AI news videos

Salesforce Says Its New AI Model Makes Three Times Fewer Errors

Salesforce and NVIDIA announced Koa, which Salesforce calls its first CRM reasoning model for Agentforce. It is not generally available. Koa is in select customer pilots now, and Salesforce says general availability is expected in winter 2026 in U.S. regions.

Watch on YouTube

Transcript

Sources

Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:

• PUNCTUATION NOTE. The source press release uses em dashes throughout its prose, including in the dateline, in both executive quotations and in the mission specific intelligence section. This script file forbids dash characters of any kind, so every quotation reproduced below renders the source's em dashes as commas. The source also uses curly typographic apostrophes and the HTML entities ’ and  , which are rendered below as plain straight apostrophes and plain spaces. No word inside any quotation was added, removed or changed. SENTENCE BY SENTENCE AUDIT. Hook, two sentences: "Salesforce and NVIDIA announced Koa, a reasoning model for Agentforce agents. It is not generally available yet." The first sentence is supported by the article's lead, "SAN FRANCISCO, September 15, 2026 , Salesforce and NVIDIA today announced Koa, Salesforce's first CRM reasoning model for Agentforce, built on NVIDIA Nemotron." The hook is cut into exactly two sentences because the hook card renders exactly two beats. The word "first" was dropped from "Salesforce's first CRM reasoning model" for length, which understates rather than inflates the announcement. The source's marketing framing "purpose-built" from the following sentence is deliberately not narrated anywhere in this Short, and neither are "unprecedented", "trusted" or "agentic enterprise". No attribution verb is attached to the first hook sentence because the fact being stated is only that the two companies made an announcement, which is a plain fact about the world and not a capability claim; the capability claims that follow are all attributed out loud. The second hook sentence, "It is not generally available yet", is supported by the article's own Availability section, "Koa: Available to select pilot customers now in Agentforce; general availability expected winter 2026 in U.S. regions." That status line was deliberately promoted into the hook rather than left for a later scene, because the largest risk in a sixty second Short about a vendor launch is that a viewer hears an announced product as a shipped, buyable one. Scene one, single sentence: "Salesforce says it trained Koa on NVIDIA's Nemotron open model using synthetic data modeled on twenty seven years of Salesforce deployments." Supported by "Koa was built by post-training NVIDIA Nemotron 3 Super with a proprietary synthetic dataset modeled on enterprise knowledge from nearly three decades of CRM deployments", by the article's first summary bullet, "Koa is trained with 27 years of Salesforce CRM intelligence to enable agents to reason through the complex, multistep tasks required for enterprise work", and by "NVIDIA Nemotron open models gave Salesforce the control to build CRM intelligence into the model itself." Three compressions were made. First, the base model is narrated as "NVIDIA's Nemotron open model" and not as the source's "NVIDIA Nemotron 3 Super", because a glued brand and digit token of that shape is garbled by the speech engine; the word "open" is not ours, it is the source's own description of Nemotron in the line quoted immediately above. Second, the source's verb "post-training" is narrated as plain "trained", because the hyphenated compound is stuttered by the speech engine; this is a slight loss of technical precision, since post-training is a specific stage applied to an existing base model, but the sentence already says the training was done on top of NVIDIA's model, so the meaning survives. Third, the digits 27 are written as "twenty seven" per the numbers as words rule, and "27 years of Salesforce CRM intelligence" is paired with the body's "nearly three decades of CRM deployments" and rendered as "twenty seven years of Salesforce deployments"; the two lines are the same fact stated twice in the same document, and the smaller, more specific figure was chosen over the vaguer "nearly three decades" so that nothing is rounded up. The acronym CRM is not spoken anywhere in the narration and "CRM deployments" is narrated as "Salesforce deployments", because the acronym is not reliably pronounced by the speech engine and because a viewer does not need it; the keyword field and the title candidates carry CRM instead. The attribution verb is "Salesforce says" because the training recipe, the synthetic dataset and the twenty seven year figure are the vendor's own description of its own unpublished training process, with no dataset card, no model card and no external verification in the source. Scene two, single sentence: "Salesforce says that on its own benchmark, Koa matches or exceeds a leading model with three times fewer errors, but never names it." Supported by "In Salesforce's CRM benchmark, a model benchmark that includes a suite of real-world tasks like updating an opportunity, routing a case, or scheduling a follow-up, Koa already matches or exceeds leading model performance on CRM actions with three times fewer errors." The phrase "on its own benchmark" is not a characterisation we invented; the source itself names the benchmark "Salesforce's CRM benchmark", so the ownership of the benchmark is the source's own wording, and stating it out loud is the single most important attribution in this Short. The digits in "three times fewer errors" are written as words. "Leading model performance" is narrated as "a leading model" so the sentence reads as plain speech; the comparison remains as vague in our narration as it is in the source, which is the point of the closing clause. The word "already" was dropped for length. The task examples, updating an opportunity, routing