The third category: what an integrated agentic system is

One question does the sorting

The question is: who starts the work? It is a question about the product, never about the technology inside it — a distinction covered in agentic AI or an agentic system.

The first two rows share something that matters more than their technical differences: nothing happens until a person shows up.

First category: SaaS with AI on top

This is by far the most common, and the easiest to identify. Take the AI away in your head and a complete product is still standing. The AI almost always does the same job in these products — it speeds up input. It drafts the message you were going to write, finds the contact you were going to look up, summarizes the document you were going to read. That is useful, and often very well executed.

It is tempting to read this as laziness or opportunism. That reading is wrong. Bolting AI on is the only rational move available to an established vendor, for four reasons that come down to architecture rather than ambition.

• The data model. A sequencing product is organized around a campaign: a list, some steps, some delays. An agentic system is organized around an objective and a state of the world — the accounts, the situations, what has been established, what is still uncertain. You do not get the second by adding columns to the first.
• Which side starts. Conventional software is request-and-response: the user acts, the system reacts. Agentic software inverts that. But the existing codebase assumes a click starts everything — sessions, interface state, how work gets scheduled. Inverting the initiative reaches the foundations.
• What the interface is centered on. A configuration surface would have to become a queue of decisions. That is not a redesign. That is a different product.
• The business model. Charging per seat, per send or per credit has nothing to do with the value of software that works while nobody is logged in. Changing the pricing breaks the revenue that would fund the rebuild — and abandons customers who are happy with what they bought.

Put plainly: in this transition the existing product is not an asset, it is a constraint. None of this is new — it is the innovator's dilemma, pointed at agents. The tests that identify this category during a demo are laid out in how to spot a real agentic system.

Second category: the connected general assistant

The second path is newer and considerably more interesting. It takes the intelligence out of the tools altogether: a general assistant — ChatGPT, Claude — drives your business software through connectors, most often over MCP today. You ask, it acts inside the CRM, the contact database, the mailbox.

What stands out is that the vendors themselves are going there. Many now publish their own MCP server. HubSpot describes its own as a way to give any MCP-compatible AI tool or agent read and write access to CRM data. Outbound vendors have gone as far as publishing blog roundups of the best MCP servers for sales.

The clearest case is lemlist, a company that otherwise markets itself as an AI-agent outbound platform. Its MCP page is titled Stop doing outbound. Start prompting it. and offers one-click connection to Claude so an outside assistant can reach its contacts, its enrichment and its sequences. A vendor selling built-in agents is simultaneously showing customers how to drive it from somewhere else.

Read this as clear sight, not surrender. These vendors worked out before their competitors that their value lies in their data and the actions it enables, not in the screens you reach them through. They are conceding that the interface becomes an execution layer something else calls, and they would rather organize that shift than be caught by it. It is the right call.

That headline also happens to be the cleanest available definition of this second category, written by one of its own participants: the work no longer starts with a click, it starts with a request. The person doing the starting has not changed.

What the MCP pattern does not solve

It solves something real, and that deserves saying plainly: the interface is out of the way, and the intelligence is no longer trapped inside each tool. Four things remain untouched.

You are still the trigger

A connected assistant does not exist between conversations. It did nothing last night and nothing while you were away. The request simply changed shape — it used to be a click, now it is a sentence. As long as work only happens when you show up, the burden of knowing what to ask for and when stays entirely with you. That burden is the heavy one.

It accumulates no model of your business

An assistant has conversation context, not a memory of your world. It does not know what was established about this account three weeks ago, what was promised, what was already tried and failed. Every session starts cold and re-reads. Access to data is not accumulated understanding.

It has no declared boundary

An assistant will do what you ask, across every tool it can reach. The line between what runs on its own and what waits for approval is whatever you remembered to put in your prompt. An integrated system carries that split in its design, where it is explicit, auditable and something you can point to afterward — the subject of our French guide on where to put the checkpoint.

