AI Agents complete guide

AI Agents in depth, in 65 chapters: creating, instructions, tools, knowledge, behavior, tester, inbox, workflows, analytics and usage, with every setting explained.

Oct 19, 2026

Video chapters

  1. 00:00 Introduction
  2. 00:38 The AI Agents screen Read this step →
  3. 01:13 Activate, test, metrics, duplicate and delete
  4. 01:43 Creating an agent: the three paths Read this step →
  5. 02:07 From a template Read this step →
  6. 02:40 Describe your business
  7. 03:20 From scratch
  8. 03:32 The editor: top bar and tabs
  9. 04:08 Identity: name, description and photo
  10. 04:35 Language and status Read this step →
  11. 05:04 Connections: how people reach it Read this step →
  12. 05:31 Connecting and disconnecting a channel
  13. 05:55 The AI in the channel settings Read this step →
  14. 06:32 Instructions: the editor Read this step →
  15. 07:11 How to write good instructions Read this step →
  16. 07:40 Using a template inside the editor
  17. 07:58 Version history
  18. 08:17 Optimize with AI Read this step →
  19. 08:41 The editor tool bar
  20. 08:59 Tools: what the agent understands
  21. 09:23 Close conversation, assign and distribute Read this step →
  22. 09:52 Transfer to another AI agent and dismiss Read this step →
  23. 10:16 Internal comments and email alerts
  24. 10:43 Catalog, store, orders and quotes
  25. 11:17 Services and appointments Read this step →
  26. 11:48 Tools by industry
  27. 12:15 Contact, tags and CRM Read this step →
  28. 12:42 Triggering workflows and sending files
  29. 13:00 Integrations, HTTP requests and calls
  30. 13:23 Custom actions: creating one Read this step →
  31. 13:50 Custom actions: parameters, variables and testing Read this step →
  32. 14:24 Knowledge: the agent's documents Read this step →
  33. 14:53 The knowledge base: URLs and files Read this step →
  34. 15:21 FAQs with files to send
  35. 15:39 Voice on calls Read this step →
  36. 16:05 Engine level
  37. 16:38 Behavior: temperature and reply length Read this step →
  38. 17:00 Escalation to a human: what to do and who to Read this step →
  39. 17:25 Failed tools and keywords
  40. 17:45 The handoff notice and never handing off
  41. 18:12 Contact allowlist
  42. 18:33 Supervision: human approval Read this step →
  43. 18:55 Confidence threshold and escalating after N replies Read this step →
  44. 19:19 Service hours Read this step →
  45. 19:43 Outside service hours
  46. 20:03 Customer inactivity Read this step →
  47. 20:35 Memory across conversations and summary on close Read this step →
  48. 21:00 Reply style and anti-spam Read this step →
  49. 21:23 Taking over on assignment and quality scoring
  50. 21:55 The tester: chat, channel and contact Read this step →
  51. 22:29 The tester: trace, real mode and HTTP actions Read this step →
  52. 22:59 The tester: attachments, voice notes and after hours
  53. 23:22 Activating the agent Read this step →
  54. 23:34 Real example: an appointment from the website
  55. 24:01 Real example: when they ask for a person
  56. 24:21 The AI in the inbox
  57. 24:47 Drafts to approve in the inbox
  58. 25:05 The conversation summary
  59. 25:22 Several agents working together Read this step →
  60. 25:44 AI agents in workflows
  61. 26:10 The agent's analytics Read this step →
  62. 26:33 Reply quality Read this step →
  63. 26:53 AI usage and credits Read this step →
  64. 27:25 Permissions and plans Read this step →
  65. 27:47 Best practices
  66. 28:09 Troubleshooting Read this step →
  67. 28:37 Wrap-up

This is the complete guide to AI Agents, in 65 chapters: every tab and every agent setting, explained one by one. Just getting started? Watch the short video in the «Start here» article first. The PDF manual has the same chapters with screenshots, steps and FAQs.

