AI in customer service is two products

AI in customer service looks like one category until you watch it run inside a real support team.

Then it splits in two.

One kind of AI talks to customers. It answers questions in chat or email and tries to keep the ticket away from a human. The other kind talks to your team. It runs inside the helpdesk, pulls context from connected systems, and gives the rep a brief and a draft before they reply.

Both are useful. They are not the same purchase.

The confusion is mostly a vendor-category problem. Every product gets called “AI for support,” the pitches describe similar-sounding outcomes, and a CX leader trying to pick one is asked to compare a tool that talks to customers against a tool that talks to reps.

That is not a fair comparison. It is two products pretending to be one category.

What is agent assist?

Agent assist is AI built for your support team. It sits inside the helpdesk, whether that is Gorgias, Kustomer, Zendesk, Intercom, or another inbox, and runs the work a rep used to do before they could type a reply: opening the order in Shopify, scanning the return status, checking the previous contact, reading the policy, and picking the right response.

In practice, that looks like a brief attached to the ticket. Customer name and order, what they bought, what shipped, what they have contacted you about before, what the relevant policy says, what your team usually does in this situation, and a draft reply. The rep opens the ticket, reads the brief, edits the reply, and sends. The work that used to take six or eight tabs and four or five minutes of hunting becomes a single page of context.

Success looks like time-to-context dropping and first-reply quality going up. Containment, the metric deflection AI is measured on, does not apply. Agent assist never tries to keep a human out of the loop. The point is to make the loop faster and better.

What is deflection AI?

Deflection AI is the customer-facing bot. It sits in chat, and increasingly email, and answers a defined set of question categories without a human. Order status, password reset, store hours, return policy lookups: the questions that have stable answers and do not carry much brand risk. When it works, the ticket never reaches your team and the customer gets an answer in seconds.

The category where it works is real and worth deploying for. The category where it does not work is also real. Refunds, subscription disputes, damage, complaints, anything emotional or brand-sensitive sits below the deflection ceiling for structural reasons, not because the model needs one more round of tuning. Gorgias’s own data across 16,000 brands puts AI-alone handling of refunds and returns at 8%.

Deflection AI is not the villain in this article. It is the right tool for half the inbox. The mistake is asking it to handle the other half.

How do they compare?

The row that matters is the third one. A brief is not an answer. Agent assist hands the rep something they can edit, send, or override. Deflection AI hands the customer the answer directly. Different output, different risk profile, different category of work.

When does each fit?

The cleanest way to decide is by ticket category, not by inbox.

If 86% of AI conversations eventually involve a human, the deployment question is not “deflection or agent assist.” It is which categories each tool handles. Order status, tracking, password resets, simple policy questions: deflection AI carries them. Refunds, cancellations, subscription disputes, damage, complaints: a human carries them, and agent assist makes that human fast and informed.

The per-category ceilings make this concrete. Order status sits in a 60-70% deflection bucket. Refunds sit under 10%. A brand running deflection across both is selectively reporting the first number and absorbing the cost of the second one quietly. A brand running agent assist on the categories below the deflection ceiling is doing the math the other way: accepting that the human is in the loop and investing in making the loop better.

What does running both look like?

The mature pattern is routing by ticket category, with a clean handover between the two systems.

The customer hits chat. Deflection AI answers if the question fits a category it handles well. If it doesn’t, or the customer asks for a human, the ticket gets created in the helpdesk. Agent assist runs against the new ticket immediately, pulls the context the customer has already shared, adds order and policy and history, and writes a draft reply. The rep opens a fully-briefed ticket instead of a cold one.

The seam between the two products is the internal note inside the helpdesk. Deflection’s transcript becomes context. Agent assist’s brief sits next to it. The rep sees both before they type. Nothing the customer told the bot has to be told again.

This is where the team-side AI thesis lands operationally. Not “replace the rep.” Not “deflect harder.” Hand the rep a ticket that already has the lookups done. The 86% Gorgias number is not a deflection failure. It is the size of the surface agent assist is built for.

What should you look for?

Three things, in either category.

Clean handover. The deflection product and the helpdesk should share context instead of handing off a transcript alone. If the rep has to ask the customer for their order number again because the bot has it but did not pass it through, the integration is the gap, not the tool.

Per-category metrics, not blended ones. A vendor that can only report one deflection number is not measuring the inbox the way the customer experiences it. The same applies to agent assist: time-to-context per ticket type beats an average across the whole queue.

No replacement pitches. Any vendor still selling “replace your team with AI” in 2026 is selling the bet the data has walked back. Tools that work for the team are tools the team will use. Tools that work against them get worked around.

Two products, two jobs, one inbox. The brands getting AI in customer service right are not picking between deflection and agent assist. They are routing by ticket category and measuring each tool on the metric that actually applies to it.

Sources

• Gorgias 2026 State of Conversational Commerce. 16,000 brands, 350M conversations. 86% of AI conversations eventually involve a human; 8% of refund and return requests handled by AI alone.
• The right deflection KPI. Per-category deflection ceilings: 60-70% on order status, under 10% on refunds, under 5% on damage and complaints.
• Are people fed up with chatbots?. The rollout failures, the wins, and what 50/50 deflection looks like in practice.