It has no judgment about your trade

Calling a CRM API does not tell you that a new head of operations at a two-hundred-person manufacturer matters more, to a firm selling operational transformation, than a funding round at a startup. That discrimination is vertical. It is built on a trade, and no API documentation contains it.

Third category: the integrated agentic system

The third path is neither software with intelligence added nor general intelligence pointed at software. It is a system designed around the agent from the start, where the work and the intelligence are not two separate layers.

Four properties define it, and each one can be checked.

• It works without being asked. If nobody logs in for a week, something describable has happened: accounts watched, situations qualified, files prepared.
• It maintains a model of your world. What it knows about an account accumulates instead of being re-read each time.
• It carries a declared boundary. What runs alone and what waits for approval is settled in advance, not left to the phrasing of a request.
• It comes back at the right moment. Whatever needs judgment surfaces in one place that can actually be emptied — the job of the agentic inbox.

This category is vertical out of necessity rather than commercial preference. Trade judgment — which signals count, what they let you conclude, what you write and to whom — cannot be acquired horizontally. That is why an agentic system gets built for one line of business, and why it is honest to say it suits others poorly.

“But an inbox is an interface too”

Fair objection, and it deserves a direct answer. If the argument is that interfaces are becoming execution layers, why does an agentic system keep one?

Because two very different things share the name. A surface that exists so you can operate the software — configuration screens, forms, settings, lists to filter — really does disappear: an agent reaches it faster and more reliably than you do, which is exactly what MCP servers make possible. A surface that exists so you can decide does not, because deciding is the part that cannot be delegated.

The distinction is between operating and adjudicating. The operating work is leaving human hands, and that is good news. The adjudicating work lands squarely back on them, and it needs somewhere to land. An agentic inbox is not one more control panel. It is what is left once all the control panels are gone.

Frequently asked questions

What is the difference between a SaaS product with AI and an integrated agentic system?

In a SaaS product with AI added, the intelligence sits on top of something that already existed. It speeds up input — drafting the sequence, finding the contact, rewording the follow-up — but the work still starts when you open the tool. An integrated agentic system is built around the agent from the start: it pursues an objective continuously and comes back to you when a decision is yours to make. The difference is not how much AI is involved. It is who starts the work.

Why haven't established vendors rebuilt their products around agentic AI?

Because it is not in their short-term interest, and that is a rational position. Their data model is organized around objects you configure rather than a state of the world. Their code assumes a click starts everything. Their interface is a configuration surface. And their pricing rests on seats, sends or credits, none of which track the value of software that works while nobody is logged in. Rebuilding would break the revenue that would fund the rebuild, and walk away from customers who are perfectly happy with what they bought.

Isn't ChatGPT or Claude connected to my tools through MCP enough?

It solves a real problem: the interface is no longer in the way, and the intelligence is no longer trapped inside each tool. But it leaves a bigger one untouched — you are still the trigger. As long as work only happens when you type a request, the burden of knowing what to ask for, and when, stays with you. Three things are also missing: no accumulated model of your business between sessions, no declared boundary between what runs on its own and what waits for approval, and no judgment specific to your trade.

What does it mean that vendors are shipping their own MCP servers?

It is clear-eyed rather than defeatist. Those vendors understood before their competitors that the value of their product lies in their data and the actions it enables, not in the screens you reach them through. HubSpot, for instance, presents its MCP server as a way to give any MCP-compatible AI tool or agent read and write access to CRM data. That is an explicit admission that the interface is becoming an execution layer something else calls.

If interfaces are becoming obsolete, why does an agentic system still have one?

Because two very different things go by the name interface. A surface that exists so you can operate the software — configuration screens, forms, filtered lists — does go away, since an agent reaches it faster and more reliably than you do. A surface that exists so you can decide does not, because deciding is precisely the part that cannot be delegated. The distinction is between operating and adjudicating.