Beyond the Chatbot: What It Actually Takes to Make an AI Agent Work

Every retail and hospitality brand hits the same wall eventually. A customer emails about a return, or a guest asks about a booking, and someone on the service team has to stop, find the right context, draft a reply that sounds like the brand, and hope the next hundred cases don't arrive before they've finished the first. It isn't really a technology problem. It's a capacity problem. The knowledge exists. The policies exist. What's missing is the ability to apply them at the speed customers now expect.

This is the problem Salesforce's Agentforce is built to solve, and it's one we've spent the last year working on in practice. Our team builds agents across multiple areas of the business, service and loyalty among them, and what follows isn't theory, it's deployed capability for brands in the two sectors we know best: hospitality and retail. Here's what we've actually learned doing it, rather than what the vendor deck says.

What Agentforce Actually Changes

It helps to be precise about what's new here, because the term “AI agent” has become loose enough to mean almost anything. Agentforce isn't a scripted chatbot with better copy, and it isn't a general-purpose copilot that drafts suggestions for a human to rewrite from scratch. It sits somewhere more useful in between. A reasoning layer, which Salesforce calls the Atlas Reasoning Engine, takes a request, breaks it into steps, and pulls from your live business data (case history, knowledge articles, order records) to work out what should actually happen next.

The distinction that matters in practice is this. A good agent doesn't just suggest an answer. It drafts a response grounded in the right knowledge, routes the case to the right place, and knows when to stop and hand off to a person. That last part, knowing when to stop, turned out to be the most important design decision in both of the customers below.

Two Deployments, One Underlying Pattern

We build Agentforce agents across a number of different areas of a  business, service and loyalty among them, and we scope each one around a specific business problem rather than a generic “add AI” brief. Service is where we have our clearest, most measurable client results today: a UK hotel group and Harvey Nichols. The problems looked different on the surface. One was about response speed and personalisation, the other about volume and capacity. But the underlying approach was the same in both, and it's the same approach we carry into every agent we build, regardless of which part of the business it sits in. Ground the agent in unified knowledge, automate the highest-frequency, lowest-risk work first, and keep a human reviewing anything customer-facing until the pattern is proven.

A UK Hotel Group: Faster, More Consistent Sales and Service Emails

The problem: sales and service teams were spending significant time drafting and personalising communications, which slowed response, and knowledge assets were fragmented across systems, making consistent, efficient resolution difficult even for experienced agents.

Rather than one single feature, we implemented three connected pieces of Agentforce capability, each targeting a different point of friction in the sales and service workflow:

• AI-generated record summaries. Leads, opportunities, accounts and case records were automatically summarised so that anyone opening a record got the context immediately, rather than having to read through every field individually to understand what was going on.

• AI-drafted sales follow-up emails. Based on each customer's follow-up date, Agentforce automatically drafted a personalised follow-up email and created a task for the sales rep to review it, edit it if needed, and send. This took a task that previously meant starting from a blank page every time and turned it into a quick review, saving the sales team significant time.

The email drafting piece specifically delivered a 40% faster response time and a lift in CSAT.

• A knowledge-grounded AI agent on the website. Deployed directly on the client's website, this agent answered customer questions using the knowledge articles held in Salesforce. Where no relevant article existed, it escalated the conversation to a human agent rather than guessing at an answer.

Together, these three pieces reduced the manual work sitting between a customer's question and a useful answer, whether that question came in through the website, a case, or a sales follow-up, and gave the team a consistent, brand-aligned AI foundation they can build further self-service capability on.

‍Harvey Nichols: Automating the Return/Refund Queue

The problem: agents were manually categorising high-volume, repetitive case types (returns, refunds, order tracking) and drafting templated responses by hand. That drained capacity that should have gone toward genuinely complex customer issues.

Rather than trying to automate every case type at once, we scoped an Agentforce MVP around the highest-frequency category first, return and refund enquiries, and built out from there. The agent drafts context-aware responses grounded in existing knowledge articles for agent review and personalisation, and intelligent routing prioritises the cases that are ready for automation so they don't sit in the same queue as everything else. The MVP also established baseline metrics, which now form the roadmap for expanding automation into further case categories.

‍

‍

What We'd Tell Anyone Starting This Journey

A few things hold true across both projects, and across most of the Agentforce work we see going well or badly elsewhere in the market:

• Start with one high-frequency, low-risk case type, not everything at once. Harvey Nichols scoped an MVP around returns and refunds specifically, rather than trying to automate the whole service inbox in one pass. That's what made the roadmap credible rather than aspirational.

• Fix the knowledge problem before you automate on top of it. Both builds treated unified, accessible knowledge as the foundation the agent sits on, not an afterthought. An agent grounded in fragmented or duplicated knowledge will confidently produce inconsistent answers, which is worse than no automation at all.

• Keep a human reviewing anything customer-facing until you trust the pattern. Generative drafts for agent review and personalisation, not unsupervised sends, is the model that's actually earned trust internally in these deployments.

• Treat thresholds and guardrails as business decisions, not technical ones. The controls that decide what an agent can do automatically, and where it must hand off to a person, need to sit with the people who understand the risk, not be buried in configuration only a developer can reach.

• Measure a baseline before you expand scope. Harvey Nichols' MVP wasn't the end state; it was designed to produce the metrics that justify the next phase. That's a more durable way to build a business case than a single big-bang rollout.

Where This Goes Next

The two examples above happen to sit within service, the queue where response time and consistency are most visible to a customer or guest. But service is only one part of what we do. The same pattern, grounded knowledge, staged automation, human-reviewed output, business-owned guardrails, is what we carry into our other agent work too, loyalty included, and it's the same discipline we'll apply as we build agents that hand off to one another across different parts of the business rather than operating in isolation. That cross-agent world is still taking shape across the industry, and we're building toward it alongside our clients. For now, the point of this piece is simpler. The technology works when it's scoped properly, and the discipline behind it is what actually determines whether it delivers.

‍

If you're evaluating where an AI agent could genuinely reduce response time or free up your team's capacity, in retail, hospitality or elsewhere, we'd be glad to talk through what we've learned in more detail.

‍