Context is the copilot

Author(s): Steve West and Matt Baker

Every customer care copilot on the market runs on the same handful of frontier models. They can all write a clean reply. What separates them is how much of the customer they can see before they write a word.

That is an unfashionable thing to say in a market that sells on model quality. But sit beside a team member on the conversations that AI hands them, the disputes and exceptions and edge cases automation could not close, and you see it right away. The prose is rarely the problem. The problem is how little the copilot can see. Give an AI copilot one message and a confident tone, and it will draft a reply that looks helpful, but amounts to a fluent guess.

Three eras of agent assistance in customer care

The first era was the snippet. Canned macros and keyword search over a knowledge base. The team member typed a phrase, scanned a list of articles, and pasted the one that looked closest. It was useful for direct lookups, and for direct lookups it still is. But it could only return the article you already knew how to ask for. The context it carried was whatever you found yourself.

The second era was the fluent stranger. Large language models arrived and draft quality jumped. The copilot could read the open conversation and produce a clean, on-tone reply in seconds. The catch being what it can read. Most AI copilots are built on top of a single ticket or session, so they see the interaction in front of them and nothing behind it. They write beautifully about a customer they have just met. On a simple question, that is fine, but on the conversations that reach a human, it is the difference between a correct answer and a plausible-sounding mistake.

The third era is the situated copilot. Team Assist is the AI copilot inside the Gladly Team workspace: it reads the live conversation, the customer's full history, and the knowledge base, then drafts a grounded response the team member reviews and sends. A situated copilot develops a rich understanding of the customer before it drafts: every prior conversation, across every channel, every order, alongside the full knowledge base. Instead of retrieving one article, it synthesizes across many and shows its sources. The model is the same one everyone runs. What's different is the context it has access to before it writes.

What a situated draft actually requires

Consider a warranty dispute. The customer bought a product months ago, bought an accessory for it in a separate transaction weeks later, and has already contacted support once about the same issue with a resolution that did not hold. A team member picks this up after AI hands it off.

A ticket-scoped copilot sees the newest message and drafts from it. A session-scoped one sees the open chat. Both will write something fluent, and both will miss that the accessory was a different purchase under a different return window, and that this customer has been here before. Team Assist reads the full relationship via the Gladly Context Graph, the continuous record of every conversation a customer has ever had with the brand. The policy that applies here is not in one article. It lives across three: the base-product warranty, the accessory exception, and the handling for a repeat contact. Team Assist combines them into one grounded answer and attaches inline citations, [1] [2], linked to the exact articles it used, so the team member can verify each claim before anything goes out.

That is the part the model leaves to you. Writing the sentence is easy. Knowing which three facts belong in it is the whole job, and that knowledge comes from the context.

The difference context makes

What the AI copilot can see or do

Snippet and KB search

Generic copilot (ticket or session)

Gladly Team Assist

The current message

yes

yes

yes

The open conversation

no

yes

yes

Full lifelong history, every channel

no

no

yes

Live order and transaction data

no

no

yes

Synthesis across the knowledge base, with sources cited

no

partial

yes

Execute an action, with a human verification step

no

varies

yes

Every row below the second is something the model cannot infer and can only be given. The copilots in the first two columns write just as well. They simply have less to work with.

When the model doesn't know the customer

At low volume, a strong model papers over thin context. Ask it something simple and the single message is enough. The gap stays invisible right up until the conversation is hard, and the hard conversations are exactly the ones AI routes to a person. By design, the easy work is already done. What reaches a team member is the dispute, the exception, the high-value customer whose loyalty is on the line. Those are the conversations where missing context turns a confident draft into a wrong one, and they are the most valuable conversations a brand has. A copilot that is strongest on the easy cases and weakest on the hard ones is optimized backwards.

There is a strategic version of the same point. When every copilot draws from the same models, model quality stops being a differentiator and becomes a floor. What is left is context: how much of the customer the copilot can see, and how much of the brand's own knowledge it can safely act on. That is not a prompt you can copy. It is built over years. Gladly's record runs across eleven years and 450 million retail conversations, and Team Assist reads from it natively. The copilot grows sharper as that record and that knowledge grow, which means the advantage compounds in the one direction model quality cannot.

The model writes the sentence. The context decides whether it is true.

On the easy questions, any fluent copilot will do. The hard ones, the disputes and loyalty calls that AI hands to a person, go to whoever's copilot actually knows the customer, not just the message in front of it.

Fluent was never the same as right.