Chapter 13
Chat AI
Chat is the organization's multi-model AI assistant: pick GPT, Claude or Gemini or leave it on Auto; attach documents and images; search the web; and in Agent mode read and act on Gmail, Facebook or any other connected account, with an approval card for every outgoing action.
Fourteen videos in three groups: chat basics (the screen, models, files, web search); working on connected accounts (Ask, Plan and Agent modes; sending email; connecting a Facebook Page; publishing a post; bulk-replying to comments); then image and video generation, Memories, managing conversations and credits.
14 videos · 29:31
Recognise the conversation rail, the credit bar, the Model and Memories tabs, the composer and the mode row, and the way back to the portal.
Use it when: The first time you open Chat, or to learn what each control does before typing.
Steps
- 1Chat is the organization's multi-model AI assistant: Q&A, reading documents, web search, image and video generation, and acting on your connected Gmail or Facebook.
- 2On the left are all your conversations, grouped by date, with pinning and search. Conversations are private: colleagues cannot read them.
- 3Top bar: the organization's remaining credits, a Top up button, the Model tab for the catalog, Memories for what the assistant remembers about you, and the way back to the portal.
- 4The composer has a plus for attaching files or adding a connected account, a globe for web search, and the model button, which defaults to Auto.
- 5Type a question and press Enter. The answer streams in, with the token count and the real cost of that turn. The conversation is saved and named automatically.
- 6The row under the composer is the mode: Ask only answers; Plan reads connected accounts and proposes; only Agent may act. Details in a later video.
- 7Back to portal returns you to the admin console.
Good to know
- Conversations are private to each member; owners see usage, never content.
Outcome: Recognise the conversation rail, the credit bar, the Model and Memories tabs, the composer and the mode row, and the way back to the portal.
Open the model picker by group and filter, read prices, switch models mid-conversation, use the Model catalog and toggle models for the organization.
Use it when: When a question needs a stronger model, or you want to save credits on simple, high-volume work.
Steps
- 1Auto lets StepX pick the best model for each message and charges the monthly credit pool. It is the default and fits most questions.
- 2The Text group has GPT, Claude, Gemini, DeepSeek and Grok. Each row shows the price per million tokens and its abilities: vision, tools, extended reasoning.
- 3With many models, search by name or filter by ability or provider. Filters do not apply to Auto.
- 4Pick a model and ask. For the same question an expensive model answers in more depth and costs more; a cheap one like DeepSeek suits simple, high-volume work.
- 5The Model tab opens the full catalog: compare benchmarks, context size and price, sort by cheapest. Click Use this model to switch right inside the conversation.
- 6Owners click Manage models to decide which models members may use. Prices differ by tens of times and video is billed per second, so switch off what you do not need.
Prompt: Write 3 sentences introducing StepX to new customers
Gemini 3.6 Flash: fuller answer, higher cost
DeepSeek V4 Pro: shorter answer, several times cheaper
Good to know
- Auto is the safe default: StepX picks the model per message and charges the monthly pool.
- Owners can switch off expensive models under Manage models so members cannot pick them by mistake.
Outcome: Open the model picker by group and filter, read prices, switch models mid-conversation, use the Model catalog and toggle models for the organization.
Attach a PDF and ask about it, paste an image to ask, know that long pasted text becomes a file and what happens when a model cannot see images.
Use it when: When you need to ask about a document, a screenshot or a long text you just copied.
Steps
- 1Click the plus, Attach file and pick a PDF. It uploads as a chip on the composer; each message takes up to five files of ten megabytes each.
- 2Ask about the document as you would a colleague who has read it. The assistant reads the full PDF and answers from its real content; click the chip to preview or download.
- 3Images paste straight from the clipboard: take a screenshot, then Cmd V. A model with vision describes and answers about the picture.
- 4If the selected model cannot see images, StepX says so and blocks the send. Switch to a model with the Vision icon, or remove the picture.
- 5Paste a very long text and StepX wraps it into a Pasted text file, keeping the composer tidy while the model still reads all of it. Ask as with any file.
Good to know
- Each message takes up to 5 files of 10 MB each; images and PDFs.
- A scanned PDF without a text layer is read with OCR, about 2 USD per 1,000 pages, and StepX says so first.
Outcome: Attach a PDF and ask about it, paste an image to ask, know that long pasted text becomes a file and what happens when a model cannot see images.
Turn on web search for one message, read an answer with inline source chips and the Sources list, understand the cost.
Use it when: When a question needs fresh information: events, prices, this week's news.
