← Back to search

An AI Colleague That Learns to Know Your Business

Werk Sneller met AI Daily · 2026-11-16 · 11 min
relevance 55 2044 words Episode page ↗ Audio ↗
Show full episode description
Een AI-assistent die je elke keer opnieuw moet uitleggen hoe je bedrijf werkt, kost je veel tijd. In deze aflevering kijk ik hoe je een AI-collega inricht die jouw bedrijf leert kennen, met OpenClaw en Hermes Agent als voorbeeld. Ik bespreek de vier onderdelen van zo'n systeem, de drie soorten bedrijfskennis, het verschil tussen instructie, geheugen en kennisbron, twee praktijkvoorbeelden, vijf taken voor kleine ondernemers, toegang in vier niveaus en de kosten. Invulkaart voor je eerste AI-collega: TAAK: wat moet de assistent voorbereiden? AANLEIDING: wanneer gaat hij aan het werk? INFORMATIE: welke documenten of gegevens heeft hij nodig? WERKWIJZE: welke stappen volgt hij? RESULTAAT: wat levert hij op en in welke vorm? BIJ TWIJFEL: wat vraagt hij als er informatie ontbreekt? GRENZEN: wat mag hij alleen na jouw goedkeuring doen? CONTROLE: hoe bepaal je of het resultaat goed genoeg is? Bronnen: OpenClaw, wat is het: https://docs.openclaw.ai/help/faq/what-is-openclaw OpenClaw, werkruimte en geheugen: https://docs.openclaw.ai/concepts/agent-workspace OpenClaw, beveiliging: https://docs.openclaw.ai/gateway/security Hermes Agent, documentatie: https://hermes-agent.nousresearch.com/docs Hermes Agent, verhalen uit de gebruikersgroep: https://hermes-agent.nousresearch.com/docs/user-stories Praktijkverhaal OpenClaw op een Mac mini: https://annuaire-agents-ia.com/en/cas-pratiques/retour-experience-assistante-ia-mac-mini/ Werk sneller met AI in je eigen bedrijf: https://werksnellermetai.nl Luister ook op Spotify: https://open.spotify.com/show/01DNvgXWIcsPtiGz7kmmeR En op Apple Podcasts: https://podcasts.apple.com/us/podcast/werk-sneller-met-ai-daily/id6825790507
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Entrepreneurs re-explain their business to a chatbot every day; an AI agent like OpenClaw or Hermes Agent learns business knowledge once and reuses it.
Benefits
  • Tell the agent about your company once; it retrieves that knowledge per task
  • Agent opens documents, compares data, and prepares drafts from folders, calendar, CRM
  • Scheduled tasks (Hermes) deliver a Monday overview automatically
  • Onboarding an AI colleague surfaces undocumented knowledge stuck in heads and old emails
  • Four-level access model keeps sending, paying, and data changes under human control
Use cases
  • Real estate entrepreneur runs OpenClaw on a Mac Mini to compare each contract with sales-system data and report differences
  • Small drainage company handles quotes, invoices, customer data, and inventory via Hermes over Telegram from the construction site or van
  • Quotation assistant: extract wishes, list 'still needed' info, use current price list, never promise a date, send nothing
  • Sales assistant finds quotes open over 7 days, checks for replies first, drafts friendly follow-ups
  • Monday overview of new requests, open quotes, appointments sorted into decide today / follow up this week / for information
KPIs / results
  • Four components: AI model, agent software, company knowledge, tools
  • Test the agent with twenty past scenarios before trusting it
  • Work Faster with Claude course: 24 lessons, 99 euros one-time, recouped in two hours at 50 euros/hour
  • Saving one hour/week on customer questions only pays if monthly maintenance stays low
Tools / build
  • OpenClaw on a Mac Mini for contract-vs-CRM comparison
  • Hermes Agent via Telegram for quotes, invoices, and inventory
  • Hermes scheduled tasks for a weekly Monday overview
  • Quotation, customer-service, onboarding, and sales assistant workflows
  • Work Faster with Claude online course
0:00 / 0:00
🌐 This transcript was automatically translated to English from the original.