1. The AI Agents screen

Agents live in Settings, AI Agents. At the top you see three workspace numbers: how many agents are active out of the total, how many messages the AI answered this month and in how many conversations, and the average satisfaction survey score.

Below, you can search by name or description, filter between all, active or inactive, and switch between card view and list view, which shows the same thing in a table.

Each card shows the name, the description, where customers reach it, whether it's online, today's messages, this month's conversations, what percentage went to a person, and a chart of the last thirty days.

2. Activate, test, metrics, duplicate and delete

The switch on the card turns the agent on or pauses it instantly, without losing any of its settings.

The three-dot menu has: Test, which opens the agent's tester; Metrics, a quick summary of messages, satisfaction and recent conversations; Duplicate, which creates a switched-off copy so you can make changes without touching the original; and Delete.

And if you have questions, the How it works button brings you to this video, the short getting-started video and the PDF manual.

3. Creating an agent: the three paths

Click Create AI agent. Omnifox gives you three paths, and whichever you choose, you can change absolutely everything later in the editor.

From a template, with an agent already written for a specific job. Describe your business, where the AI builds the agent for you in seconds, free of charge. Or from scratch, a blank agent. They're all born switched off: nobody sees them until you turn them on.

4. From a template

There are more than one hundred fifty templates, sorted by category: sales, lead qualification, reception, support, collection, satisfaction surveys and more. You can filter by category or search by keyword.

When you pick one, you see the tools it turns on, what you need ready for it to work, for example your services loaded or your calendar set up, and a preview of its full instructions.

If it works for you, click Use this template and the agent is created with its instructions and tools. Today, though, we'll go back and try the fastest path.

5. Describe your business

In Describe your business you write, in your own words, what you do, who messages you, what details you always ask the customer for, and when you want it to hand over to a person. The more specific, the better.

Click Build my agent. In a few seconds the AI writes the full instructions and picks only the tools it needs. Generating it doesn't use your AI balance, and you get five tries a day.

Here's the proposal: the name, which you can change, the tools it turns on, what you need ready, and the instructions, which you can also edit before creating it. If you don't like it, ask for another version.

Click Create this agent and you go straight to the editor.

6. From scratch

The third option, Start blank, creates an agent with no instructions and no tools, so you can write it entirely your own way. It's ideal if you already know exactly what you want, or if you're going to copy another agent's instructions.

7. The editor: top bar and tabs

This is the editor. On the left, the tabs: Identity, Connections, Instructions, Tools, Knowledge, Voice on calls, Engine level and Behavior. On the right, the tester, which you can show or hide.

At the top you have Save and, while the agent is off, Save and activate, which saves and turns it on in one step. When there are unsaved changes, an orange notice appears; you can also save with Control S.

One important thing: the tester always replies with the saved version. Save before testing. And from here you can also open the agent's analytics and quality.

8. Identity: name, description and photo

In Identity you set the name, which is what the customer sees in your website chat and what other agents use to hand conversations to it. A short, clear name works best.

The description is just for your team: it shows on the card in the list and doesn't change what the AI replies.

With Change photo you upload a PNG or JPEG image of up to five megabytes, then crop and rotate it. The photo is saved instantly and shows in your website chat and in the inbox.

9. Language and status

In Language you choose the agent's main language: Spanish, English, Portuguese, French, Italian or German. By default, the agent detects the language each customer writes in and replies in that language.

If your business serves customers in just one language, turn on Always reply in the agent's language: that way an okay or an emoji won't make it switch languages.

And in Status, the Active switch turns it on or pauses it. When it's off it doesn't reply to anyone, but it keeps all its settings.

10. Connections: how people reach it

The Connections tab answers the first question: how do people reach this agent? An agent with no way in is written, but it doesn't serve anyone.

Here you see each channel it's connected to, and whether it replies to every message, only handles new conversations, or has the AI switched off on that channel. You also see the workflows and phone menus that route conversations to it.