Steps
- 1A model does not know today's news on its own. Click the globe to search the web for this message; each search adds about two US cents.
- 2The answer draws on real web pages and cites them inline. Click a chip to open the source, or use the arrows to move between sources.
- 3The Sources line lists every page the assistant read for this answer, including ones it did not quote directly.
- 4The globe switches itself off after each send so you never pay by accident. Some models search on their own when needed; the globe forces a search.
Good to know
- The globe switches itself off after each send; each search adds about 0.02 USD.
- Some models search on their own when needed; the globe forces a search for that message.
Outcome: Turn on web search for one message, read an answer with inline source chips and the Sources list, understand the cost.
Tell the three modes and Agent's three permission levels apart, arm an account for the conversation and read the account chips.
Use it when: Before letting the assistant touch Gmail, Facebook or any other organization account.
Steps
- 1Three modes. Ask: questions only, no account touched. Plan: reads real data and proposes a plan, does nothing yet. Agent: may actually act.
- 2Choose Agent and its permission appears: Restricted reads only; Ask first has you approve every create, edit or send; Allowed acts without asking. The default is Ask first.
- 3Plus, Add integration: every account the organization connected on the Integrations page is here, grouped by app. Unconnected apps sit at the bottom and can be connected right here.
- 4Click an account to arm it for the conversation, up to three at once. Armed accounts show as chips on the composer; click the cross to stop using one.
- 5Arming an account while in Ask switches the mode to Agent, since arming means you want it used. Plan stays Plan.
- 6Armed accounts and the mode are remembered for this organization on this device, so you do not re-pick them every conversation. Each send still records what it used.
Good to know
- Up to 3 accounts armed at once; armed accounts and the mode are remembered per organization on this device.
- Agent defaults to Ask first: every create, edit or send waits for your approval.
Outcome: Tell the three modes and Agent's three permission levels apart, arm an account for the conversation and read the account chips.
Have the assistant read new mail, draft and send a real email after approving the confirmation card, read the tool timeline and check the mail in Gmail.
Use it when: When you want an inbox summary or to send an email without leaving StepX.
Steps
- 1With Gmail armed in Agent mode, ask about the inbox directly. The assistant uses real Gmail tools instead of guessing.
- 2Ask first means the first read on each account also asks you once. Approve, and the remaining read-only steps of this turn run on their own.
- 3The block above the answer is the record of what actually ran, taken from the system rather than from the model's words. Expand it to see each step.
- 4Ask it to draft and send an email. The assistant writes the message, then stops at the send step, because sending changes data.
- 5The approval card shows the exact recipient, subject and body. Check it and Approve, or Reject so the assistant revises. Left alone, it rejects itself.
- 6The email is out. Ask to send it again and the assistant remembers it already did, asking with a choice card instead of sending a duplicate.
- 7In the recipient's inbox the email is there, with the subject, sender and body exactly as the approval card showed.
Any new email in the last 24 hours? Summarise each briefly.
Draft an email to tue.nguyen@tecurix.com titled 'StepX demo schedule next week', proposing Tuesday and Thursday at 10:00, signed StepX Team. Send it.
Good to know
- With Ask first, the first read on each account asks you once; remaining read-only steps run on their own.
- The assistant remembers actions already run in the conversation: asking to resend gets a confirmation instead of a duplicate.
Outcome: Have the assistant read new mail, draft and send a real email after approving the confirmation card, read the tool timeline and check the mail in Gmail.
Understand each Facebook Page as its own connection, pick a Page in the integration dialog, connect a new app right inside chat and ask about the Page.
Use it when: Before publishing posts or replying to comments from chat for the first time.
Steps
- 1Facebook in chat works per Page. Every Page you manage is its own connection, so the assistant knows exactly which Page it posts or replies on.
- 2No Facebook or other app connected yet? Click Connect right in this dialog: sign in in a small window, the chat stays put, and the new account appears when done.
- 3For a freshly signed-in Facebook account, StepX asks which Pages to use. Pick them; each becomes its own row, and more Pages can be added any time.
- 4Click the Page to arm it for the conversation. The chip carries the Page name: from now on every post, comment and message is pinned to this Page.
- 5Try asking about the Page: followers, latest posts. Approve the first read and the assistant pulls real data from Facebook and summarises it.
Good to know
- Every Page is its own connection; the chip carrying the Page name pins every post, comment and message to that Page.
- Unconnected apps can be connected right inside the Use integration dialog, no trip to the Integrations page.
Outcome: Understand each Facebook Page as its own connection, pick a Page in the integration dialog, connect a new app right inside chat and ask about the Page.
