A factory usually needs light because people need to see what they are doing. Machines, bots and software systems do not depend on light in the same way. This is where the idea of the “dark factory” begins: an environment where operational work can continue in the background, even when humans are not manually touching every process. In today’s economy, there is a lot of discussion around autonomous businesses, AI agents and intelligent operating systems.
We would be careful to claim that the fully autonomous company is already ready for today. But as a long-term direction, it is clearly becoming important. As part of our internal R&D, supported by FFG Project Start, we are exploring how companies can build private knowledge centers, connect operational data sources, define AI agents, and validate which processes can be automated, which should stay suggestion-based, and which are not yet worth automating.
A dark factory in a digital context is not simply a set of automations. It is a new operational structure. Data from different sources is collected, stored, connected and made usable for AI agents. This does not only create faster workflows. It creates a growing company memory that can be improved continuously.
From AI tools to owned AI infrastructure
Many companies start with ChatGPT, Claude or similar tools. That makes sense, but it is not the same as owning AI infrastructure. Individual AI projects have limits: context windows, token costs, package restrictions, missing integrations and scattered data. The strategic step is to avoid storing company knowledge inside one tool and instead build a private knowledge center.
The knowledge center is the central knowledge layer. It can connect documents, campaigns, product data, tasks, customer communication, support cases, technical events and historical decisions. AI models then become interchangeable. The company remains independent from a single provider and can use different models depending on the task.
Why context is not permanently solved inside chat projects
It is a misunderstanding to believe that a project inside an AI tool automatically contains the full company context forever. The larger the data volume becomes, the more technical and economic limits appear. Context must be actively structured, stored, updated and made retrievable. That is exactly why a dedicated knowledge architecture is needed.
AI models become replaceable, data stays inside the company
The long-term advantage lies in separating knowledge from the model. The knowledge lives inside the company’s own system. The models access it. One process may use Claude today, ChatGPT tomorrow, Gemini or a locally hosted open-source model later. This creates less dependency on single platforms and more flexibility for cost, quality, privacy and scaling.
An agent is not simply a chatbot with a name. A useful agent has a role, task areas, access to specific knowledge sources, clear boundaries and a feedback system. An SEO agent works differently from a Google Ads agent, a customer service agent or an e-commerce agent. Through interaction, feedback and new data, the quality of agents can improve over time.
Automate, suggest or intentionally leave untouched
An important part of our R&D is the validation of real processes. Not every task should be automated. Some processes can be fully automated when risk and complexity are low. Others should only generate suggestions that a human reviews. Others are currently too sensitive, unclear or insufficiently standardized to automate in a meaningful way.
Local LLMs for routine, API models for complex work
Not every task requires the strongest external model. For recurring processes such as classification, extraction, tagging, simple summaries or data preparation, local LLMs can make sense. For strategy, complex analysis, creativity and high-quality decision support, stronger API models can be used. This makes cost, privacy and quality easier to control.
An autonomous business platform does not appear overnight. It grows through real use cases, clear data models, secure infrastructure and repeated learning. The realistic path starts with observation and structuring, moves toward suggestions and approvals, and can later become controlled automation for suitable processes.
Why e-commerce companies are especially interesting
E-commerce companies have many operational data sources: Shopify, product data, inventory, Google Ads, Search Console, Analytics, newsletters, support, returns, suppliers and content. This is exactly where value appears when data is connected and continuously evaluated. A knowledge center can better support marketing, product management, SEO, support and management decisions.
R&D supported by FFG Project Start
This topic is part of our internal research and development work. With support from FFG Project Start, we are validating which architecture, processes and agent models make sense for such systems. The focus is not to prematurely claim a finished product, but to test reality: Which data sources create value? Which agent roles are useful? Which tasks can be automated? Where is human control required?
Area | Single AI Tool | Private Knowledge Center |
Knowledge storage | Knowledge lives in chats or tool projects | Knowledge lives in controlled company infrastructure |
Model dependency | Depends on one provider and its limits | Different models can be used via API or locally |
Context | Limited by tokens, packages and project logic | Scales through retrieval, graph and data structure |
Data sources | Often pasted manually or fragmented | Sources can be connected systematically |
Agents | Mostly general chat assistants | Role-based agents with tasks and boundaries |
Feedback | Feedback often stays inside chat history | Feedback can improve agents and processes |
Automation | Individual workflows or prompts | Process logic with approvals and control points |
Security | Depends on external SaaS configuration | Can be planned privately, locally or tenant-separated |
Long-term value | Quick productivity | Operational company memory |
Which data sources really create value?