And you can always assign it a conversation by hand from the inbox, or through the API.

11. Connecting and disconnecting a channel

Click Connect to a channel and pick one from the list. If that channel already replies with another agent, you'll get a warning: each channel has just one chat agent, so the previous one stops replying there.

Click Connect and you're done: it saves instantly and the agent will reply to every message on that channel. To remove it, use Disconnect. Let's also connect the website chat, which we'll use in the real examples.

12. The AI in the channel settings

You can see the same thing from the other side: in Channels, each channel has its Artificial Intelligence tab. There you choose the channel's chat agent and the response strategy.

Disabled: the AI doesn't reply and everything goes to your team. Always: the AI replies while the conversation is in its hands; if a person takes it over or it's transferred to a team, it stops replying. And New conversations only: the AI takes conversations that start on that channel, but not the ones that already existed.

Below, it reminds you of the agent's schedule: the channel strategy decides when the AI steps in, but outside the agent's own hours it won't reply.

13. Instructions: the editor

In Instructions you tell the agent what your business does, what to answer and when to hand over to a person. It's the most important part of the whole agent.

Type the dollar sign to insert variables, like the customer's name or the agent's name, and the exclamation mark to name a tool. They show up as colored tags. Note: naming a tool doesn't turn it on; you do that in the Tools tab.

Below, you see how many characters and tokens it has, and how many of your plan's credits each reply spends just by reading these instructions. That's why they should be clear and free of filler. The maximum is thirty thousand characters on the free trial and two hundred thousand with a plan.

14. How to write good instructions

Good instructions have five parts: who the agent is and who it works for; what it can do and with which tools; how it talks; what it must never do; and when it hands over to a person.

Be specific: instead of be helpful, write ask for the pet's name before booking. Say when to use each tool, and ask it to never make up prices, hours or links: it should look them up with its tools or hand over.

And always test: write the way a real customer would, check the trace and adjust. Instructions get better with use.

15. Using a template inside the editor

With Use template you open the same library, in whatever language you want. When you pick one, its instructions replace the current ones and its tools are turned on.

It asks you to confirm first. And don't worry: when you save, the previous version stays in the history.

16. Version history

Every time you save a change to the instructions, the tone, the temperature or the maximum length, the previous version is kept in History, with the date and who made it. The last fifty are kept.

With Restore this version you go back to an earlier one in one click. The current one is saved too, so you never lose anything.

17. Optimize with AI

Optimize with AI reviews your instructions and suggests an improved version, side by side with the original.

It points out what doesn't add up, for example two rules that contradict each other, the tools worth turning on, and the reasoning and main changes.

If you like it, click Apply optimized version and save. It doesn't use your balance and you get up to ten a day.

18. The editor tool bar

On the right side of the editor there's a bar with one icon per tool. A green dot means it's on, a gray one means it's off.

Click one to see when to use it and an example to paste into your instructions. With Enable and add example you turn it on and add the example in a single step.

19. Tools: what the agent understands

The Tools tab has two parts. First, what the agent understands from what it's sent: Image analysis, so it can see photos, screenshots or receipts; Voice note transcription; and Document reading for PDF and Word files.

All three come switched off because they use quite a few more credits than text. Turn them on if your customers often send audio or photos, and keep an eye on your usage for the first few days.

20. Close conversation, assign and distribute

Next, what the agent can do. Close conversation: the AI closes the chat when the customer doesn't need anything else; first it asks if there's anything more it can help with.

Assign to agent or team is the key tool for passing the chat to a person or a team. Turn it on whenever you want the AI to hand over.

When you turn it on, you choose how it's distributed within the team: with the team's own strategy, in a shared inbox, round robin, to the least busy member or to the first one available, and whether only to people who are online.

21. Transfer to another AI agent and dismiss

Transfer to another AI agent passes the conversation to another specialized agent, for example from reception to grooming. The other agent takes it and carries on the chat.

