25 Live Chat Scripts That Make AI Support Feel Human in 2026

A real-time chat box is still the highest-leverage surface in customer support. The same visitor who would never open a ticket will type a one-line question into a widget - and that one line is where renewals, returns, and abandoned carts get decided.

The trouble is that the words on the other side of the chat are hard to write under pressure. New agents fumble the opening. Senior agents end up rephrasing the same answer twenty times a day. And AI agents - even the very capable ones - sound robotic the moment you stop guiding them with concrete examples of how your team actually talks.

Scripts solve all three problems at once. They cut response time, hold a consistent tone across humans and bots, and give an AI agent a tangible voice to imitate rather than a generic "be helpful" instruction. What follows is a working library: the lines we see real support teams use, organized by the situation that triggers them, written so you can paste them straight into your help center, your macros, or - if you are running an AI agent on Berrydesk - your training set.

What a live chat script actually is

A live chat script is a pre-written response, or a small family of responses, attached to a recurring situation: a greeting, a refund request, a shipping question, a checkout abandonment. It is not a rigid call-center read; it is a starting point that an agent (human or AI) bends to fit the conversation in front of them.

Picture the moment a visitor opens a widget on a pricing page. The first message they see - "Hi! Want a quick walk-through of the plans?" - is almost never improvised. Either someone wrote it once and saved it as a macro, or it is one of a few rotating openers a chatbot has been trained on. That is a script doing its job.

For a human team, scripts compress the loop between "I see the question" and "I send a useful answer." For an AI agent, scripts do something subtler: they teach the model your house style. Modern systems like Claude Opus 4.7, GPT‑5.5, and DeepSeek V4 are excellent at following an instruction like "sound friendly and concise" - but they are dramatically better when you show them ten real examples of how your senior agent opens a refund conversation. Scripts are the cheapest, fastest training signal you can give a support agent.

Why scripts matter even more in 2026

It is fair to ask whether scripts are obsolete now that models can hold a million tokens of context and reason for hours about a thread. They are not - and the reason is worth pausing on.

The frontier got smarter, not more opinionated. Anthropic's Claude Opus 4.7 leads SWE‑bench Pro at 64.3% for complex coding work, and Gemini 3.1 Ultra runs natively across text, image, audio, and video with a 2M‑token context window. On the open-weight side, DeepSeek V4 Flash now serves at $0.14 per million input tokens and $0.28 per million output tokens, GLM‑5.1 from Z.ai posts 58.4 on SWE‑Bench Pro under an MIT license, and Moonshot's Kimi K2.6 can coordinate up to 300 sub-agents across 4,000 steps in a single autonomous run. The capability ceiling is no longer the bottleneck.

What the models still need is grounded specifics about your business. A 2M-token window is meaningless if you have not told the agent how you greet returning customers, what tone to take with a frustrated buyer, when to offer a discount, and when to escalate. Scripts are where that knowledge lives. Pair them with a long-context model on Berrydesk and you get an agent that holds an entire policy library in working memory and speaks like your team.

The economics also push toward scripted, AI-driven chat rather than away from it. Routing routine traffic through an open-weight model like DeepSeek V4 Flash or MiniMax M2 - the latter runs at roughly 8% the price of Claude Sonnet at twice the speed - pushes per-resolution cost into fractions of a cent. Reserve Claude Opus 4.7, GPT‑5.5 Pro, or Gemini 3.1 Ultra for the genuinely hard escalations and your support P&L stops being a cost-per-ticket problem.

Greeting a customer

A greeting sets the temperature for the entire conversation. Aim for warm, specific, and short.

• Standard opener: "Hi there - welcome to [Company]. What can I help you with today?"
• Returning customer: "Welcome back, [Name]. Good to see you again. Picking up something new, or following up on [last topic]?"
• Page-aware opener: "Hi! I noticed you were looking at [product] - happy to answer anything about it."
• Seasonal opener: "Happy [Holiday]! Quick heads-up: our team is on a slightly slower schedule this week, but I'll get to your message as soon as I can."

Train an AI agent on three or four greetings rather than one. Conversational variety is what separates a chatbot people use from a chatbot people screenshot to mock on social media. With a Berrydesk agent, you can also feed it the visitor's page URL and prior conversations, so the model picks the right opener instead of you having to branch by hand.

Handling a complaint

When someone is upset, the first job of a script is to slow the conversation down and acknowledge the problem before reaching for a fix.