Have the assistant draft a post, refine it by talking, approve the Create post card, check the live post on Facebook and edit it afterwards.
Use it when: When you need a quick post, edited by talking, published without leaving StepX.
Steps
- 1Describe the post you want: topic, tone, length, hashtags. The assistant drafts it first and publishes nothing yet.
- 2Edit by talking, as with a content writer: add a call to action, change the tone, shorten. When happy, say publish.
- 3When it needs your decision, the assistant asks with a card of ready options: tap a button instead of typing an answer.
- 4The approval card shows the exact text going onto the Page. Approve and it is published, with a link back. Reject and nothing goes out.
- 5On the Facebook Page, the post is live with the exact text and hashtags from the approval card.
- 6Need a change after publishing? Tell the assistant. One approval card again, and the post on the Page is updated in place.
Write a post introducing StepX's new Chat AI: many models, Gmail and Facebook connections, image and video tools. Friendly tone, under 120 words, 3 hashtags. Draft only, do not publish.
Add a call to try it free at stepx.vn at the end, then publish to the Page
Edit the post just published: change the last hashtag to #StepXChat
Good to know
- The approval card shows the exact text going onto the Page; Reject and nothing goes out.
- Editing after publishing also goes through an Update post approval card; the post is updated in place.
Outcome: Have the assistant draft a post, refine it by talking, approve the Create post card, check the live post on Facebook and edit it afterwards.
Find unanswered comments on the Page, approve a reply plan, approve one bulk-send card, follow progress and check on Facebook.
Use it when: When the Page has a backlog of comments after a post or a campaign.
Steps
- 1Ask which comments on the Page are still unanswered. One step scans recent posts and lists every comment the Page has not replied to.
- 2For a multi-step job the assistant proposes a plan first: how many comments, on which Page, which steps send data. Read it and Approve the plan.
- 3Reject with a reason and the assistant adjusts and proposes again. Approve the new plan when it fits.
- 4One approval card for the whole batch: each comment with its reply. Approve once and StepX sends them in turn, reporting how many are done.
- 5On Facebook, each comment now carries the Page's reply, exactly as approved.
- 6Doing this daily? Switch to Allowed: the assistant replies without asking each time. Use it when you hand over the whole job; otherwise keep Ask first.
Does the Page have unanswered comments? List each comment and its post.
Draft an individual, friendly reply for each comment, signed -StepX Team, then send them all.
Reject the plan with a reason: Shorter, each reply under 25 words
Good to know
- One call replies to up to 300 comments; StepX sends in batches and reports progress.
- Doing this daily, switch to Allowed so the assistant stops asking each time; otherwise keep Ask first.
Outcome: Find unanswered comments on the Page, approve a reply plan, approve one bulk-send card, follow progress and check on Facebook.
Pick an image model, write the prompt, get a new picture from another prompt, know how to edit an existing one with a reference image, download or save to Media and read the per-image cost.
Use it when: When you need an illustration for a post, a static message or an internal document.
Steps
- 1In Auto, typing draw or create an image does not switch to an image model on its own, since one picture costs far more than text. It asks you to pick an image model, or answer in text.
- 2Pick an image model; the request is re-sent and the picture appears in seconds. The image model stays selected for later messages until you change it.
- 3Type another prompt and the assistant makes a new picture from it; it does not edit the previous one. To edit an existing picture, attach it to the composer and describe the change.
- 4Happy with a picture? Download it, or Save to Media to reuse it in assistants and static messages. Saving twice never duplicates.
- 5The Image group offers several models, priced per picture or per image token, a few thousand credits each. The real cost sits under every picture.
Draw a friendly robot answering a customer on a phone
Change the background to deep blue, add a StepX wordmark in the corner, flat style
Good to know
- In Auto, typing draw or create an image does not switch models by itself: StepX asks you to pick an image model, since a picture costs far more than text.
- A picture attached to the composer is a reference image: the image model edits that picture instead of drawing from scratch.
Outcome: Pick an image model, write the prompt, get a new picture from another prompt, know how to edit an existing one with a reference image, download or save to Media and read the per-image cost.
Pick a video model, set length, resolution, aspect ratio and audio, send the prompt, wait, watch and save the clip, read the per-second cost.
Use it when: When you need a clip of a few seconds for a post or a video message on a channel.
Steps
- 1The Video group: Veo, Kling, Hailuo, Seedance. Priced per second of footage, with the maximum length shown. Pick Veo 3.1 Lite, the cheapest.
- 2With a video model selected, the composer gains options: length, resolution, aspect ratio, audio on or off. Each changes the price; video models take no attachments.