Dismiss: not a potential customer closes the conversation without bothering your team when it's clearly not a customer: spam, jokes, insults or someone looking for a job. It says goodbye once, leaves a note and, if they keep insisting, it can block them.

22. Internal comments and email alerts

Add internal comment leaves internal notes in the conversation that the customer doesn't see: for example, a summary of the case before handing it to your team.

Send internal email alert sends an email to a department in your company with the details the agent collected. You define the destinations: a short label, like sales, and who receives it. The agent can only choose from those destinations, never type an address, and there's a maximum number of emails per conversation.

23. Catalog, store, orders and quotes

If you sell products: Search the catalog to answer prices and descriptions, Send the store link, See the customer's cart, and Build the customer's cart during the chat.

Send product photo sends the photo with the name and price, instead of a link. Take orders closes the order inside the chat with the details you require. And Send quotes builds the quote with your real prices and gives the customer the link to accept it.

Prices always come from your product records, never from the model. Each of these tools needs its module: catalog, online store or orders.

24. Services and appointments

For services and bookings: Search services answers what you offer, how much it costs and how long it takes. Check appointment availability looks for free slots in your calendar and offers three at most.

Book appointment makes the booking when the customer confirms, and View, move or cancel the customer's appointment manages the one they already have. These are the ones our vet clinic uses.

And Book a meeting sends your Calendly or Cal.com link, if you have them connected in Integrations.

25. Tools by industry

There are tools for every industry. Real estate: search properties, send their photos by room and schedule viewings. Vehicles: search your inventory, send photos, simulate financing and book test drives.

Courses: search courses and enroll people in a group or on the waiting list. And medical: search doctors by specialty, location or insurance, and book with a doctor. Each one only works if you have that module set up.

26. Contact, tags and CRM

Update contact fields saves to the contact record what the customer shares in the chat: their name, their email, their city or your custom fields. Update tags adds or removes tags.

Create CRM opportunity opens an opportunity in your pipeline when the customer shows buying intent. You set the rules: if they ask the price, to this stage; if they ask for a quote, to another; with an optional amount. It never duplicates or moves stages backward.

27. Triggering workflows and sending files

Trigger workflows lets the AI fire your automations: for example, sending a template or adding the customer to a campaign.

And Send a knowledge base file sends the images, videos, PDFs and links you uploaded, instead of describing them in words. It turns on by itself when the agent has files to send.

28. Integrations, HTTP requests and calls

The Azur tools check balances and invoices, send invoice copies, live prices and stock, or create pro forma quotes, if you have that integration connected.

Make HTTP requests is the key to your custom actions: it has to be on for the agent to use them. And Manage calls lets it transfer, record, hang up or play messages during a voice call.

29. Custom actions: creating one

Below are the Custom actions: your agent can call your own APIs, for example to check an order's status in your system. Click New action.

The technical name, in lowercase with underscores, is the one you use in the instructions. The label is the visible name. And the description tells the model when to use it: that's the most important part.

Then the request: the method, the URL, which has to be https, and the headers, for example your API key.

30. Custom actions: parameters, variables and testing

Input parameters are the details the model has to fill in before calling, for example the order number. Each one has a name, a type, a description and whether it's required.

In the body you can insert variables: your parameters and data from the contact, the conversation, the channel or the agent. With the field to extract you return to the model only the part of the response that matters, and the timeout goes from one to sixty seconds.

Save the action and try it with Test: you see the real response from your system. Each action has its own switch to turn it on or pause it.

31. Knowledge: the agent's documents

In Knowledge you choose which documents the agent checks to answer. For each reply it looks for the passages closest to the customer's question and answers with that information.

Here you see the linked articles; one click opens them for editing. Below, the workspace documents that aren't linked yet, with a button to add them, and the trash icon to unlink.

Note: every new document is automatically linked to all agents in the workspace. If an agent shouldn't use it, unlink it here.

