AI agents with a name and a job title: what Salesforce's announcement teaches SMBs

On September 11, 2026, Salesforce introduced seven "job-ready" AI agents, each with its own name and a defined role. This isn't a marketing detail: it's the clearest signal yet of where AI in business is heading. The "assistant that knows everything" is over; what's coming are specialized agents that fill a specific business function.

What exactly was announced

Agent	Function	Status
Casey	Customer service via voice, SMS, WhatsApp and web: questions, returns, escalation	Available
Paige	Internal IT and Human Resources requests	Available
Carter	E-commerce: product discovery and checkout	Available
Hunter	Outbound sales prospecting in multi-week campaigns	Pilot (available in November 2026)
Marshall	Supply chain and back office, with an audit trail	Available
Piper	Handling inbound leads from the website and email	Available
Fin	Complex customer service workflows	Available

The most important technical novelty is the long-running runtime: agents that pursue a goal over days or weeks, keep memory across sessions and course-correct, instead of forgetting everything when the chat closes. Hunter is the first to use it: it doesn't answer a question, it works a pipeline.

Salesforce backed the announcement with customer results: Hibbett automated 90% of its shopping journeys in six weeks, Autism Queensland resolves 70% of its administrative requests with Paige, and Perk attributes 60% of its pipeline to Hunter. These are the vendor's own figures and should be read as early evidence, not as a guarantee. But the direction is unmistakable.

Why it matters even if you'll never use Salesforce

A Mexican SMB isn't going to buy seven enterprise agents. What it can — and should — copy is the mental model:

1. One agent, one job. An agent that "does everything" fails at everything. One that only follows up on quotes knows exactly what data it needs, what it can promise and when it should hand the case to a person.
2. Goals, not conversations. The value isn't in answering a message well; it's in taking an account from "quote sent" to "order closed" over two weeks.
3. Business data, not model memory. Marshall works because it sits on top of real inventory and orders. Without connected data, a specialized agent is just a chatbot with a nice name.
4. Auditing from day one. Every action is logged: who requested it, what data it queried, what it changed.

The four "jobs" an SMB can fill first

If we translate Salesforce's list to the reality of a distributor, a professional firm or a services company in Mexico, the roles with the fastest return are these:

• The SMB's "Piper": immediate lead response. Reply in under a minute via WhatsApp or web, qualify with three questions, and schedule or quote. This is where the most sales are lost to sheer slowness.
• The "Hunter": quote follow-up. Check daily which quotes have gone more than three days without a response and follow up with context: current price, stock, delivery date.
• The "Marshall": inventory and purchasing. Detect products below minimum, propose the purchase order to the usual supplier and flag it before an order goes unfilled.
• The "Casey": collections and after-sales. Reminders before the due date, reconciliation of applied payments and payment links, with a clean handoff to a person when the customer asks for one.

Owner: Which quotes over 50 thousand pesos are still unanswered this week?

Follow-up agent: There are four. The oldest is COT-1187 for Constructora Altavista, 86,200 pesos, sent 9 days ago; the customer opened the PDF twice. Shall I schedule a follow-up call today at 11 and let Mariana know?

What separates a specialized agent from a chatbot with a name

• Permissions scoped to its job. The collections agent shouldn't be able to change prices; the sales agent shouldn't be able to cancel invoices.
• Explicit autonomy limits. What it can do on its own (send reminders, quote at list price), what requires approval (discounts, extensions) and what it never does.
• Its own metric. Each role is measured by its number: time to first response, quote close rate, days of inventory, pesos recovered.
• Handoff with context. When a case escalates, the person receives the full history, not a "the customer wants to talk to someone."

In Centella, the assistant operates on the same database where customers, quotes, inventory, invoices and collections live, with the same roles and permissions as the rest of the system. That's what makes it possible to assign it concrete "jobs" without it inventing figures or overstepping its authority.

Checklist for hiring your first specialized agent

1. Which role has the most repetitive work and a result measurable in pesos?
2. Is the data that role needs in a single system and up to date?
3. What can the agent do without approval, and who approves the rest?
4. How will I know in 30 days whether it worked? Define the metric before you start.
5. Who reviews the action log every week?

Frequently asked questions

Do I need several agents from the start? No. Start with a single role — usually lead response or quote follow-up — and add the next one once the first has stable metrics.

Isn't a long-running agent risky? It is if it has no limits. With role-based permissions, approvals for anything sensitive and a complete log, an agent that works for weeks is more controllable than a salesperson without a CRM.

Does this replace my team? In the SMBs we work with, the typical effect is different: the team stops chasing pending tasks and focuses on closing. The agent does the follow-up nobody had time for.

---

Large companies are already "hiring" agents by role. An SMB can do the same with a smaller scope and a faster return. Discover Centella, explore our AI agent services or tell us which role you'd like to fill first.