With Guidelines, you can influence the general behavior of your AI Agents and Copilots. They apply to every action of the AI, which clearly distinguishes them from entries in the knowledge database:
  • The knowledge database contains (potentially a very large number of) information or knowledge that your AI Agent or Copilot can use to generate answers. This information is retrieved dynamically by the AI searching the knowledge database (exclusively) as needed.
  • Guidelines, on the other hand, represent fixed specifications that apply to every single action of the AI application.
Further down you will find how long a guideline can be, how many guidelines an AI Agent or Copilot takes, and how to tell whether the guidelines, actions, and brand and tone of a setup add up to too much. A setup is one AI Agent or Copilot with its settings; Connect AI lists them under "Setups".
You configure the following things directly in the respective Setup and not in the Guidelines:
  • Response length
  • Formality
  • Role
  • Tone and communication style
Please note: If you create multiple Guidelines, you should pay very close attention to ensure they do not contradict each other or exist in similar form multiple times, as this can significantly impair the quality of responses. Additionally, you should try to limit the number and content of Guidelines to the essentials.
In your account, you will find pre-configured Guidelines created by our AI team, which are not editable.
To add a new Guideline, click on the blue "+ Add" button:
Image without caption
Give the Guideline an internal name and prompt the instruction in the following field:
Image without caption
In the last step, assign the newly created Guideline to the desired AI Agents and Copilots:
Image without caption
To do so, switch to the "Not connected" tab under "Connected setups" and click "Connect" next to each setup you want, or "Connect all" to connect every setup at once.
Image without caption
The setups you connected now appear in the "Connected" tab. Then click "Save".
Image without caption
πŸ’‘
You can connect up to 20 guidelines to every AI Agent or Copilot. The pre-configured guidelines count too, because they are connected to every new setup automatically. If you do not need one of them, switch it off in the setup's "Guidelines" tab.
To reach a setup's "Guidelines" tab, open "Setups" > "AI Agent" or "Copilot" under Connect AI, click the setup's name and switch to the "Guidelines" tab. It lists every guideline in your account with a switch; switched on means connected to this setup. The "Guidelines" menu item under Connect AI, by contrast, is where you create and edit guidelines.
What the warning icon in the "Setups" column of the Guidelines overview means is explained under "How do I know if a setup has too many instructions?".

What happens when a setup already has 20 guidelines?

For such a setup, the "Connect" button is disabled. Hovering over it shows "This setup already has 20 guidelines. Remove one from the setup first, or pick another setup." "Connect all" skips full setups; they stay in the "Not connected" tab. If you switch on another guideline in a full setup's "Guidelines" tab, the error "Setup '<setup name>' has reached the limit of 20 guidelines. Please remove some guidelines before adding new ones." appears.
Some setups still have more than 20 guidelines from before this limit existed. Those guidelines keep working and stay editable. You can remove guidelines from such a setup, but you can only connect one again once it has fewer than 20. That includes a guideline you have just switched off: it cannot be switched back on until the setup has fewer than 20 guidelines.
Which guidelines a setup has is shown by the switches that are on in its "Guidelines" tab. The list shows no count, and it can run over several pages.
To remove a guideline from a setup, switch it off in the "Status" column of the setup's "Guidelines" tab. The change is saved immediately. Alternatively, in the guideline, go to the "Connected" tab under "Connected setups", click "Remove" next to the setup, and then click "Save".

How do I know if a setup has too many instructions?