This is Work Faster with AI Daily. Every workday, in a few minutes, what you as an entrepreneur need to know about AI and, most importantly, what you can do with it. I'm Erik Seibersma. Hey, and welcome to Work Faster with AI Daily. Today we're going deeper. It's about an AI colleague who learns to know your business. How does such a thing work? How do you give it your business knowledge? And what tasks do you give it as a small entrepreneur so that you really save time with it? Imagine that on Monday a new colleague starts. On the first day, he knows nothing. Not what you sell, not how you deal with customers, and not which appointments you have with suppliers. So what do you do? You explain. You show where everything stands. You tell them how you make quotes, what questions customers often ask, and when someone needs to present something to you. And after a few weeks, that colleague knows your business better and better. Precisely that idea you see now with a new generation of AI assistants. An assistant that remembers things, searches your documents, and, if you allow it, sets work ready for you. Today I'll take two as examples. OpenClaw and Hermes Agent. First, the problem. Most entrepreneurs use AI in this way now. You open a chat and ask for an email to a customer. Then you have to explain who that customer is, what your company does, what was discussed earlier, and what tone you want. And the next day you start again from the beginning. A large part of your time is spent in the same explanation over and over. An AI agent handles this differently. You tell it once. This is our company, these are our services, this is how we handle customer inquiries, these are the current terms and conditions, this you can prepare, and this you always tell me. For every new task, it retrieves that knowledge itself. And if you give it access to a folder, a calendar, or your customer system, it can also open documents, compare data, and prepare a draft. How does such a system work? Four components work together. The first component is the AI model, the part that understands language and formulates answers. This can be a model from a major provider or a model running on your own computer. The second component is the agent software, such as OpenClaw or Hermes. This manages how the work is organized, including tasks, memory, and tools. The third component is your company knowledge: documents, price lists, frequently asked questions, and work instructions. The fourth component are the tools: the mailbox, calendar, CRM, or messaging app that it has access to. Compare this to a new employee. The model is the cognitive ability. The software is the way of working. The company knowledge is the onboarding folder. And the tools are the systems where you give him a password. Briefly, here are the two main ones. OpenClaw is open source and is managed by an independent foundation. You run it on your own computer or server. You choose which AI model you use and you can reach it via, among others, WhatsApp, Telegram, and Slack. Hermes Agent comes from News Research, also open source, with memory, a reusable workflow that they call skills, and tasks that run on a fixed schedule. There is also a desktop app. Important to know: both ask for technical knowledge about the setup. You work with configuration files and keys from AI providers. Can't do that yourself? Then you need someone to set it up and maintain it for you. Then the most important part: your business knowledge. The greatest advantage lies in the knowledge already present in your company. Which customers fit your business? When do you offer discounts? What is standard in a quote? How do you handle a complaint? When must an employee first consult with someone else? Much of that knowledge resides in your head, in old emails, and in scattered documents. And as soon as you want to onboard an AI colleague, you realize how much of it is nowhere to be found. That is immediately a valuable exercise. I would divide that knowledge into three groups. The first is fixed company information. What do we do? For whom do we work? What tone do we use? For example, we are an installation company with six people. We work for private and small businesses in the region. And we promise no date before the schedule is checked. The second group is current information. Prices, delivery times, terms and conditions. You must be careful there. If your assistant finds three price lists and it is not clear which one is valid, he might take the wrong one. So agree in advance what the valid version is, who maintains it, and where it is located. And the third group are workflows. The steps that follow. For example, when a quotation request comes in, first check whether name, contact details, tasks, and location are known. List the missing information, create a draft response, and send nothing. Such a small, clear workflow is far more valuable than one large rule like "my quotations." Let's do some shameless promotion. Yes, we have something new. An online course. Work faster with Claude. You learn to let Claude handle your recurring tasks without having to program. From your first good assignment to skills, connectors, and scheduled tasks. And ultimately, you build your own small tool, such as a price calculator or an intake form. These are 24 short lessons. You see exactly how it works on the screen and do it immediately with your own work. What does it cost? 99 euros. One-time. If you calculate your own hour at 50 euros, you have already recouped your investment after two hours. And the first lesson is free. Go to werksnellermet.ai.nl under Academy. Okay, back to that AI colleague. What does such an agent remember? There is often confusion. People say, I have taught him everything. There are three different places. The instructions tell him how it works. The memory stores preferences and appointments. And a knowledge source is a document or system that he picks up when needed. Do not throw them all in one pile. Write our emails friendly and direct. That is an instruction. A price list belongs in a managed source. And when handling a complaint, first