a case and scheduling a follow up, were dropped for length, and "follow-up" is in any case a hyphenated compound the speech engine stutters. The closing clause "but never names it" is our own observation and it was verified against the full snapshot: the article names no competing model anywhere, gives no error rate for either side, gives no sample size, gives no methodology beyond the one clause quoted above, publishes no benchmark results table, and cites no third party evaluation. That clause exists specifically so that a viewer cannot hear the three times figure as a measured, checkable result. Scene three, single sentence: "Salesforce says Koa already runs internally in a Slack agent and in select customer pilots, and general availability is only expected this winter." The first half is supported by "Koa is already in use inside Salesforce, including an agent in Slack that helps employees find information and complete everyday tasks in Slack, and is now moving into customer pilots with 1-800Accountant, Baxter Credit Union (BCU), Engine, Formula 1, UChicago Medicine, and Xero." The six named pilot customers are deliberately not spoken: the list is six items against a twenty five word scene ceiling, and the first name is a glued brand and digit token that the speech engine garbles badly. The second half is supported by the Availability section, "Koa: Available to select pilot customers now in Agentforce; general availability expected winter 2026 in U.S. regions." The year in "winter 2026" is not narrated as digits and is not narrated as a year at all; because the story date and the publication date are both September 15, 2026, winter 2026 is the coming winter, so it is narrated as "this winter", which avoids reading a date aloud while keeping the timing exact. The qualifier "in U.S. regions" was dropped from this scene for length, which makes the availability sound slightly broader than the source states, so the word "only" was kept in front of "expected" to hold the scene on the conservative side, and the hook has already said plainly that the product is not generally available. The word "expected" is the source's own word and is not upgraded to "will". The attribution verb is "Salesforce says" because the internal usage claim and the pilot status are entirely the vendor's account of its own deployments, with no customer confirmation quoted for the Slack agent and no usage figures of any kind. Scene four, single sentence: "Salesforce also says no customer data trained Koa and that it runs on Salesforce systems, but no outside audit is cited." The first clause is supported verbatim in substance by "No customer data was used to train the Koa reasoning model", and by "Its training corpus was built entirely from synthetic scenarios." The second clause is supported by "Salesforce controls the model weights and performs post-training and inference entirely within its own trust boundary", and by "Salesforce controls the weights and runs Koa within its own infrastructure, ensuring that no customer data crosses the trust boundary during inference." The source's phrase "trust boundary" is deliberately not spoken, for two reasons: it is vendor security framing rather than a plain description, and it is not self explanatory to a viewer in one pass. It is narrated instead as the literal thing the source describes, that the model runs on Salesforce systems. The closing clause "but no outside audit is cited" is our own observation and was verified against the full snapshot: the article cites no auditor, no certification, no compliance attestation, no penetration test and no third party review of either the training data claim or the inference isolation claim. Both claims are a company describing its own internal practices. The attribution verb is "Salesforce also says" for the same reason. Endcard: fixed channel boilerplate, carries no claim about the story. SENTENCE BUDGET NOTE. The script runs eight spoken sentences in total, two in the hook, one per scene and two in the endcard, and one hundred and twenty two narrated words. Each scene was written as a single sentence from the start, joined with the connectives "and" and "but" rather than split at full stops, because full stops cost more spoken time than words do on this voice. No fact was added, removed, softened or strengthened by that choice. HONESTY CAVEAT. This is a vendor press release. It is two companies, Salesforce and NVIDIA, jointly announcing a product they built together, published on Salesforce's own newsroom, quoting Salesforce's own chair and chief executive Marc Benioff and NVIDIA's own founder and chief executive Jensen Huang, and quoting four customers who are participating in the pilot programme being announced. It is not independent reporting, it is not a benchmark paper, it is not a peer reviewed result and it is not a regulatory filing. The central performance claim, three times fewer errors, comes from Salesforce's own CRM benchmark, which Salesforce built, ran and reported. The model Koa is compared against is not named anywhere in the article. There is no published methodology beyond a single clause listing three example task types, no error rates for either side, no sample size, no task counts, no scoring rubric, no confidence interval, no repetition and no independent evaluation of any kind. The claims that no customer data was used in training and that inference stays inside Salesforce's own trust boundary are Salesforce's own statements about its own internal practices, with no audit, no certification and no third party attestation cited. The product is not generally available: it is in select customer pilots now, and general availability is only expected in winter 2026 and only in U.S. regions. Accordingly every capability, performance and practice claim in this Short is attributed out loud to Salesforce, the benchmark is named as Salesforce's own in the same sentence as the three times figure, the unnamed comparison model is called