Instructions are everything an AI Agent or Copilot has to keep in mind for every answer: the text of each connected guideline, the instructions of the connected actions, and brand and tone. A guideline's text is therefore only one part of the instructions. The more of it there is, the less reliably the AI follows each rule.
How much a setup carries is shown by the "Instruction load" meter. It sits under the title of every AI Agent's and Copilot's settings page and is visible on every tab; open the page through "Setups" under Connect AI, as described above. The meter measures the length of those texts and shows it as a percentage of the recommended amount. At or below 100% it stays blue and gray, with no note.
The legend under the bar shows where the instructions come from. Each part shows its percentage of the recommended amount, so the three parts add up to the total shown above the bar:
  • Guidelines: the text of every guideline connected to the setup, the pre-configured ones included
  • Actions: the "When to trigger this action" and "How to execute this action" fields of every action that is connected to the setup and not marked "Blocked". That tag appears next to the action's name, in the "Actions" menu under Connect AI and in the setup's "Actions" tab.
  • Brand and tone: the "Brand description" field, plus "Custom role instructions" and "Custom tonality instructions" if you entered your own under "Role" or "Tonality and communication style"
The recommended amount is not a fixed limit like the 20 guidelines. A setup with far fewer guidelines can still go above it when those guidelines are long.
Above 100%, the percentage turns orange and an orange bar marks the part beyond the recommended amount. Underneath the meter, the note "The AI follows each rule less reliably when it has a lot to keep in mind. Trim or merge guidelines to bring this down." appears. Clicking "guidelines" in that note opens the setup's "Guidelines" tab. The meter never blocks anything (unlike the limit of 20 guidelines): you can always save the setup.
"Guidelines" therefore means several things on this page: the menu item under Connect AI where you create and edit guidelines, the tab in a setup where you switch them on and off for that setup, the part of the legend, and the link in the note that leads to that tab.
You can also spot setups above 100% in the overviews:
  • In the overview under "Setups" > "AI Agent" or "Copilot", an "Optimize instructions" label appears next to the name of such a setup. Hover over it to see the percentage. If the setup still uses the previous AI model, the "Previous AI model" label appears in that spot instead. Such a setup can be above 100% too, so open it to check the meter.
  • In the Guidelines overview, the "Setups" column shows how many AI Agents and Copilots a guideline is connected to. If one of those setups is above 100%, a warning icon appears next to the count. Hover over it to see the name and percentage of each setup concerned.
To bring the percentage down, start with the largest part in the legend:
  • Shorten long guidelines or merge guidelines that cover similar ground. A guideline can be connected to several setups, and a change applies to all of them. Switch off what the setup does not need in its "Guidelines" tab. The pre-configured guidelines cannot be edited; switch off the ones you do not need.
  • In the "Actions" menu under Connect AI, shorten the "When to trigger this action" and "How to execute this action" fields of the connected actions. An action, too, can be connected to several setups, and a change applies to all of them. If only this setup does not need the action, switch it off in the "Status" column of the setup's "Actions" tab instead.
  • In the setup's "General" tab, shorten the "Brand description", "Custom role instructions" and "Custom tonality instructions" fields. This works the same way for AI Agents and Copilots.

Best Practices for Guidelines

Best Practices for Guidelines

  • Avoid contradictions: Make sure your guidelines don't contradict each other, as this can severely impact the quality of responses.
  • No redundancies: Ensure that guidelines are not present in similar form multiple times.
  • Keep it essential: Keep the number and content of guidelines as concise as possible - focus on the most important requirements.
  • Clear and precise formulation: Formulate guidelines unambiguously and comprehensibly so the AI can implement them reliably.
  • Don't use for setup settings: Configure response length, formality, role, and tone directly in the setup, not in the guidelines.
  • Observe character limit: Each guideline has a maximum of 2000 characters. Still, formulate concisely and keep each guideline to a single rule.

Examples of Guidelines

Support Bot Examples

Escalation for technical issues
If a customer describes a technical problem that cannot be resolved through simple troubleshooting steps, transfer the conversation to a human support agent. Inform the customer transparently about the transfer and provide a realistic wait time.
Handling negative emotions
When dealing with frustrated or angry customers, show empathy and acknowledge their situation. Apologize for any inconvenience caused and focus on solutions. Avoid defensive language and remain professional, even if the customer is rude.

Sales Bot Examples

Needs assessment before product suggestions
Before making product recommendations, ask targeted questions to understand the customer's specific needs and requirements. Only recommend products that actually match the identified needs. Avoid generic suggestions without context.
Transparency about prices and availability
Always provide complete pricing information including all relevant additional costs. Proactively inform about current availability and delivery times. If information is not available, admit this honestly and offer to obtain the information.

Accounting Bot Example

Data privacy for sensitive financial data
Never ask for complete credit card numbers, passwords, or other highly sensitive financial data in chat. For requests that require such data, point to secure alternative channels. Treat all shared financial data confidentially and don't mention it in subsequent responses.

General Examples

Prevent hallucinations
You are a support bot for end customers. Answer only based on the provided data and guidelines. If information is missing or unclear, honestly say you are not sure instead of guessing. Never invent facts, features, links, or offers. Point out conflicting information and ask clarifying questions when needed.