Salesforce Data Chief: AI Is Not Always Reliable
Salesforce published an interview with Chief Data Officer Michael Andrew on October 13, 2026, arguing that customers should be free to pick and swap the AI model behind their systems. Everything below comes from Salesforce's own newsroom page.
Transcript
Salesforce's chief data officer says its AI can make mistakes. He says customers should be free to swap models.
Andrew said traditional software gives the same result every time. He said AI systems are not always going to be reliable.
Salesforce says customers should pick whichever model fits the use case.
Andrew called AI a fuel that can come from anywhere, his phrasing not ours. Salesforce lists its custom model Koa as one option.
Andrew named Anthropic and OpenAI as frontier partners, but Salesforce published no performance comparison.
This is an interview, not a product launch. Salesforce gave no dates or prices.
Inboxsmith tracks the claims behind company announcements. Please like and subscribe for more news.
Sources
Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:
- Line-by-line backing. Hook sentence one, 'Salesforce's chief data officer says its AI can make mistakes', is backed by Andrew's transcript answer describing nondeterministic AI: 'You can get a new employee, you can try to train them to do the best possible, but that new employee might make mistakes. That's what the nondeterministic systems of AI are like', and by his line 'they're not always going to be reliable'; the speaker's title is given in the standfirst as 'Salesforce Chief Data Officer Michael Andrew'. Hook sentence two, 'He says customers should be free to swap models', is backed by Andrew's verbatim 'We believe in giving our customers choice, that you can pick the model that's right for your particular use case' and 'you have the flexibility to keep evolving and changing what model you use'. Scene 1 sentence one, on traditional software giving the same result every time, is backed verbatim by 'Businesses have always run on deterministic systems, which guarantee you're going to get the same thing every time, everywhere in the world' and by 'Software is deterministic. If you open a case about your customer, if a customer files an issue, every time you press that button, you're going to open a case.' Scene 1 sentence two, that AI systems are not always going to be reliable, is Andrew's exact wording: 'You can train them, you can guide them, but they're not always going to be reliable.' Scene 2, that Salesforce says customers should pick whichever model fits the use case, is backed by the article body 'rather than forcing customers to stick with a single tool, Salesforce gives them flexibility to use whichever model works best for any given use case' and by the opening line 'Not every job calls for the same AI model.' Scene 3 sentence one, the fuel line, is quoted and explicitly flagged in narration as Andrew's own phrasing rather than Inboxsmith's, and is backed verbatim by 'AI is a fuel; it's more fuel into your engine, and we want to let that fuel come from anywhere.' Scene 3 sentence two, naming Koa as one custom-model option, is backed by 'whether it's a frontier, open, or custom model like Koa' and by 'It's orchestrated alongside frontier models rather than in place of them'; Koa is mentioned in a single clause only, as an option, with no construction detail. Scene 4 sentence, naming Anthropic and OpenAI as frontier partners and stating that no performance comparison was published, is backed by 'you have big frontier foundation models produced from partners like Anthropic and OpenAI' and by the absence from the page of any model-performance comparison, benchmark figure, speed number, or cost number. Scene 5, that this is an interview and not a product launch with no dates or prices, is backed by the page format itself, an interview transcript in the Newsroom's Agentic Enterprise section, and by the confirmed absence of any availability date, pricing, customer count, benchmark result, third-party audit, or GA status for Koa anywhere on the page. The endcard carries no factual claim about Salesforce and is Inboxsmith house copy only. Deliberate exclusions: the Koa construction details, including the NVIDIA collaboration, the training simulations, the 27 years of CRM workflow experience, the Salesforce trust boundary, and the benchmark error-rate claim, were all excluded because Inboxsmith already published a separate Short on them, 'salesforce-koa-crm-reasoning', on September 15 2026. The multiplayer AI in Slack thesis, including shared channels, shared context, people and agents working in the open, and the digital colleague framing, was also excluded because Inboxsmith already published a separate Short on it, 'slack-multiplayer-ai', on October 1 2026. The angle retained here, the model-choice position itself plus Andrew's candid reliability admission, is covered by neither prior Short.(Primary and sole fact source for this Short is the Salesforce Newsroom article 'Salesforce Chief Data Officer Michael Andrew explains how purpose-built models like Koa and multiplayer AI in Slack fit together', bylined Cody Snell and published October 13 2026 in the Newsroom's Agentic Enterprise section, labeled a 4 minute read. This is a first-party page: the publisher is Salesforce and the speaker is Salesforce's own Chief Data Officer, so it is the company describing its own products and positioning. The URL was fetched live at production time and returned HTTP 200 on the same host with no cross-host redirect, and the live body was byte-compared against the watcher snapshot captured at alert time at scripts/data/article-cache/f6b69be639cff898.txt, which matched, so there is no stale or rotted-stub risk. The story is same-day relative to the production date of October 13 2026. The owner clicked APPROVE on the alert, and the alert was rescored at 76 HOT with no regional demotion applied. Zero of five candidate corroborating YouTube transcripts were retrievable because the installed yt-dlp build cannot fetch captions, so no secondary video source informs the script and the article stands alone as the fact base. Because the publisher is an interested party describing its own models, every statement in the narration is attributed to Salesforce or to Michael Andrew by name and nothing is asserted as independently verified capability. The fetched page contained no prompt-injection content, no embedded instructions, and no third-party user-generated text; all quoted material is editorial prose or interview transcript published by Salesforce.)