look up the order. That is a workflow. OpenClaw works, for example, with a work folder containing separate files: one for work instructions, one for personality, one for long-term memory, and a new file each day with what happened that day. You do not need to know the file names. The point is, do not put all your knowledge into one gigantic text. How does that look in practice? Two examples. An entrepreneur with a few companies, including a real estate firm, has set up OpenClaw on a Mac Mini for his administration. One of the tasks is to compare each contract with the data in the sales team's system. The agent does this and provides an overview of the differences. He also mentions that it took a lot of time to explain, test, and improve. A nice example. Before the agent reminds someone about a missing document, he must first check whether the deadline has passed and whether that document has already been received. Logical for a human, but for an agent it must be in the workflow. He also mentions the less appealing aspects. The costs for the AI model were significant. And due to a misspelled email address, a document once went to the wrong place. The second example comes from the Hermes user group. Someone with a small drainage company writes that he handles quotes, invoices, customer data, and inventory via Hermes. All via Telegram, either on the construction site or in his bus. That is recognizable for anyone who works outdoors frequently. A painter, an installer, a gardener, a trainer on-site. The information is generated during work, and the administration is done in the evenings. With such an assistant, you immediately speak in, create a draft for the quote to the Dorpstraat. Replace outdoor crane, materials, and two hours of labor. Confirm the schedule. He sets it up, states what is still missing, and you review it in the evening. What tasks can you assign to such an AI colleague? Five ideas. 1. The quotation assistant. Read the application, extract the wishes, list missing information under the heading "still needed," use only the current price list and the quotation template, do not promise a date, and send nothing. You give him a task, a source, and a boundary. 2. The customer service assistant. Start small. No access to your entire mailbox. An employee pastes a customer inquiry and receives a draft response, along with the conditions on which it is based. If that works well, then look further. 3. The onboarding assistant. A new employee has hundreds of small questions. Where is the form? Who approves a quotation? Describe the main tasks separately and let the assistant answer those questions. If multiple people get stuck at the same point, improve that instruction. 4. The sales assistant. Let him search for quotations that have been open for more than 7 days. First check whether there is already a response. Then draft a concept for a friendly follow-up email. That word first prevents you from bothering customers who have already responded for a long time. And 5. The Monday overview. Every Monday morning, a list of new applications, outstanding quotes, appointments, and customer inquiries that are waiting. Divided into decide today, follow up this week, and for information. Hermes has planned tasks that run on fixed schedules. And the access? I would build it in four levels. First, he reads only a few documents and gives answers. Then he creates concepts that someone reviews. Then he generates overviews from coupled systems. And only then may he execute a few actions that you have predefined. When sending, paying, and modifying customer data, you want to know very precisely what he can do and where someone is watching. Openclaw states in its own documentation that an agent with access to messages can be routed to other conversations and channels by default. Take this carefully into account. And be aware! If the software runs on a computer in your office, your information does not stay there automatically. If you use an AI model from an external provider, your information goes to that provider. Therefore, look at the entire path of your data. What does it cost? The software is free. You pay for the AI model, often per use. Sometimes there are costs for a computer or server, connections, and assistance for setup. Do not forget your own time for gathering knowledge and testing. Calculate it accordingly. If you assign someone two hours per week to the same customer questions and the assistant saves one hour there, you gain one hour per week. But if the system requires significant maintenance every month, you must deduct that. And if you spend ten minutes reviewing an answer that you could have written yourself in two minutes, you gain nothing. What can you do with this today? First choose the task, then choose the system. Pick one task that recurs every week. Gather ten good examples of past work and the accompanying documents, without including customer data that you do not need. Then answer five questions: What should it deliver? What information is it allowed to use? What steps does it follow? What does it do if something is missing? And what must it not do on its own? Test it afterward with twenty scenarios from the past. Is the price correct? Does it use the right terms? Does it notice what is missing? Does it promise anything you cannot fulfill? The show notes include a fill-in-the-blank card that you can use immediately. And my question to you? Which question do you get every week again, while the answer is already somewhere in your company? That is where I would start. Do you need help setting up such an AI colleague? From the first task to a working and tested version? That is what we do at Werk Sneller met AI. Visit worksnellermetAI.nl. Want you want to keep receiving practical tips like this? Subscribe to the Werksneller met AI Daily. Until tomorrow! You listened to the Werksneller met AI Daily. I am Erik Seibersma. Follow the show on Spotify or in your own podcast app. Then you won't miss an episode. And if you found something useful in this episode? Forward it to another entrepreneur who can also benefit from it. TV Gelderland 2021