The AI support stack, decoded for non-engineers

Somewhere in your queue right now, an agent is pasting a customer's question into a chatbot on their personal phone, copying the answer back, and tidying it up before they hit send. Nobody told them to. They're just faster that way — and the research says they're right.

That quiet workaround is the whole AI support debate in miniature: real productivity, no oversight, and a vocabulary — LLMs, RAG, agents, guardrails — sold to leaders as if it were one thing. It isn't. It's a stack of four distinct layers, and knowing which one a vendor is actually selling you is the difference between a useful tool and an expensive apology.

Layer one: the model, which is confidently wrong

A large language model (LLM) is a text-prediction engine. Feed it words and it returns the statistically likely next words. That is the whole trick. It has ingested much of the public internet, so it sounds fluent about almost anything — including your refund policy, which it will cheerfully invent. On its own, an LLM knows nothing about your product, your tickets, or the customer in front of it.

What it is genuinely good at is language: rephrasing, summarising, drafting. That alone moves the needle. In the largest field study to date, support agents given a generative-AI assistant resolved measurably more issues per hour — and the lift was wildly uneven.

The gain concentrated among newer and lower-skilled agents; veterans barely budged.

Read that as the assistant bottling the tacit knowledge of your best people and handing it to everyone else. It is a floor-raiser, not a ceiling-lifter — and that alone should reframe it from a headcount play into the fastest onboarding programme you have.

Layer two: RAG, or teaching the model your business

To stop the confident invention, you connect the model to your own knowledge — help centre, macros, past tickets, order data. The technique is retrieval-augmented generation (RAG): before answering, the system retrieves the relevant documents and instructs the model to answer only from them.

This is the layer that matters most and gets discussed least. A model with good retrieval over a clean knowledge base is genuinely useful; the same model pointed at a stale, contradictory help centre will cite the wrong policy with total confidence. Most of what gets blamed on "the AI" starts right here — not in the model, but in the mess it was told to trust. RAG doesn't fix bad documentation. It broadcasts it.

Layer three: agents, which actually do things

An "AI agent" is a model wired up to take actions — look up an order, issue the refund, update the ticket — not just talk about them. This is where "autonomous resolution" lives: cases closed end to end with no human. It is real, and it reaches further than sceptics expect.

That is a genuine ceiling, from a company with tidy processes and a narrow, high-volume problem set. The industry-wide expectation is more modest, but still steep.

The trap is treating that figure as a target rather than an outcome. Autonomy scales with how structured and repetitive your contacts already are. "Where's my order" automates beautifully; "I was double-charged during a house move and I'm furious" does not.

Layer four: guardrails, the part everyone skips

The final layer is the least glamorous and the most likely to sink you: the rules governing when the AI acts, what it discloses, and how cleanly it hands a conversation to a human. Customers have firm opinions here.

They want to know a decision was made by AI, and why — and they notice a clumsy handoff that makes them repeat everything. Your own staff will route around missing guardrails, too: shadow AI, the unapproved phone in that opening scene, spreads fastest exactly where sanctioned tools are thin. If you don't build the assisted layer, your agents will build it for you — off the books, with your customers' data.

What to actually ask a vendor

You don't need to read the model weights. You need four questions. Which layer is this — assist, retrieval, or autonomy? What does it read, and how fresh is that source? When it's unsure, does it stop or does it guess? And how does the handoff feel from the customer's seat?

Get straight answers and the vocabulary stops being intimidating. It is four layers, each with its own payoff and its own failure mode — and each is a separate decision, not a single "do we do AI" switch. For where autonomy pays off and where a human has to keep the pen, read agent assist vs. autonomous; for the metrics to judge any of it honestly rather than optimistically, start with the benchmarks.