AI customer service software: what each category actually solves

There is a conversation that repeats in every support software evaluation: three vendors, three demos, and the feeling that all three do the same thing.

The feeling is correct in the demo. In the demo they all answer a frequent question well, they all show a dashboard with an automation percentage, and they all have a WhatsApp integration.

What the demo does not show is the case nobody prepared for.

Necessary note: what this post says about other platforms is what each one publishes about itself, read on 17 October 2026. We do not compare performance or pricing, which none of them publish in comparable form.

Three categories that look alike in a demo

The help desk with AI built in. The unit of work is still the ticket, and AI arrives as a layer that resolves or assists. Zendesk presents itself as "an AI first service platform" and talks about AI agents that "resolve issues end to end". Intercom describes itself as "a complete system for AI and human customer service", with its Fin agent integrated into the helpdesk.

The agent platform inside a suite. The unit is the business record, and the agent lives inside the ecosystem where the data already is. Salesforce's Agentforce describes itself as "the AI agent platform that delivers 24/7 autonomous support at enterprise scale", with full integration into the Salesforce ecosystem.

The AI native agent platform. The unit is the conversation or the process, not the ticket. Sierra presents itself as "the leading conversational AI platform for businesses". StudioChat sits here, with the process as the unit of implementation.

All three categories work. Which one works for you depends on where your operation is today, not on which one has the better AI.

The axis that really separates them: what happens when the agent doesn't know

There is a distinction here that the industry itself is pushing, and it is worth taking seriously, because it changes which number to look at.

Deflecting means the case does not reach a person. Resolving means the case does not come back. They are not the same thing and the dashboard usually shows the first.

Worth saying that this is not our criticism of the category: it is what the category says about itself. Zendesk headlines its own page with "move beyond deflection with AI agents that resolve issues end to end". So the distinction is being drawn by the largest platform in the business. The question is whether the dashboard they show you in the demo respects it.

And there is an arithmetic reason to take it seriously: a deflected but unresolved case does not come back as one case. It comes back as several.

The arithmetic of that is further down, in the demo section: the resolution rate below which an agent stops removing work and starts adding it.

The comparison by what each one solves

The column that decides most is the third one. If your operation thinks in tickets, a help desk with AI will fit with no friction. If it thinks in processes that cross systems, the ticket will feel small, because the case does not close when it is answered but when it is executed.

Why StudioChat works in all three cases

There is an asymmetry between the three categories that decides the comparison, and it is worth stating plainly.

The smaller unit of work always fits inside the larger one. An agent whose unit is the process also resolves a ticket, because a ticket is a one step process. The reverse does not happen: a help desk whose unit is the ticket does not start writing into your ERP because the project matured.

Which means choosing the platform that executes leaves you short in no scenario. If all you need is answering questions, it answers them. If tomorrow you also need it to change a delivery date, it changes it. Choosing either of the other two leaves you short exactly when the case stops closing with an answer, which is when it starts to matter.

StudioChat is in that category. The unit of implementation is the process, and the agent connects to the systems that process lives in: changing a delivery date, taking an order, recording an attendance event, moving an appointment, sending the invoice they asked for.

The client figure we have published comes from Takenos: 70% of tickets resolved with no human involvement, with a response time of around one minute. It is the only one we publish, and it is the one we have.

What is worth asking any vendor, ourselves included, is the question from the previous section: what happens when the agent does not know, and what share of the cases it took came back within thirty days.

The questions to take into any demo

Four, and none of them is answered by a screen.

What the system does when it does not know. The four possible answers are answering anyway, saying it does not know, handing over with full context, or handing over blind, and only the middle two are acceptable.

What share of the cases it took reappeared within the following thirty days. If nobody measures that, the automation number cannot be interpreted.

That second question decides the most and gets answered the worst, so it is worth walking into the demo knowing why it matters. Put your volume in below and watch the resolution rate at which an agent stops removing work and starts adding it.

The threshold is the thing worth taking into all five meetings. If a returning case generates one extra contact, resolution has to clear fifty percent just to break even. With two extra contacts, the threshold rises to sixty six.

That turns a vague question about quality into a number you can demand from Zendesk, from Intercom, from Agentforce, from Sierra and from us.

What it can write into our systems and with which permissions. Reading is easy and writing is the real conversation.

And who looks at the conversations that went wrong. If the answer is nobody, or a monthly report, the system does not improve: it repeats.

None of the four is a question about artificial intelligence. All four are questions about operations, which is where this gets decided.

This object also has a page of its own, in case you want to send just that.