How to Describe an AI Agent: 7 Listing Items That Help Buyers Decide

What buyers decide from a listing page

Buyers looking for an AI agent read the listing page to decide three things.

• Does it fit our work: what can be delegated, and what stays with us
• Will it run in our environment: can we prepare the accounts and APIs it needs, and what does it cost
• Can we trust it: how data is handled, and who responds when there is a problem

To answer these three questions, write the seven items below. In this order, buyers can make their decision as they read down the page.

Writing the name and opening

Buyers look at the name and the first few lines in the list to decide whether to open the page. Put the target and the job in the name.

Dify's documentation recommends naming marketplace templates so the name alone shows where the template runs and what it does. The suggested form combines the channel or target with the core task. The same idea works for AI agent names.

• Example name: AI that registers contact form submissions in your CRM and drafts replies
• What to put in the opening: whose work and which job it takes on, and what work stays with people
• What to avoid: names such as 'all-purpose' or 'does anything' that hide what it is for

1. What it can do

Describe what it can do in the order of input, processing and output: what triggers it, what it receives, what it does, and what it leaves where.

Phrases like 'makes your operations more efficient' do not let buyers map the agent to their own work. Describe concrete behavior.

Also name the tools it supports. Whether email means Gmail or Outlook, and whether spreadsheets mean Google Sheets or Excel, decides whether the buyer can use it.

2. What it cannot do

Stating what it cannot do reduces misunderstandings after adoption. Buyers read the limits to estimate the work that will stay in-house.

• Out-of-scope work: for example, it does not read the contents of attachments, or it does not classify inquiries in languages other than Japanese and English
• Points where a person checks: actions such as sending, publishing or paying that proceed only after human approval
• Preconditions: input format, supported languages, the amount it can process at once and so on
• When things go wrong: how it notifies you when it stops on an error, and how much processing is left undone

Here is an example. It does not read the contents of attachments. Inquiries in languages other than Japanese and English are sorted into 'Other'. Replies stop at the draft stage, and a staff member sends them. If processing stops partway, it notifies the staff member by email and leaves the stopped inquiry unprocessed.

AI output is not always correct. When you state where a person checks, the buyer can assign someone in-house to do the checking.

3. Required accounts and APIs

AI agents often run connected to the buyer's own accounts. List what needs to be prepared, together with who pays for it.

Set things up so API keys and passwords are never exchanged directly between seller and buyer. Write steps for buyers to issue and configure them themselves.

The operating environment is also information required when selling by mail order. The guide to the Act on Specified Commercial Transactions (Japan) from Japan's Consumer Affairs Agency lists the operating environment as a required disclosure in mail-order advertising for software transactions.

4. Data handling

Many buyers handle customer and employee information. When you describe the data flow, buyers can move ahead with their internal review.

• Data it reads: what it reads, such as email bodies or specific columns in a sheet
• Where data is stored: on the seller's server, or processed only in the buyer's environment
• Where data is sent: which AI model provider receives the data
• Use for training: whether the data sent is set to be used for training AI models
• Retention and deletion: how long data is kept, and how it is deleted on cancellation

Here is an example. It reads only the body of the contact form and the sender's email address. Nothing is stored on the seller's server. The body is sent to the OpenAI API for classification. Retention and use for training follow the recipient's terms and settings, and these are described in the setup instructions.

In an alert issued on June 2, 2023, Japan's Personal Information Protection Commission (PPC) set out points for businesses to note when entering prompts containing personal information into generative AI services. These include confirming that the use is within the stated purpose, and confirming that the provider does not use the data for machine learning. State where data is sent and whether it is used for training so buyers can make these checks.

5. Setup steps and deliverables

State what you deliver and what the buyer must do to start using it. Make the description of deliverables detailed enough that the buyer can get started without asking anyone.

• Deliverables: an account per buyer, a time-limited link, workflow files and so on
• Buyer's tasks: issuing accounts, granting access to connections, initial settings
• Steps to get started: write them in order with numbers
• How to confirm it works: the first action to try and the result when it works correctly

Do not deliver a URL that anyone can open. Use a form that limits access to the buyer, such as an account issued per buyer or a time-limited link.

6. Support scope

If you do not state the scope of support, the buyer's expectations and the seller's workload drift apart. Write the following.

• What you handle: setup consultation during adoption, investigating bugs and so on
• What you do not handle: changes to the buyer's internal systems, adding out-of-scope work and so on
• Support hours: weekdays only or not, and a rough response time
• Updates: whether the seller fixes the agent when connected services or AI models change, and what you announce when you do

7. Cost breakdown

State what the monthly fee includes and what costs extra. If AI model usage fees or connected tool plans cost extra and you do not say so, buyers cannot compare costs.

AI model usage fees can change with the volume processed. If you cannot state an exact amount, state what makes it go up.

Wording to avoid

Do not use wording that could mislead buyers' decisions.

• Numbers without a track record: time savings or conversion rates you have not measured
• Absolute claims: 'guaranteed', '100%', 'fully automatic' and the like
• Wording that suggests you are a company: adding a corporate designation to your name when you are not incorporated, for example
• URLs anyone can open, or the API keys themselves

Checking against the AI Guidelines for Business

The AI Guidelines for Business (version 1.2) from Japan's Ministry of Internal Affairs and Communications (MIC) and Ministry of Economy, Trade and Industry (METI) give examples of information that businesses providing AI services should share with users and others. They say to communicate, in plain and easy-to-understand form, items such as the following.

• The fact that AI is used, the scope of its use, and appropriate and inappropriate ways to use it
• Safety information such as technical characteristics, foreseeable risks and how they are mitigated
• When the AI system is updated, what changed and why

The same guidelines also mention telling users what to watch for when entering personal information. The seven items in this article overlap with these points.

What to check when you finish writing

• What it can do is written in the order of input, processing and output
• What it cannot do, and where a person checks, are stated
• Required accounts and APIs, and who pays for them, are stated
• The scope of data it reads, where data is stored, where it is sent and use for training are stated
• Deliverables and the steps to get started are stated
• What support covers and does not cover is stated
• Costs included in the monthly fee and costs charged separately are stated
• There are no unmeasured numbers or absolute claims

List on Employee Store

On Employee Store, Employee Store Operations reviews what you submit through the listing form, and approved listings are published. Any format is accepted, including n8n or Dify workflows and agents you developed yourself. GPTs are also an accepted listing format, but OpenAI has announced that it plans to end custom GPTs on December 11, 2026.

There is no listing fee and no upfront cost. The commission is 20% of the deal amount and is charged only when a deal closes. How to compare sales channels is covered in How to sell your own AI agent, and listing terms in the seller guide.