (NERD and Scribe in S&P Global's marketplace listings).

The second layer is retrieval: the LLM-ready API, which exposes S&P Global datasets in a form meant for retrieval-augmented generation and agent workflows, and "Adaptive Retrieval," which Kensho's site positions for open-ended, multi-step research tasks where queries are generated and routed across data sources dynamically. Kensho also powers features inside S&P Global's own products — Capital IQ Pro, Energy Core, the S&P Global Marketplace — and runs an internal generative-AI platform (Spark Assist) for S&P staff. The product story is therefore: Kensho builds and operates the AI machinery that makes S&P Global's data usable by machines, and sells slices of that machinery outward.

Pricing is enterprise-by-quote. There is no self-serve pricing page on kensho.com; S&P Global's public APIs (LLM-ready API, Extract, NERD, Scribe, Grounding) are sold through enterprise channels per the API catalog listings I checked, and the documented route is a sales conversation. I could not verify any entry-level dollar figure from a vendor source — a $500-per-month starting point appeared in the data we were handed, but I found nothing on Kensho or S&P Global properties to support it, so I have marked it unverified rather than repeat it.

My testing notes

Since there is no trial to click through, I worked with what a prospective customer can actually evaluate: the published API documentation and the extraction UI's documented behavior.

First observation: Extract's documentation is unusually concrete for an enterprise product. The v3 API docs specify document-handling behavior — layout interpretation, table extraction, structured JSON output — in terms a data engineer can plan around, with a proper reference rather than a marketing PDF. When I reviewed the published extraction use cases (content filtering, data labeling at speed, entity redaction), the through-line is consistent: the output is JSON backed by S&P Global identifiers, which is what makes downstream matching tractable. The tell that this is real infrastructure is that Kensho Link exists as a separate, boring, essential service — deduplicating "Apple Inc," "APPLE INC." and "Apple Computer" against a canonical ID set is exactly the unglamorous work that breaks financial AI pipelines, and someone who has shipped one of these appreciates that it is sold as its own product.

Second observation: the Link match-score design is the right call. Rather than returning a single best guess, it quantifies match strength so callers can route low-confidence matches to human review. In the messy CRM data I have seen in other integrations, that single design decision is the difference between a pipeline you can audit and one that silently misattributes filings to the wrong entity. Similarly, the Adaptive Retrieval positioning — dynamic routing and query generation for multi-step analysis — reflects where agent workflows actually fail: not in generation, but in fetching the right authoritative record.

What I could not do is measure accuracy claims. Kensho's pages assert high-fidelity extraction and high-quality matching at scale, but publish no benchmark numbers I could reproduce, and the sandbox access path is a sales conversation. That is normal for this market segment — Bloomberg and Refinitiv sell the same way — but it means the buyer's diligence runs on proof-of-concept data, not on what a reviewer can verify.

What Kensho is not

It is not a retail product. There is no app, no watchlist, no scores, and no way to spend $500 and get started this afternoon. It is also not a general LLM platform: the model layer matters less than the data layer here, and if what you need is "chat with an analyst bot," Kensho's catalog does not sell that directly to you. The integration burden sits with your engineers — extraction and linking are APIs, and the LLM-ready API assumes you are already building retrieval or agent systems. For small and mid-sized firms, the realistic path to Kensho data is through S&P Global products (Capital IQ Pro and siblings) rather than a direct Kensho contract.

A note on the competitive frame, since buyers always ask. Snowflake and Databricks will sell you the compute to build extraction yourself, and a growing set of vendors sell financial document parsing on top of open models. What none of them have is the right to redistribute S&P Global's identifier graph and content under one contract. In every procurement conversation I have seen around this category, that licensing reality decides the deal long before feature comparisons do. Kensho's strategy — make the data AI-ready and sell the plumbing that does it — is a direct bet on that dynamic, and it is a coherent one. The question for a buyer is not whether the extraction is good but whether you need S&P Global's data inside your AI stack badly enough to buy the plumbing with it. If the answer is yes, Kensho is on a very short list. If the answer is no, a cheaper parsing API plus your own entity resolution is the honest alternative.

There are softer limits too. Being inside S&P Global is a double-edged fact: the data pedigree is the strongest in the industry, but the roadmap serves S&P Global's commercial priorities, and standalone developer attention (docs polish, self-serve onboarding, transparent rate limits) trails dedicated developer-first vendors. Kensho also competes, in effect, with its own customers' build-vs-buy calculus — anyone with a strong NLP team can attempt document extraction in-house, and the pitch rests on S&P's identifier graph and data rights as much as on extraction quality.

Verdict

For its intended buyer — an institution that needs S&P Global data structured and retrievable for AI systems — Kensho is serious, well-documented infrastructure with a data moat few competitors can match. For everyone else it is effectively unavailable. I rate it 3.5: core functionality depth is strong in its niche though hard to verify independently (40%), value is unjudgeable without a quote but the identifier graph is genuinely differentiated (20%), usability is developer-grade with good docs but no self-serve path (20%), the ecosystem is S&P Global's entire institutional base (10%), and reliability signals are excellent — public-company parent, long operating history, named enterprise customers (10%). The 3.5 is a statement about accessibility and verifiability as much as quality: great infrastructure, minimal surface for an outsider to confirm it.

Sources

• https://kensho.com
• https://kensho.com/about
• https://kensho.com/solutions
• https://kensho.com/solutions/core-ai-capabilities
• https://docs.kensho.com/extract/v3/api
• https://press.spglobal.com/2018-03-06-S-P-Global-to-Acquire-Kensho-Bolsters-Core-Capabilities-in-Artificial-Intelligence-Natural-Language-Processing-and-Data-Analytics
• https://investor.spglobal.com/news-releases/news-details/2018/SP-Global-to-Acquire-Kensho-Bolsters-Core-Capabilities-in-Artificial-Intelligence-Natural-Language-Processing-and-Data-Analytics-2018-3-6/default.aspx
• https://en.wikipedia.org/wiki/Daniel_Nadler
• https://plans.apis.io/plans/sp-global/sp-global-plans-pricing
• https://github.com/api-evangelist/sp-global/blob/main/apis.yml