AI at work
1. Agentforce takes on longer corporate workflows
Salesforce’s September 11 Agentforce expansion gives corporations a more practical starting point: agents designed for customer support, employee service, shopping and sales. Several are generally available, along with coordination between specialized agents. Hunter, its outbound sales agent, remains a pilot, with general availability planned for November.
The potential impact is in the handoffs: finding customer context, checking a request, updating records and bringing in a person when needed. Salesforce’s longer-running agents are designed to pursue goals across days or weeks. If they work reliably, employees could spend less effort coordinating systems and more on judgment and relationships. Buyers should test the workflow they need against capabilities available today.
Salesforce announcement, Sept. 11
2. Results need a denominator and a full cost
Smarsh reports that its Agentforce support agent achieved 72% self-service deflection across 405 interactions in the second quarter of 2026. That is useful deployment evidence, with a disclosed sample, but remains one company’s reported result rather than a universal resolution rate.
Costs need the same precision. Salesforce lists Flex Credits at US$500 per 100,000; a standard or custom production action uses 20 credits. Ten such actions therefore cost US$1 in action credits. Licensing, data services, implementation, retries and human review can add to the total. A pilot should measure successful completion and corrections alongside cost per completed task.
Smarsh results, Sept. 3 · Salesforce pricing · Aug. 31 rate card
3. Start with the customer journey
I’d start by looking beyond the answer. Chrispy’s September design discussions about AI and Agentforce emphasized gathering useful context, helping someone move through a task and giving a human representative a better starting point when assistance is needed.
Chrispy’s lesson is to distinguish predictable business steps from conversational flexibility. The wider customer journey matters too: where people get stuck, change direction and ask for help. These are his learning and design priorities, not a completed deployment or measured business results. When you evaluate agents, ask: what should happen after the answer?
Agentforce platform background
4. ChatGPT dots bring delegation to everyday admin
OpenAI introduced dots on September 29 as agents that can take on ongoing responsibilities, work across connected apps and bring results back for review. Access is rolling out gradually to eligible accounts.
Here’s one small-business example from Chrispy: a dot drafted an invoice reminder, revised the wording and sent it through Messages on his connected Mac after approval. The verified outcome was a sent reminder; payment collection was not established.
If you juggle delivery, sales and administration, that pattern could reduce the attention consumed by small unfinished tasks. I’d start with one responsibility, clear source material and a reviewable result. Record review time and corrections before claiming a productivity gain.
5. Define the boundaries before handing over work
A useful assignment names the goal, permitted sources, review points and evidence of completion. OpenAI’s controls documentation says a request to draft does not authorize sending. A dot’s cloud computer and access to the owner’s computer are separate; local work needs the connected computer online with ChatGPT open.
Corporate deployments need equally concrete checks. Salesforce warns that custom Apex and Flow actions can run in system mode and bypass the running user’s permissions. Inspect the tools and access behind the conversation. Begin with a narrow task, keep consequential decisions with a person and verify what actually happened.
OpenAI controls · Computer access · Salesforce architecture guidance
Practical AI projects
Three short experiments to try
Use invented data and existing tools. These are proposed tests, with estimated effort; no purchase or live action is required.
- Quote gap finder. Compare a fictional photography request with a draft quote. Ask AI to flag missing quantities, dates and delivery details, citing both texts. Plant three omissions and check how many it catches without inventing terms. About 20 minutes in text chat; nothing is sent.
- FAQ contradiction finder. Create three short fictional business FAQs with conflicting opening hours and turnaround times. Ask AI for a conflict list with exact source references and questions for the owner. Success means surfacing disagreements without choosing an unsupported answer. About 20 minutes in text chat.
- Delivery pack checker. Give AI six invented filenames and a promised-deliverables list. Include an outdated duplicate and a missing item. Request a client-ready index and a separate exception list, then compare with the planted defects. About 25 minutes using chat or a spreadsheet; no files are moved or deleted.