• Acknowledge: "I'm really sorry that happened, [Name]. Let me take a look right now."
• Empathize: "That's a frustrating place to be - I'd feel the same. Let's get this sorted."
• Propose: "Here's what I can do straight away: [solution]. Does that work, or would you prefer something different?"
• Escalate: "This needs a closer look from our [team] - I'm flagging it now and you'll hear back within [time window]. I'll stay copied so nothing falls through."

The mistake to avoid is jumping to the solution before the customer feels heard. Even with an agentic model that could refund instantly, leading with empathy lands better than leading with the action. A line like "I'd feel the same" is small, but it costs nothing and consistently lifts CSAT scores in chats we've audited.

Following up after an interaction

Follow-ups are the cheapest, most underused move in support. They turn one-off ticket-closers into relationships.

• Post-resolution check: "Hi [Name], circling back - is everything still working as expected? Happy to dig in if anything's off."
• Open-thread check: "Hey [Name], still on this with you. The team is working on [issue] and I'll have an update by [time]."
• Satisfaction probe: "Quick one: did the answer earlier actually solve it for you? Honest feedback helps us tune things."
• Soft cross-sell: "While I'm here - you mentioned [adjacent need] earlier. We do [feature] for that, want me to walk you through it?"

A long-context AI agent has a structural advantage on follow-ups. Because Berrydesk lets the agent keep entire conversation histories in memory, it can reach back to a problem from three weeks ago and bring it up by name without a human prompting it.

Welcoming a first-time customer

New visitors are conversion gold and they signal themselves clearly: no account, fresh session, often loitering on a pricing or product page. Greet them with an offer to guide, not just to answer.

• Welcome: "Hi! Looks like this might be your first visit - welcome. What brought you over today?"
• Proactive tour: "Want a 60-second tour of how [product] works? I can show you the parts most people care about first."
• Onboarding nudge: "Happy to walk you through your first [order/setup]. I can stay on the chat while you go through it."
• First-time offer: "Heads-up: first-time customers get [bonus]. Want me to apply it for you when you're ready to check out?"

Notice that none of these are pushy. They offer help, then back off. AI agents are particularly prone to over-selling, so script the hand-off moments - "let me know when you're ready" - as carefully as the openers.

De-escalating an angry customer

Escalation scripts have a different rhythm: shorter sentences, more "I" statements, fewer exclamation marks.

• Acknowledge anger: "You're right to be frustrated, [Name]. I want to fix this."
• Own it: "This is on us. I'm sorry. Let me make it right."
• Quick win: "The fastest thing I can do right now is [action]. I'll do that, then we can talk about [bigger fix]."
• Hold the line: "I hear you, and I'm not going anywhere - I'll keep updating you as I work this. Last update was [time], next will be [time]."

This is the section where AI tone matters most and where cheaper models tend to fall down. We recommend routing detected-frustration conversations to a frontier model on Berrydesk - Claude Opus 4.7 or GPT‑5.5 - even if the rest of the queue runs on a cheaper open-weight model. The CSAT difference on hard conversations is real, and the volume is small enough that the cost barely registers.

Providing order and shipping updates

Status questions are usually quick - but they're also the conversations where a wrong answer creates the worst downstream pain.

• Asking for the order number: "Happy to check - what's your order number? (It usually starts with #.)"
• Sharing tracking: "You're all set - tracking link: [link]. Estimated delivery is [date]. Anything else you'd like me to look at?"
• Explaining a delay: "I see a delay on this one - [reason]. New estimate is [date]. I'm watching it; I'll ping you the moment it ships."
• Confirming delivery: "It looks like this delivered [time]. Did it actually arrive in good shape, or is something off?"

This is where AI Actions earn their keep on Berrydesk. Rather than scripting the agent to ask a human to look up the order, you wire the agent to your order system directly: it pulls live status, posts a tracking link, and updates the customer in one turn. The agent uses the script for tone; the action handles the data.

Picking up where you left off

Returning visitors should never feel like they're starting from scratch.

• "Welcome back, [Name] - last time we were sorting out [issue]. Did that resolve, or want to keep going?"
• "Hey again. I saw you tried [solution] last week - did it stick?"
• "Picking up where we left off - you were considering [plan/product]. Any new questions on it?"

This is where the 1M‑token context windows on Claude Sonnet 4.6 and DeepSeek V4 actually shine. With Berrydesk, you can let the agent keep the full transcript across sessions and just recall prior context naturally, instead of building brittle "session memory" infrastructure on top.