32. The knowledge base: URLs and files

The knowledge base belongs to the workspace and all agents share it. You can upload PDF, text or Word files of up to twenty-five megabytes.

With Add URL the agent learns from a page on your website: Omnifox reads it once and saves its text. If the page changes, add it again.

Each document shows its type and status: processing, ready or failed, and how many agents use it.

33. FAQs with files to send

The format that works best is frequently asked questions: a question and its answer, one by one. That way the agent finds exactly what it needs.

To each question you can attach files the agent should send, like a photo, a video or a PDF, and mark whether it's always sent or the agent decides when.

34. Voice on calls

The same agent works for chat and for calls. In Voice on calls you choose the voice for each language and listen to it with Play sample, plus the maximum length of each spoken reply: shorter means snappier.

And the transfer to an advisor: direct, warm with a summary whispered to the advisor before connecting, or let the agent decide per call; the default destination team, and the messages to the customer and to the advisor. Voice on calls depends on your plan.

35. Engine level

In Engine level you choose how much brainpower the agent uses. Swift Fox is fast and very affordable, more than enough for FAQs, orders and collecting details. Pro Fox reasons better in tangled conversations.

Max Fox is a leap in judgment, for agents that work on their own. And Deluxe Fox is the very best, and the one that uses the most credits. The number next to each one shows how much it spends compared to Swift.

You can switch whenever you want and it applies from the next reply. During the free trial you can use Swift and Pro; Max and Deluxe unlock with a plan.

36. Behavior: temperature and reply length

Behavior is where you decide how the agent acts. Let's start with temperature: close to zero, it always replies in a similar, more precise way; higher, more creative. For customer service, between zero point three and zero point seven.

And max tokens per reply caps the length of each message. One thousand twenty-four is plenty for a chat.

37. Escalation to a human: what to do and who to

In Escalation to a human you decide what happens when the AI has to let go of the conversation: transfer it to a person, to a team, close it, or take no action.

Choose the destination, for example the Front desk team, and how it's distributed within the team. If you don't choose a destination, the conversation stays unassigned, marked as unread and with high priority, but with no owner. And remember to turn on the Assign to agent or team tool.

38. Failed tools and keywords

After N failed calls hands over to a person if the tools fail several times in a row in the conversation. Zero turns it off.

And the escalation keywords: if the customer types, for example, human, person or agent, it transfers instantly, without going through the model. It ignores capital letters and accents, and matches the whole word.

39. The handoff notice and never handing off

When the AI hands the conversation over, it can notify the customer: with the platform notice in their language, with your own text, with a workflow, or with no notice. If the AI already wrote something at that moment, that gets sent and not both.

And AI always handles it does the opposite: it never hands off, even if the customer asks. Instead, it asks them to explain it another way. Only use it if there's no one serving customers behind it.

40. Contact allowlist

The contact allowlist is perfect for launching an agent: if you turn it on, the AI only replies to the phone numbers or emails you put here, one per line. Everyone else is left for your team.

Add your own number, test on the real channel, and when you're happy, turn it off to open it to everyone. Careful: with the list on and empty, the AI won't reply to anyone.

41. Supervision: human approval

Require human approval before sending turns every AI reply into a draft: a person approves it, edits it or rejects it from the inbox before it reaches the customer.

You choose how long a draft can wait, three hours by default, and what happens if nobody approves it in time: discard it and notify the team, hand the conversation to a person, or send it anyway. It's ideal for the first few weeks of a new agent.

42. Confidence threshold and escalating after N replies

With the confidence threshold, the model rates how sure it is of each reply, and if it falls below the value you set, for example zero point seven, it hands over to a person. It's a self-assessment, so use it as a safety net.

And Escalate to a human after N replies without resolving is a fixed lock: after that number of replies in the same stretch, the system hands the conversation to your team, without asking the model.

