AI agents in Markonex: what your company can automate on the B2B platform

What are Markonex Agents?

An AI agent is an assistant that does not wait for you to type in a chat: it receives your instructions once, written in plain language, and then works on its own, at the time you set or when you press "Run now". In Markonex, agents are built on Anthropic's Claude Managed Agents and are created in a separate web app, Markonex Agents, which you open from your company account without a new password.

The agent belongs to the company, not to a single employee, and sees Markonex exactly like a company user: the requests for quotation on the marketplace, the catalogue, offers, negotiations, agreements, contracts, messages, the CRM and received invoices, only as far as the permissions you grant. It does not browse the internet, run programs or open files: its only tool is the Markonex MCP server.

AI agents are available to companies verified by the platform, meaning companies checked with ANAF by their CUI and then validated by the Markonex team based on documents.

What can an AI agent automate directly on the platform?

Any repetitive task you do today by opening the same pages every morning can become an agent's job. A few examples of tasks companies can give agents in Markonex:

• Daily monitoring of new RFQs that match what you sell, with a report of the best opportunities and of those expiring soon.
• Preparing offers at catalogue prices for matching RFQs, so a colleague can approve or reject them with one click.
• Following open negotiations: what changed in the last round, where you are expected to reply and which offers expire in the next few days.
• Preparing counter-offers within the limits you write, for example "never go below the catalogue minimum price and never promise delivery in under 5 days".
• Comparing the offers received on a multi-line RFQ, line by line, and proposing the right supplier for each line.
• Turning a product list written as text into a draft multi-line RFQ that you review before publishing.
• Summarising agreements and contracts in progress: what waits for your confirmation, what waits for the partner and what is completed.
• A weekly report for management: active RFQs, offers received, agreements concluded and invoices received on agreements.
• Keeping the CRM up to date after transactions: notes on the partner record, tags, relationship type and price comparison across partners.
• Creating following filters for your categories, so you are notified when a matching RFQ appears.

What does an agent at work actually look like?

The example below is illustrative: a building materials distributor writes the agent's instructions, and the agent runs every working day in the morning. The report appears on the agent's "Runs" page, and the prepared offer waits for a colleague's approval.

1. Agent instructions

   Every working day, look for new RFQs from the last 24 hours for building materials in Cluj, Alba and Mureș counties. List the ones that best match our catalogue and prepare offers only for requests where we have every product, at the catalogue list price.

2. The agent

   I found 6 new matching RFQs. The best 3: a 4-line list for a site in Turda (we have every product), a request for tile adhesive in Cluj-Napoca (we have the product, it expires in 2 days) and a request for AAC blocks in Alba Iulia (we lack the requested thickness). I prepared an offer for the Turda request at catalogue prices.

3. Markonex Agents - approval

   The agent wants to run a prepared action: send the offer on the RFQ "Turda site materials", 4 lines, at catalogue prices. Approve / Reject

Who approves what the agent does?

The agent can read, and it can do only three things on its own that have no commercial effect: save draft RFQs, create following filters and mark notifications as read. Every other action, from publishing an RFQ and sending an offer to confirming an agreement, adding a catalogue product, sending a message or changing the CRM, is only prepared.

The prepared action appears in Markonex Agents as a card "The agent wants to run a prepared action", showing everything that is about to happen. Colleagues with the right to manage agents are notified in the app, by push and by email and press "Approve" or "Reject". The approval is valid for 24 hours from preparation and only while the agent's run is still active; after that the card shows "Expired" and nothing is executed.

An approved action runs in the name of the person who approved it, with their Markonex rights, and only within the permissions granted to the agent. So an agent can never do something the approving person is not allowed to do.

Step 1: turn on AI assistants for your company

The company administrator goes to My company → AI assistants → Agents in Markonex, ticks the information notice and turns Agents on. The MCP server must also be enabled for the company, because agents work through it. This is done in the web app; the mobile app only has the on/off switch.

Step 2: add your company's Anthropic API key

The "Open Markonex Agents" button takes you to the agents app with single sign-on from your Markonex account. Under "API keys" you add the key of your company's Anthropic account. The key is stored encrypted and only its last 4 characters are shown.

AI model usage is paid directly to Anthropic, from the prepaid credits of your Anthropic account. Markonex does not bill or resell this usage; the app only shows a cost estimate.

Step 3: create the agent with the "New agent" wizard

The wizard takes you through a few simple steps: you choose the API key, the model and the effort level, then write the agent's name, description and instructions, in Romanian or English, just as you would explain the job to a new colleague.

Under "Permissions" you choose what the agent can see and prepare: reading RFQs, offers or the CRM, and writing only if you tick it explicitly. An agent that only reports can have read permissions only.

Step 4: set the schedule and the budget limits

Under "Schedules" you choose when the agent runs, for example every working day at 8:00 or on Mondays at 9:00, Romanian time. An agent can run at most once an hour, and a run may start a few minutes late. You can always start it manually with "Run now".

Under "Limits" you set the monthly budget, the budget per run and the maximum number of runs per day, so there are no cost surprises. Finally, "Summary" shows everything you chose; every setting can be changed later.

Step 5: test it in the test environment, then switch it on

Before letting the agent work on real data, you can run it in your company's test environment: a separate copy with your test company and up to two test partner companies, with no emails, no ANAF and no e-Factura. There, test agents confirm their own actions, so you can see the whole flow quickly.

When you are happy with the result, you switch the agent on for the real account. On the agent's page you follow "Runs" (what it did on each run), "Chat" (you can talk to it directly), "Schedules", "Costs" and "Settings".

How much does an AI agent cost and who pays?

The agent uses the Claude model through your company's Anthropic account, so you pay the model cost directly to Anthropic, from prepaid credits. Markonex does not add a charge on model usage and does not resell it.

The cost depends on how often the agent runs, how much it has to read and which model you choose. Monthly and per-run budget limits, plus the maximum number of runs per day, stop the agent before it exceeds what you set, and the "Costs" page shows the estimate per run and per month.

How safe is an AI agent in Markonex?

The agent is tied to a company user, whose identity the administrator can change with "Change identity". Everything it does appears in the company activity log as "Name (via Anthropic agent #id)", and that user sees it as "You (via Anthropic agent #id)".

• It works only through the Markonex MCP server: no access to the internet, files or other systems.
• It sees and prepares only what both the agent's permissions and the user's Markonex rights allow.
• It cannot issue, edit or credit invoices, sends nothing to e-Factura, makes no payments and does not administer the company, its users or permissions.
• It cannot change passwords, emails or phone numbers and cannot manage tokens or the MCP and Agents settings.
• The Anthropic API key is stored encrypted; turning Agents off for the company immediately revokes all their access.

What does a Markonex agent not do?

An agent has no memory between runs: each run starts from your instructions and the current platform data, which is why clear instructions matter so much. Agents use Anthropic's Claude models only; agents on OpenAI models are not available.

The agent starts on its schedule or manually, not automatically on an event, and executes nothing with a commercial effect without a person's approval. The Markonex Agents app is used in the browser, on the web.

AI agents or MCP: which should you choose?

With MCP you talk to ChatGPT or Claude yourself, in a conversation, and the assistant works in Markonex while you chat: it suits one-off questions and actions, such as "create an RFQ from this list". An AI agent works on its own, on a schedule, without you opening the chat: it suits tasks that repeat daily or weekly.

Many companies use both: the agent prepares the report and the offers in the morning, and the person in charge reviews them and asks follow-up questions from ChatGPT or Claude. Connection details are on the MCP page, and complete examples are in the case studies.