Proactive messages by page

A proactive message is a script that fires based on where the visitor is, what they've done, or how long they've been still. Match the line to the page.

On the pricing page

• "Looking at plans? I can break down the differences in plain English if it helps."
• "Two questions and I can usually narrow it to the right plan: how many seats, and what's the main thing you're trying to do?"

On the homepage

• "Hi! Just exploring, or looking for something specific? Either is fine - I'm here either way."

On the checkout page

• "You're almost there. Anything blocking you from finishing? I can help if a code, address, or payment is being weird."

On product or category pages

• "If you're between two options on this page, tell me what you'll mostly use it for and I'll tell you which one I'd pick."
• "Stuck filtering? I can shortlist three picks based on what you've told me."

The trick with proactive messages is restraint. Fire too early and you annoy people; fire too late and you miss the moment of doubt. A reasonable default: ten to fifteen seconds of dwell, or a clear hesitation signal like a back-and-forth between two products.

Promotions and discounts

Promo scripts have to feel like a heads-up, not a hard sell. Lead with the value, mention the constraint, then get out of the way.

Promoting a product

• "Quick heads-up - [product] is [X%] off through [date]. Code [PROMO] at checkout."
• "You've got a [X%] discount waiting on your account - want me to apply it?"
• "Sale starts [date] on [category]. Want me to drop you a reminder when it goes live?"

Recovering a checkout

• "Looks like you got most of the way through - here's [X%] off if it helps you decide: [code]."
• "If shipping is what's holding you up, I can offer free shipping on this order. Worth a try?"

These scripts are also the ones to be most careful about with an AI agent. Make sure your training set is explicit about who gets which discount and under what conditions - otherwise the model will discover, the way models do, that giving discounts ends conversations quickly and start handing them out for sport.

Refunds and returns

Refunds and returns are a process, and the script's job is to make a process feel like a conversation.

Refund request

• "I'm sorry [product] didn't work out. I can start the refund - could you share your order number?"
• "Got it. To make sure I refund the right charge: was it the [date/amount] order?"
• "All set - refund is processing now. You'll see it in [X] business days, and I've emailed the receipt."

Return request

• "I can help you return this. Quick check: was it damaged, wrong item, or just not the right fit? That changes which option I can offer."
• "Easiest way is our [returns portal] - I'll generate a label and email it to you. Anything I should flag for the warehouse?"
• "Return received and processed. Refund/exchange is on its way; you'll get a confirmation shortly."

For an AI agent, this is the cleanest case for AI Actions. Wire the refund flow to your billing system directly so the script and the action are one motion: the agent says "refunding now" because it is refunding now, not because it hopes a human will get to it.

Technical support assistance

Tech support scripts have to do two things at once: confirm the problem precisely, and walk the user through fixes without sounding like a wiki dump.

Confirming the issue

• "Sorry you're hitting this. Can you tell me what you see - error message, blank screen, weird behavior - so I can match it to a known fix?"
• "Got it. Just to be sure: it happens every time you do [action], or only sometimes?"
• "Thanks for the screenshot. That looks like [diagnosis]. Want me to walk you through the fix?"

Walking through a fix

• "Try this first: [step]. Tell me what happens - even if nothing changes."
• "That fix usually clears it. If not, the next thing I'd try is [step]. Want to keep going?"
• "If neither works, the cleanest path is [escalation]. I can open that for you now."

A capable model - GPT‑5.5, Claude Opus 4.7, or an agentic open-weight like Kimi K2.6 - can run an entire diagnostic loop autonomously, including reading logs and proposing fixes. The script's job is to make sure it narrates what it is doing in a tone the customer can follow, instead of dumping a stack trace.

Out-of-stock and restock scripts

Out-of-stock scripts should always do two things: tell the truth, and offer a path forward.

When the item is unavailable

• "That one's sold out right now - sorry. Restock is expected around [date]; want me to put you on the alert list?"
• "Out of stock at the moment. We've got two close alternatives in stock - want a quick look?"
• "Currently unavailable, no firm restock date yet. Drop your email and I'll make sure you're first to hear."

Offering alternatives

• "Three options sit in roughly the same spot as the one you're looking at: [A], [B], [C]. Want a one-line difference for each?"
• "If color/size is flexible, I can show you the same model in what's actually in stock."

The pitfall to avoid: AI agents that confidently invent restock dates. If you do not know, say so. Train explicitly with a script line for that case - "no firm restock date yet" - so the model has a graceful default instead of hallucinating one.