Not every integration is automatically useful. In the R&D process, we validate which sources actually improve decisions. These may include Shopify, Google Ads, Search Console, Analytics, CRM, email, support, documents, project management and internal databases. The key question is the concrete operational value.
Which agent roles make sense?
A good agent needs a clear role. An SEO agent, Google Ads agent, customer service agent, e-commerce agent or project manager agent each has different tasks, data sources and boundaries. The question is not only whether an agent is technically possible, but whether it creates real value in daily work.
Which tasks can be fully automated?
Some tasks are recurring, low-risk and rule-based. These are better candidates for automation. Examples may include data structuring, simple classification, monitoring, technical checks or internal summaries. Even then, logging, control and proper error handling are required.
Where are suggestions better than autopilot?
Many valuable tasks should not be executed automatically right away. AI can prepare recommendations, reports, drafts or warnings. Humans review and approve. This is especially relevant for marketing budgets, customer communication, strategic recommendations, content approvals and technical changes.
Which processes should not be automated yet?
An honest R&D approach also recognizes limits. Some processes are too sensitive, too complex, too unclear or too dependent on human context. In such cases, it may be better to collect knowledge, observe processes and reassess later instead of forcing automation too early.
How does feedback improve agents?
When humans work with agents, valuable feedback signals appear. Was a recommendation accepted, changed or rejected? Which information was missing? Which wording was useful? This feedback can improve prompts, roles, data models and workflows over time.
How does the system stay independent and secure?
The goal is not to become dependent on one AI model. The knowledge layer should remain independent and able to use different models. At the same time, security concepts are required for API tokens, data access, backups, tenant separation and approval processes. Without these foundations, AI infrastructure can quickly become risky.
We would be careful with that claim. Individual processes can already be automated or supported by AI very well today. A fully autonomous company is much more complex because data quality, responsibility, security, legal aspects, process logic and human judgment all matter. The realistic path is step-by-step validation instead of big promises.
Automation often solves individual tasks. A dark factory describes a larger operational structure where data, processes, AI agents, knowledge storage and control points work together. It is not only about doing work faster. It is about building a system that learns, recognizes relationships and prepares decisions better.
A knowledge center ensures that company knowledge does not remain scattered across chats, documents or individual tools. It consolidates data, relationships, history and context in a controlled structure. This allows different AI models to access the same knowledge without the company becoming fully dependent on one provider.
AI agents take on specialized roles. They can review SEO data, analyze campaigns, prepare support drafts, structure product data, observe project progress or create management reports. The key is that each agent has clear boundaries, suitable knowledge sources and a defined approval process.
Partly, but not completely. Local LLMs can be very interesting for recurring and structuring tasks, especially when privacy or cost matters. For complex strategy, high-quality writing, creative evaluation or advanced analysis, external API models may still be more useful. The combination of both approaches is often strongest.
E-commerce companies have many data sources and recurring decisions: products, inventory, campaigns, SEO, analytics, support, returns, content and customer behavior. When this information is connected, AI can help create better recommendations, detect issues earlier and prepare operational decisions in a more structured way.
AI can be powerful for analysis, structuring and recommendations. Still, mistakes, wrong priorities or missing context can happen. For budgets, customer communication, legal topics, technical deployments or sensitive data, humans should still review and decide. Good AI infrastructure does not replace control. It makes control more efficient.
The future of AI inside companies is not about solving everything inside one chat window. It is about building a secure, structured knowledge system that can work with different AI models. As part of our R&D, supported by FFG Project Start, we are exploring exactly this development: how companies can structure their data, processes and decisions better long term without becoming dependent on one specific AI tool.
The answer will not appear overnight. It will emerge through real use cases, technical validation, secure infrastructure and honest process analysis. Some tasks will be automated. Some will be prepared as suggestions. Some will intentionally remain human. This differentiation is the realistic path toward the next generation of AI-supported companies.