- 3A video takes a few minutes. The waiting frame has a dinosaur game to pass the time; you may leave the page, the clip will be here when you return.
- 4The clip is ready: play it in chat, download it, or save it to Media for video messages on your channels.
- 5Four seconds of Veo Lite without audio is roughly forty thousand credits, the price of hundreds of text answers. Audio, 1080p or Kling and Seedance raise the per-second price.
- 6Short on credits, StepX warns before generating and charges nothing. Try a shorter clip, a lower resolution, or top up. When done, switch the model back to Auto.
The StepX wordmark fades in among rising chat bubbles, blue gradient background, soft light, slow camera pull-back
Veo 3.1 Lite · 4 s · 720p · 16:9 · No audio
Good to know
- Priced per second of footage and changes with resolution and audio; four seconds of Veo Lite without audio is roughly forty thousand credits.
- Short on credits, StepX warns before generating and charges nothing; try a shorter clip or top up.
Outcome: Pick a video model, set length, resolution, aspect ratio and audio, send the prompt, wait, watch and save the clip, read the per-second cost.
Open the Memories tab, read auto-learned entries and categories, add a memory by hand, switch off or delete unwanted ones and see memories shape answers.
Use it when: When you want the assistant to always remember how to address you, your answer style or work context without repeating it.
Steps
- 1Memories are what the assistant has learned about you across conversations: who you are, how you like answers, what you are working on. Private to you.
- 2Auto-learned entries are drawn from your own conversations. Each has a category, an on/off switch, and edit and delete buttons.
- 3Add by hand what you want the assistant to always remember: how to address you, the answer style, your work context. Keep it short and durable.
- 4Switch an entry off to pause it without losing it; Edit to refine it. At the limit, delete one entry to add another.
- 5Back in chat the answer follows your memories: the right form of address, short, bulleted, without being told again.
- 6Delete an entry that no longer holds; the assistant forgets it.
Good to know
- Memories are private to you in this organization; auto-learned entries come from your own conversations.
- At the limit, delete one entry to add another; switch one off to pause it without losing it.
Outcome: Open the Memories tab, read auto-learned entries and categories, add a memory by hand, switch off or delete unwanted ones and see memories shape answers.
Rename and pin conversations, search, queue messages, stop an answer, delete a conversation and hide the rail.
Use it when: When the conversation rail has grown long, or you want to send more while the assistant is still answering.
Steps
- 1Titles come from the first message; rename them to find things later. Hover an entry and click the three dots.
- 2Pin a conversation you use often to keep it on top. With many entries, search by title.
- 3While the assistant is answering you can keep typing and press Enter: the message queues and runs when the current answer ends. Send now interrupts instead.
- 4Click Stop to cut a long answer short; what was written stays, and only the generated part is charged.
- 5Deleting a conversation removes it from the list; the credits it spent stay in the organization's usage records.
- 6The top-left button hides the conversation rail when you want room to read.
Good to know
- Sending while the assistant is answering queues the message; Send now interrupts instead.
- Deleting a conversation only removes it from the list; credits spent stay in usage records.
Outcome: Rename and pin conversations, search, queue messages, stop an answer, delete a conversation and hide the rail.
Read the credit bar and per-turn cost, open Usage with its three tabs, allocate credits per member, find Top up and know what happens when credits run out.
Use it when: When an owner wants to know who spends what, or to cap each member's share.
Steps
- 1Every chat turn draws on the organization's credits, the same pool as Legal and Tasks. The top bar is what remains this month; each answer shows its real cost.
- 2Owners click Usage: total credits spent, messages, tokens, a chart by day, week or month, and the share by model. Conversation content never appears here.
- 3The Members tab ranks who spent how much and on which models. Enough to see where credits go without reading anyone's content.
- 4Credit allocation: Split equally, then adjust each member by percent or credits. Once saved, a member who uses up their share stops, without affecting others. Here I undo without saving.
- 5Out of credits, the message is not sent and nothing more is charged; StepX says so plainly. Top up right here or upgrade under Limits and plans; top-up credits do not expire monthly.
- 6That is all of Chat AI: Q&A, files, web search, Gmail and Facebook, images, video, memories and credits. Enjoy working with StepX.
Good to know
- Chat, Legal and Tasks share the organization's credits; every answer shows its real cost.
- Out of credits, the message is not sent and nothing more is charged; top up here or upgrade under Limits and plans.
Outcome: Read the credit bar and per-turn cost, open Usage with its three tabs, allocate credits per member, find Top up and know what happens when credits run out.