Payment issues

Payment problems are sensitive. Scripts here should be calm, neutral, and clear about what the customer should try next.

Failed payment

• "Looks like the payment didn't go through. That's usually a bank-side check - want to try a different card or method?"
• "Card was declined. Most common reasons are address mismatch or a fraud hold. Either ring a bell?"

Double-charge

• "That shouldn't have happened. Send me a screenshot and your order number - I'll trace and refund the duplicate today."
• "Confirmed - I see two charges. Reversing the extra one now; you'll see it back within [X] business days."

If you wire your AI agent to your payment system on Berrydesk via AI Actions, double-charge resolution moves from a multi-day ticket to a one-turn fix. The script keeps the human-feeling part; the action does the heavy lifting.

Restock alerts and back-in-stock follow-ups

Closing the loop on a back-in-stock matters more than the original out-of-stock conversation.

• "Good news - [item] is back in stock. You asked me to ping when it returned. Here's the link: [link]."
• "Restock landed earlier than expected. If you still want one, I'd grab it soon - they tend to sell out within a day."
• "It's back. Want me to apply the discount we'd talked about, or has that expired on your end?"

Common pitfalls when using scripts

A few patterns sink even well-written scripts in practice. Watch for them.

Sounding canned. A script is a starting point, not a recital. Train AI agents on families of three to five variants per situation, and let the model pick or recombine. Single-line scripts repeated verbatim are why "AI sounds AI."

Over-promising. Scripts written by marketing tend to use phrases like "right away" or "instant" that operations cannot back up. Audit your scripts against what you can actually deliver and rewrite the ones that overshoot.

Stale references. Scripts mentioning specific carriers, hours, or pricing tend to drift. Keep them in a single source - your knowledge base, Notion, or Drive - and let your AI agent pull from there. With Berrydesk, you can connect Notion or a Drive folder once and have the agent stay in sync as you edit.

No escalation rule. Every script library should have at least one explicit "stop and escalate" line, with the criteria that trigger it. Without it, an AI agent will keep trying to solve things that should be escalated, and a tired human agent will too.

Treating AI like a single capability. The 2026 model landscape rewards routing. A typical Berrydesk deployment runs DeepSeek V4 Flash or MiniMax M2 on routine traffic, GLM‑5.1 or Qwen3.6 for tool-heavy AI Actions, and Claude Opus 4.7 or GPT‑5.5 Pro for hard escalations. Scripts give all of those models the same voice; routing keeps the cost sane.

Open-weight versus closed frontier for chat scripts

A practical question keeps coming up: which model should run the scripted conversations?

For most support chats, open-weight frontier models are now the right default. DeepSeek V4 Flash at $0.14 / $0.28 per million tokens, MiniMax M2 at roughly 8% the price of Claude Sonnet, and Alibaba's Qwen3.6‑27B (Apache 2.0) all comfortably handle the scripted conversation patterns above. They follow tone guidance, stay on policy, and call AI Actions reliably.

Closed frontier still wins on the long tail. When a conversation goes off-script - a creative refund negotiation, a multi-issue thread, a customer who is genuinely unhappy and articulate about it - Claude Opus 4.7, GPT‑5.5 Pro, and Gemini 3.1 Ultra still pull ahead on judgment and recovery. The good news is that you do not have to choose. On Berrydesk you can pick the model per agent, and route within a single agent based on the kind of conversation it's in.

For regulated industries, the MIT-licensed Chinese open weights - GLM‑5.1, Qwen3.6‑27B, MiMo‑V2 - also unlock on-prem and air-gapped deploys. If your compliance team has a hard rule against shipping conversations to a third-party API, you no longer have to give up frontier capability to honor it.

Bring your scripts to life

Scripts on a page are useful. Scripts wired into an AI agent that is trained on your docs, branded with your colors, and connected to your booking and payment flows are something else entirely.

That is what Berrydesk does. Pick a model - GPT‑5.5, Claude Opus 4.7, Gemini 3.1, DeepSeek V4, Kimi K2.6, GLM‑5.1, Qwen3.6, MiniMax M2, and others. Train it on your sites, PDFs, Notion, Drive, or YouTube. Drop in scripts like the ones above as examples. Brand the widget. Add AI Actions for refunds, bookings, and order lookups. Deploy to your site, Slack, Discord, WhatsApp, and beyond - in an afternoon, not a quarter.

Start free at berrydesk.com and turn this script library into a support agent your team would actually be proud of.