Companies are giving up on chatbots. Here's what they build instead.

For two years, the answer to "what AI should our company get?" was almost always the same: a chatbot. On the website, in customer support, for internal staff questions. In 2026, though, the centre of gravity visibly moved somewhere else — and it's a shift that makes far more sense than it first appears.

Companies haven't stopped using AI. They've stopped building it as a chatty window and started building it as a quiet clerk that does a specific piece of work and hands it to a human for approval.

Why a chatbot on its own often disappoints

A chatbot is "read-only" by nature. It interprets a question, generates an answer and stops. It writes nothing into your accounting, records no order, prepares no invoice for payment. That's fine if what you're solving is repetitive customer questions. But plenty of companies bought a chatbot expecting relief from work — and got relief only from answering.

The second problem is measurability. Return on a chatbot is hard to calculate: how many hours did it actually save? How many of those questions would nobody have asked anyway? With a bounded task, by contrast — "sort incoming e-mails and attach them to the right job" — the saving is worked out in minutes, because you know exactly what that activity cost by hand.

And third: a chatbot is the most visible part of AI, so when it gets something wrong, everyone notices. Automation running in the background with human approval before the final step has far less of that problem.

What the data shows

Analyst firm Gartner estimates that by the end of 2026, 40% of enterprise applications will include AI agents built for specific tasks — up from under 5% in 2025. In its AI maturity survey, Gartner also reports that 85% of enterprise leaders consider integrating AI into workflows their main investment priority — so not "having AI", but "having it plugged in where the work actually happens".

That's precisely the shift: from a standalone chat window to AI built into a process the company has had all along.

What it looks like in practice

A few concrete examples of tasks that run automatically in companies today — and what they have in common is that they're narrowly defined:

Sorting incoming mail. The system reads an e-mail with an attachment, works out whether it's an invoice or a contract, identifies which company it concerns, and files it where it belongs. A person only checks and confirms. On one of my projects this step runs at 98–99% classification accuracy in live operation — and saves hours of manual mail handling every week.

Preparing briefing notes. Before a meeting or a quote, AI pulls a summary out of the communication history and internal documents: what we discussed with the client, what we last offered them, what was left open. Nobody has to dig through a year of e-mail.

Routing requests. A question comes in — does it go to complaints, sales, or technical support? At higher volumes that decision alone is several hours of work a day, and it automates well.

Draft replies instead of automatic replies. A crucial nuance: the system doesn't send the answer to the customer itself, it prepares a draft that a person checks and sends in a few seconds. You keep the speed, without the risk of nonsense going out in your name.

The key phrase: bounded task

This is the most important thing in the whole article. The difference between automation that works and automation that ends up in a drawer usually isn't the model or the price. It's how precisely the task is bounded.

"Handle our admin" is a brief that cannot end well. "From an e-mail with an attachment, determine the document type and assign it to a job" is a brief with a clear input, a clear output, and a way to verify whether the result is right.

The same logic applies as with writing prompts — the more specific the brief, the better and above all the more verifiable the result (I wrote more about this in What is a prompt). Only here it applies to a whole process rather than a single answer.

Where the line is — and why a human belongs on it

Fully autonomous AI agents that plan their own steps and act unsupervised are technically available today (I wrote about them in AI agents in 2026). For most small and mid-sized companies, though, they don't make sense as a first step — for two reasons.

The first is practical: the more steps an agent takes on its own, the harder it is to trace where exactly something went wrong. When a process has one checkpoint before the final step, the error is caught before it does damage.

The second is security. An autonomous system with access to your systems is also a usable target — and that this isn't a theoretical worry was shown by the first documented fully autonomous attack this summer. The broader the permissions you give an agent, the more carefully you have to manage who can do what with them.

From a regulatory standpoint, the company that deploys AI is always answerable for its output — and from August 2026 oversight of the AI literacy obligation under the AI Act begins as well. Seen that way, a step where a human approves the output is the cheapest insurance you can buy.

How to tell which process is worth automating

1. It repeats. Daily, or at least weekly, always the same way. A one-off isn't worth automating — the setup takes longer than doing it by hand.

2. It follows a clear rule, even an unwritten one. If a colleague can explain what they base the decision on, it can be automated. If they say "you just have to feel it", it's a poor candidate.

3. You can verify whether the result is correct. With document sorting you know immediately. With "come up with a strategy for us" you don't.

4. It takes measurable time. If the activity costs an hour a week, the return will be negligible. If it costs a day a week, it's an obvious candidate.

5. A mistake there won't cause a disaster. Or can be caught by a check before sending. Starting with automation where an error means losing a client or money is needless risk.

So is the chatbot dead?

No. A chatbot makes excellent sense where you genuinely handle a large volume of similar questions — an e-shop, a clinic, a school, a service centre. If dozens of identical questions about opening hours, appointment availability or order status arrive every day, it's still one of the fastest-returning AI investments there is.

What mainly changed is that the chatbot stopped being the automatic answer to everything. It's one tool among several — and often not the first one a company should deploy. When someone tries to convince you a chatbot will fix your operational problems without asking what those problems actually are, that's a warning sign.

An honest conclusion

The shift from chatbots to bounded automation isn't a fashion wave, it's maturing. After two years of experiments, companies added up what actually saved them money — and it was the boring things: sorting, routing, preparing notes, checking. Not an impressive demo, but a process that quietly does a piece of work every single day.

If AI "hasn't landed" at your company yet, it's often not because it doesn't suit you. It's usually because it was deployed as a universal tool instead of on one specific, well-bounded task. And that's a fix that doesn't normally take months.

Frequently asked questions about workflow automation and chatbots