How an AI agent prepares a manufacturer's offers while a person only approves them

Who is the company and why do requests go unquoted?

Take a small metalwork manufacturer: shelving, gratings, brackets and made-to-order parts. The company has a catalogue of a few dozen products on Markonex, with a price range, currency and categories, and the owner writes the offers personally, between production and deliveries.

Many requests for quotation match the catalogue, but writing each offer from scratch takes time. Small requests, which could bring new customers, end up last on the list.

How was an offer made before the agent?

The owner read the request, looked for the right product in the catalogue, worked out the price for the requested quantity, added the delivery time and sent the offer. For a request with a product list, everything was repeated on every line.

• Similar offers were rewritten every time.
• Some requests expired before there was time for them.
• It was not clear which price rules the company applies to large quantities.

What do you need to start an agent like this?

Markonex Agents are Claude agents, from Anthropic, that a company creates in the Markonex Agents web app. The agent works for the company, not for a single employee, and uses the company's own Anthropic API key: model usage is paid directly to Anthropic, from the company's prepaid credits, not to Markonex. The cost shown in the app is an estimate.

• The company is verified by the platform (the "Verified by the platform" badge).
• The administrator turns on Agents and MCP under My company → AI assistants → Agents and ticks the information notice.
• In Markonex Agents, under "API keys", the company adds its Anthropic key; only the last 4 characters remain visible.
• The "New agent" wizard has these steps: API key, Model, Instructions, Permissions, Schedules, Limits (monthly budget, per-run budget, maximum runs per day) and Summary.
• Permissions in plain language: the agent can read requests for quotation, the catalogue and the company's offers, and can prepare initial offers and line offers on a request. Preparing is not sending: every offer waits for approval.

How do you try it safely, before using real data?

Every company has a separate test environment: a "(TEST)" company and up to two test partner companies, with their own data, no emails, no push notifications and no ANAF or e-Factura. There you let the agent prepare offers on the test partner's sample requests, to check that it follows the price rules before it works with real customers.

In the test environment the agent confirms its own actions, so you can see the whole flow quickly. In the real account the same actions wait for a person to approve them. The test environment can be reset or deleted at any time.

Step 1: put the price rules in the instructions

The agent does exactly what you write. The clearer the rules, the fewer corrections the prepared offers need.

1. Agent instructions

   On every run, look for active requests for quotation that match our shelving and metalwork catalogue and that we have not quoted yet. For each one, choose the right catalogue product and prepare an offer at the catalogue minimum price, with delivery in 10 working days. For quantities above 200 pieces, never go below the catalogue price and note in the offer that the price is open to discussion. If you are not sure the product matches, do not prepare the offer and explain why in the report.

Step 2: schedule the agent twice a day

The schedule can be, for example, 9:00 and 14:00 on working days. Under "Limits" you set the monthly budget, the per-run budget and the maximum number of runs per day.

Step 3: the agent prepares the offers and stops

The agent finds the requests, chooses the products and prepares the offers. At that point the run pauses and waits: company users with the agents-management permission get a notification in the app, by push or by email.

1. Markonex Agents - approval

   The agent wants to run a prepared action: initial offer on the request "Metal warehouse shelving, 120 pcs", product "5-shelf metal rack", unit price 310 lei excluding VAT, delivery in 10 working days. Approve / Reject

Step 4: you approve or reject, and Markonex runs only what you approved

The owner opens Markonex Agents, checks the offer and clicks "Approve" or "Reject". An approved offer is sent in their name, with their rights, and appears in the activity log as "Name (via Anthropic agent #id)".

The approval is valid for 24 hours from preparation and only while the run is still active. If nobody answers in time, the action shows as "Expired" and nothing is executed.

What does a small manufacturer gain from an agent that prepares offers?

The owner no longer writes offers from scratch but reviews them. Small requests get an answer too, and the price rules are applied the same way every time because they are written in the instructions. As an illustrative estimate, an offer that took a quarter of an hour to write can be reviewed in a minute or two; the real time depends on how complete the catalogue products are.

What limits and safeguards does the agent have?

The agent has a single tool: the Markonex MCP server. It does not browse the web, open files or run code. It has no memory between runs: each run starts from the data in the platform and the instructions the company wrote. The agent does not invent prices outside the rules it was given and cannot send any offer without approval; at most it prepares one.

• It works only with the permissions ticked when it was created, within the rights of the user whose identity it uses.
• Schedules run at most once an hour, on Romanian time; a run may start a few minutes late.
• Prepared commercial actions wait for approval by a user with the agents-management permission, who is notified in the app, by push or by email; the approval is valid for 24 hours and only while the run is still active.
• An approved action runs in the name of the person who approved it, and the activity log shows "Name (via Anthropic agent #id)".
• The agent cannot issue or cancel invoices, sends nothing to e-Factura, makes no payments and cannot administer the company or its users.
• If the administrator turns Agents off for the company, all agent access is revoked at once.