43. Service hours

In Service hours you decide when the AI replies on its own. Always: at any time, every day. Business hours: the same service hours as your workspace.

Custom: you choose the time zone and the windows, for example Monday to Friday from nine to seven, and Saturdays from nine to two. You can add several windows.

For our examples we'll leave it on Always.

44. Outside service hours

Outside service hours the AI doesn't call the model: it does what you set here. You can send the customer a message, like we're currently closed, we'll reply as soon as we reopen.

And Transfer to a human team hands the conversation to a team, for example whoever is on call. If you leave the message empty and the transfer off, the AI doesn't reply at all.

45. Customer inactivity

Customer inactivity is for when the customer stops replying. Turn on automatic follow-up and create a sequence of touches: each one is sent once the customer has been silent for that many minutes.

Each touch can be generated with AI, taking the conversation into account, or be a fixed message. You can add several touches.

And if they still don't reply, Transfer to a human passes the conversation to a team after the minutes you choose, with an optional message. If you leave it off, the AI keeps the conversation and replies when the customer comes back.

46. Memory across conversations and summary on close

Summarise the conversation when it closes generates an automatic summary your team sees in the inbox. Each summary is one call to the model, so turn it on if someone is going to read them.

And memory across conversations uses those summaries: when the same customer comes back, the agent remembers their previous conversations. You choose how many, up to ten, and how many days back. For it to have memory, the summary on close has to be on.

47. Reply style and anti-spam

In Reply style, the fixed tone always keeps the style of your instructions, and the adaptive tone mirrors how the customer writes: more formal, more casual, with emojis or without.

Anti-spam groups messages sent in a row: if the customer sends several short messages, the agent waits a few seconds and replies once with all the context. On WhatsApp, Instagram and Messenger the minimum wait is fifteen seconds.

48. Taking over on assignment and quality scoring

Reply when a conversation is assigned to me: when someone on your team assigns a chat to the agent, the agent replies right then to the last pending message, without waiting for the customer to write again. On WhatsApp, only within twenty-four hours of the customer's last message.

And Quality scoring: when a conversation closes, another AI rates it on relevance, safety, coherence and helpfulness, and flags the ones worth reviewing. It uses credits, and the scores show up on the Quality screen.

We set up what our vet clinic needs and click Save.

49. The tester: chat, channel and contact

Time to test. The tester uses the same engine as the real channel, with the saved version of the agent. At the top you choose the simulated channel, so the agent knows whether it's on WhatsApp, the web or Instagram.

Write the way a customer would. Under each reply you see the tokens, the cost and the time, and in amber the tools it used. Note: the tester does use your balance.

In the Contact tab you can load a real customer or type their details by hand, so the agent addresses them by name.

50. The tester: trace, real mode and HTTP actions

The Trace tab shows which tools it used, with what data and what they returned. It's the best way to understand why it replied what it replied.

By default, tools that change data are simulated: you see what would have happened, without creating appointments or assigning anything. Lookup tools, like searching services or checking times, do run.

With Real mode, and a real contact loaded, the tools actually run. And Real HTTP actions does the same with your custom actions. Only use them to test the complete flow.

51. The tester: attachments, voice notes and after hours

With the paperclip you attach an image, an audio file or a PDF, and with the microphone you record a voice note, to test what the agent understands. At the top you can copy the conversation, regenerate the last reply and reset.

And if you test outside the agent's hours, the tester shows you what would happen on the real channel, without spending anything, and offers to test as if it were within hours.

52. Activating the agent

When you're happy, click Save and activate. The agent goes online and starts replying on the connected channels. Let's go back to the list: there it is, active.

53. Real example: an appointment from the website

Let's see it with a real customer, in the Happy Paws website chat. The customer asks how much a bath for their dog costs.

The agent looks up the service and replies with the real price. The customer asks for an appointment, and the agent checks the calendar and offers free times.

The customer picks one and the agent books the appointment in the calendar. All in seconds, with no one on your team stepping in.

