Knowledge Workers Spend 6.4 Hours a Week Botsitting AI. That's Not Productivity.

AI promised to save us 11 hours a week. More than half of that disappears again.

The promise of generative AI was remarkably simple. Ask a question. Get an answer. Save time. In many ways, that promise has become reality. Recent research shows that knowledge workers estimate AI already saves them around 11 hours every week through automation.

But there's a catch.

Those same workers spend 6.4 hours every week supervising AI. Feeding it context. Checking its answers. Correcting mistakes. Running prompts again. Cleaning up responses that looked convincing but weren't quite right. And despite almost universal AI adoption, only 13% believe their organization is performing significantly better because of it.

Researchers recently gave this hidden work a name:

Botsitting.

It's one of the best descriptions I've seen for what's really happening inside companies today.

What is botsitting?

Botsitting is all the work required to make AI useful in practice.

It isn't the AI generating the answer that consumes most of the time. It's everything around it.

Employees spend hours every week:

• explaining the business context
• checking AI-generated content
• debugging incorrect responses
• rerunning prompts
• comparing different AI tools
• fixing mistakes before they reach customers

When you add all those activities together, the average knowledge worker spends almost an entire working day every week supervising AI instead of benefiting from it. The irony is difficult to ignore. AI was introduced to eliminate repetitive work. Instead, many organizations have simply replaced one repetitive task with another.

The productivity paradox

The numbers become even more interesting when you look beyond individual productivity. Almost everyone using AI feels more productive personally. Very few believe their organization has become dramatically more productive.

That distinction is important.

Saving ten minutes writing an email doesn't automatically make the business more efficient. If another employee spends fifteen minutes checking that email, adding missing information and correcting mistakes, the organization hasn't gained anything. Individual productivity has improved. Organizational productivity hasn't. That's the productivity paradox many companies are experiencing today.

The hidden tax nobody talks about

The report introduces another interesting concept: the AI Toggle Tax.

Very few employees use a single AI tool. Instead, they jump between ChatGPT, Claude, Gemini, Copilot, internal copilots and specialized AI applications. Every switch has a cost. The employee has to explain the project again. Reload the documents. Copy previous conversations. Paste company information. Compare different answers. Decide which response is actually correct. The AI may remember nothing from the previous application, so the human becomes the integration layer between disconnected systems. That hidden switching cost slowly eats away at the productivity AI initially created.

Botsitting is really a context problem

When people complain about AI, they often blame the model. In reality, the model usually isn't the biggest problem.

The problem is context.

An AI rarely knows:

• who the customer is
• what happened in previous email conversations
• what's stored in your CRM or ERP
• your company procedures
• internal documentation
• who owns a particular process
• what happened during yesterday's meeting
• which spreadsheet is actually the final version

Without that information, AI produces generic answers. Employees then spend their time filling in the missing pieces. Every prompt starts from scratch. Every conversation requires explanation. Every correction costs time.

That's botsitting.

Data isn't the same as context

One insight from the report deserves much more attention. Many companies believe that connecting AI to enterprise data solves the problem.

It doesn't.

Connecting AI to your ERP or SharePoint gives it access to information. It doesn't tell the AI which document is authoritative, which process is the current one, what an internal abbreviation means or why everyone ignores the official procedure in favour of a practical workaround.

That's enterprise context.

And context is what makes information useful. Without it, employees remain responsible for translating the business into something AI can understand.

Cognitive offloading isn't the enemy

Botsitting and cognitive offloading are closely connected.

Cognitive offloading is a well-known concept in cognitive science. It describes our natural tendency to delegate mental work to external tools. We've always done it. We use GPS instead of memorizing roads. We write shopping lists instead of remembering everything. We save contacts instead of memorizing phone numbers. AI simply becomes another external cognitive tool. Used well, this is incredibly valuable. It allows us to spend less energy on repetitive tasks and more energy on creativity, judgment and decision making.

