
What happened
At Dreamforce, Salesforce's Marc Benioff said roughly 30,000 of its about 150,000 customers use Agentforce, the feature launched in 2024 that builds digital workers for business tasks.
Why it matters
Salesforce's own adoption figures suggest the agentic vision is still a work in progress, and the brands on the floor are taking a more cautious version of it, automating more but not everything.
What to watch
The test is where each brand draws the line — AT&T is stopping short of using AI agents to handle cancellations, and Crocs' Feliz Papich said a human still needs to be in the loop while automation's complexities are learned.
WHO IT HITSEnterprise software buyers and the customer-service and support leaders deploying these agents will be watching whether vendors deliver on data quality and not just slick interfaces. Crocs' Feliz Papich pointed to the data foundation under the agent as the complexity that often gets lost.
Summaries like this, in your inbox every morning.
Salesforce's Dreamforce stage was the setting for a reality check on AI agents. CEO Marc Benioff wants agents to take on a lot more work that humans do, but the adoption numbers he cited — roughly 30,000 of about 150,000 customers on Agentforce, a feature launched in 2024 — show how far that vision still has to travel. The brands on the conference floor gave a more cautious version of it.
Two contrasting approaches stood out. AT&T is using AI to cut the time employees spend upgrading a customer's phone from about 30 minutes to roughly 10 minutes, yet it stops short of letting agents handle cancellations, wanting humans to understand why a customer is leaving. Southwest Airlines uses chatbots for servicing tasks like flight changes while reserving human employees for 'right here, right now' problems, where empathy and assurance matter. Crocs, which runs a consumer-facing chat and voice bot called Rivet and internal agents for promotions and product categorization, still keeps human hands on the tiller, as SVP Feliz Papich put it, because the capabilities and complexities of automation are still being learned.
The through-line is a tension Papich named: the rush to build sophisticated interfaces with AI vendors can bury the harder problem of making the data underneath good enough. What the next phase hinges on is whether vendors can communicate that data complexity and whether brands keep drawing the line at the tasks where human judgment pays off.
For example, today's edition would include:
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. Q&As are published on this page for other readers too.
Applied Clinical Trials Online reports on lessons drawn from hundreds of clinical trials involving AI, coverin…

Mastercard rolled out a payment option Thursday that gives an AI agent a virtual card to buy online without ch…

The RTX 5090, launched at $1,999, now costs over $6,000 at US retailers, with third-party listings above $9,00…

Palantir CEO Karp spoke bluntly in the AI risk debate, saying aloud what the article frames as the quiet part

Co-CEO Clay Magouyrk said Oracle last year hadn't found broad internal use for generative AI; after rolling ou…

Twist Bioscience signed an agreement to provide antibody characterization data and services to Eli Lilly's AI/…