54. Real example: when they ask for a person

Now the customer writes that they want to talk to a person. That's a keyword, so the agent passes the conversation to the Front desk team instantly and lets the customer know.

In the inbox, the Front desk sees the complete conversation: what the customer wrote and what the AI replied, marked with the agent's name.

55. The AI in the inbox

At the top, AI paused appears, with the reason. From here your team works as usual, and the AI stays out of it.

If a person replies in a chat the AI is handling, they take control and the AI pauses. To give it back, use Resume AI, or assign the conversation to an AI agent from the Assign button.

In the inbox filters you can also see only one AI agent's conversations.

56. Drafts to approve in the inbox

With human approval on, the AI's replies arrive in the inbox as drafts, with the text it prepared and its confidence level.

Your team decides: Approve and send, Edit it before sending, or Reject it with a reason. The customer only receives what a person approved.

57. The conversation summary

With summary on close turned on, every closed conversation leaves an AI summary in the inbox, with the customer's sentiment.

And from the Summarize conversation menu you see the summary, what was left pending and the suggested actions, and you can pin it as an internal note.

58. Several agents working together

You can have several specialized agents. Happy Paws has a second agent, Happy Paws Grooming, an expert in baths and haircuts.

With the Transfer to another AI agent tool, reception passes it the conversation when the customer asks about grooming, and the trace shows the transfer. To chain agents, give them clear names and tell each one when to transfer.

59. AI agents in workflows

Workflows have a special trigger: The AI handed off the conversation, with the reason and the destination, so you can do something every time the AI passes on a chat, like alerting a supervisor or sending your own message.

And nodes for working with AI: Assign AI agent, so a workflow hands the conversation to an agent; Respond with AI, to generate a reply inside the workflow; and the AI classifier, which picks the path based on the customer's intent.

60. The agent's analytics

In View analytics you get the agent's numbers for the last seven, thirty or ninety days: conversations handled, how many were resolved without going to a person, tokens used and response time.

Below, charts of conversations per day, response time distribution and errors, and the table of recent conversations, with a button to open them in the inbox. You can filter by channel and by status.

61. Reply quality

In Quality you see the scored conversations: the score for relevance, safety, coherence and helpfulness, the reason it was scored, and the ones flagged for review.

You can filter by agent and period, mark each one as good or bad, or open Review to leave a comment. It's how you improve the instructions with real cases.

62. AI usage and credits

Everything the AI does uses tokens from your plan. In Settings, AI usage, you see the cycle meter, when it resets, and the usage per day.

It explains why a short reply uses so many tokens: the AI reads its instructions, the knowledge base and the history for every reply. That's why clear instructions without filler save you the most.

And you can see usage by feature, by agent, by channel and by conversation. If you reach the cap, the AI pauses until the next cycle, unless you turn on overage with your Omnifox balance.

63. Permissions and plans

Creating, editing, activating and deleting agents is a job for owners and admins; supervisors can view them. There's no limit on the number of agents.

What changes by plan: some tools, like custom actions, updating contacts and tags or handling calls, voice on calls, and the Max and Deluxe levels, which unlock when you activate a plan.

64. Best practices

A few best practices to wrap up. Start with one agent per job and just a few tools. Launch it with the allowlist or with human approval. Always turn on Assign to agent or team and choose a destination for escalation.

Load your FAQs into the knowledge base, check quality and analytics every week, and improve the instructions with real cases.

65. Troubleshooting

The agent isn't replying? Check that it's active, that it's connected to the channel with a strategy other than Disabled, that it's within its hours, that the allowlist is off and that the AI credit hasn't run out. And that no one on your team has taken over the conversation.

Making things up? Tell it in the instructions to use its tools and its knowledge base, and to hand over if it doesn't know. Not using a tool? Check that it's turned on and saved, and look at the trace in the tester.

Was this helpful?

Related articles

Message us on WhatsApp