Researchers often describe two possible outcomes:

• The first is a cognitive dividend. AI removes routine work and gives us more time for higher-value thinking.
• The second is cognitive debt. AI becomes such an easy shortcut that people gradually stop exercising the skills they once developed themselves.

The objective isn't to avoid cognitive offloading. The objective is to offload the right work.

The next step after botsitting: botshitting

The report introduces another memorable term.

Botshitting.

It refers to shipping AI-generated work that hasn't been properly verified or that the employee couldn't confidently explain if someone asked. Unfortunately, this appears to be more common than most organizations would like. Many employees admit they have delivered AI-generated work they couldn't fully explain. Others acknowledge blaming AI for mistakes that were actually their own. This isn't necessarily because people are careless. It's often because botsitting becomes exhausting. When someone spends hours every day checking AI output, the temptation to accept "good enough" eventually becomes very real.

Human-in-the-loop is about more than governance

Human-in-the-loop AI is often discussed from a compliance perspective. That's certainly important. But there's another reason to keep humans involved.

Expertise.

The best AI users don't automate everything. They deliberately protect the parts of their work that require professional judgment while letting AI handle repetitive preparation, summarization and information gathering. They use AI as a thinking partner. Not as a thinking replacement. That's exactly where long-term value is created.

At ReplyFabric, we believe botsitting is mostly a context problem

When we started designing ReplyFabric, one principle quickly became obvious. Users shouldn't have to explain the same business context every single time they interact with AI.

The software should already understand:

• which mailbox is processing the email
• who the customer is
• previous conversations
• uploaded documentation
• company knowledge
• website content
• business rules
• who is responsible for each category
• which systems contain additional information

Instead of asking users to paste context into every prompt, ReplyFabric retrieves it automatically. Instead of asking employees to copy information between Outlook, SharePoint, Business Central, CRM systems and AI tools, ReplyFabric connects those sources before generating a draft. The human still reviews the result. The human still makes the final decision. But the human no longer spends the day babysitting the AI. That's a fundamentally different way of thinking about enterprise AI.

The future isn't better prompts

Prompt engineering has become an important skill over the past few years. But ideally, knowledge workers shouldn't have to become prompt engineers just to do their jobs. The future of enterprise AI isn't about writing longer prompts or discovering clever prompt tricks. It's about building AI systems that already understand enough about the business to produce useful results with minimal supervision. I believe the real breakthrough will come from invisible AI agents.

Not another chatbot. Not another browser tab. Not another application employees have to remember to open.

Instead, imagine an AI colleague that quietly does its work in the background. It reads incoming emails, retrieves the right business context, prepares a draft, categorizes requests and notifies the right colleague when action is required. By the time someone opens Outlook or Gmail, the work is already waiting for them. That's the kind of AI people actually want.

The AI you never have to wait for. The AI you don't have to prompt. The AI you don't have to babysit.

At ReplyFabric, that's exactly what we're building. Employees stay inside Outlook or Gmail because that's where they already work. They don't need to switch between AI tools, reload the same context over and over again or become prompt engineers. ReplyFabric works quietly in the background, continuously processing emails and preparing the next action before anyone asks for it. In other words, AI becomes a workhorse instead of another application.

Less prompting. Less context switching. Less botsitting. More meaningful work.

Because the best AI isn't the one you interact with the most. It's the one you hardly notice because it's already done the work.

---

Read the full report

Most of the research and numbers in this post come from the Work AI Index 2026, published by Glean's Work AI Institute. The full PDF is really rich in content: it covers where AI's time savings actually go, how botsitting turns into botshitting, what high AI achievers do differently and how organizations can close the gap between adoption and impact. Well worth the download.

👉 Download the Work AI Index 2026 (PDF)

👉 Other interesting study: Cognitive Offloading or Cognitive Overload? How AI Alters the Mental Architecture of Coping (PDF)