out in that same sentence, the absence of an audit is called out in scene four, and the pilot rather than shipped status is stated twice, once in the hook and once in scene three. No marketing adjective from the source is repeated anywhere in the narration: "purpose-built", "unprecedented", "trusted", "agentic enterprise", "deep technical collaboration", "mission-specific" and "specialized" are all absent. CORROBORATION STATUS. There is no corroboration for this story. The research step captured zero video transcripts: the box's yt-dlp build is too old and returned nothing for any candidate video, which is a tooling failure and carries no information about the story itself. The Salesforce newsroom press release, read from the local snapshot captured 2026-09-15T12:02:22Z and re-fetched live on the run date with a matching result, is therefore the single source, and every narrated sentence including both hook sentences traces to a line quoted above. No transcript was obtained, none is cited, and nothing from any candidate video title is narrated as fact. Because the story is same day, the article dateline being September 15, 2026 and the run date being 2026-09-15, no timeline block was written into this script; adding one would age the story to the catalog tier. DELIBERATE OMISSIONS. Omitted for length, for speech engine safety, or because the item is the vendor's own framing: the Marc Benioff quotation in full, including "The most valuable thing Salesforce has built isn't our platform , it's the accumulated knowledge of how enterprise business actually works" and "it runs entirely inside your trust boundary"; the Jensen Huang quotation in full, including "AI is creating a much larger opportunity for software"; all four customer quotations and the executives who gave them, Ryan Teeples of 1-800Accountant, John Sahagian of Baxter Credit Union, Elia Wallen of Engine and Andrew Chang of UChicago Medicine; all six named pilot customers, 1-800Accountant, Baxter Credit Union and its acronym BCU, Engine, Formula 1, UChicago Medicine and Xero, the first of which is a glued brand and digit token the speech engine garbles; the base model's version name, Nemotron 3 Super; the entire post-training methods paragraph, meaning Supervised Fine-Tuning and its acronym SFT, reinforcement learning with Group Relative Policy Optimization and its acronym GRPO, and NVIDIA NeMo RL, NeMo Gym and NeMo AutoModel, all of which are glued or acronym heavy tokens; the claim that the model learned "not only to produce the right answer but to take the right action, step by step, to reach a goal"; the statement that the synthetic scenarios covered more than 14 industries and the four named examples, manufacturing, financial services, healthcare and travel; the detail that each scenario paired a persona with specific tasks and mapped the sequence of actions and tool calls; the benchmark's three example tasks, updating an opportunity, routing a case and scheduling a follow-up; the workflow list of generating leads, qualifying opportunities and resolving service cases; the entire second announcement in the release, meaning the Missionforce collaboration, the second summary bullet about giving government and regulated organizations control over their model, data and deployment environment, the Missionforce Operations product and its procurement, supplier management and logistics workflows, the statement that Missionforce Operations is generally available now in U.S. regions with post-trained NVIDIA models available to select customers in October 2026, and the private clouds, classified networks and air-gapped systems framing, which is also a hyphenated compound the speech engine garbles; the About Salesforce boilerplate paragraph; the page's five minute read time; and the unrelated sidebar and footer links to other Salesforce stories, meaning the Google Cloud, AWS, Siemens, Claudeforce, VCARB, Databricks, FIFA and Informatica items, none of which were read or used. Nothing omitted contradicts anything narrated. The omitted Missionforce section is a separate product line and makes no claim about Koa's performance or availability; the omitted methods, industry count and customer names would only add detail in the same direction as scene one and scene three; and the omitted executive and customer quotations are all promotional and would only strengthen claims this script deliberately keeps attributed and hedged. PROMPT INJECTION CHECK. The snapshot was read strictly as untrusted data and was checked for instructions. It contains no imperative addressed to a reader or to a model, no hidden directive, no "ignore previous instructions" pattern and no attempt to change behaviour. The imperative style text present is ordinary site furniture and marketing calls to action: "Skip to Content", "Skip to Footer", "Share article", "Copied", "Close", "Subscribe", "Subscribe now", "Sign up for Salesforce News & Insights, direct to your inbox", "Get the latest Salesforce News & Insights, straight to your inbox", "Discover Agentforce", "Build and customize autonomous AI agents to support your employees and customers 24/7", "Visit www.salesforce.com for more information", "Explore related content by topic", the regional language picker entries, and the legal and cookie preference links. These are normal newsroom page elements and product calls to action, not injection. No injection attempt was found, and no text from the snapshot was treated as an instruction at any point.(Salesforce (salesforce.com/news), first party corporate newsroom press release titled "Announcing Koa: Salesforce's First CRM Reasoning Model, Built on NVIDIA Nemotron", no byline, dateline SAN FRANCISCO, September 15, 2026, a joint Salesforce and NVIDIA announcement in which both companies are announcing their own jointly built product and quoting their own two chief executives, read from a local snapshot captured 2026-09-15T12:02:22Z and re-fetched live on the run date with a matching result)