The complete guide to AI customer support tools

This guide is for support leaders, operations people and founders who run a help desk and are being asked to put AI in front of customers. You need to know what these tools do, what they cost and how to tell whether they work.

Afterwards you should be able to compare the main tools on how they bill, read each vendor’s definition of a “resolution” skeptically, and plan a rollout that protects customers and budget. We include a worked cost example at list prices as of October 2026, with the arithmetic labelled as ours.

Our position: the per-unit price on the pricing page matters less than the rule behind it. Every vendor counts something different, and the same rule that decides your invoice also produces the resolution rate on your dashboard.

Summary

• AI support tools fall into four jobs: resolving customer questions on their own, assisting human agents, triaging incoming tickets and reviewing quality. Most buying decisions are about the first.
• Modern AI agents answer from your help content using retrieval, take actions through APIs, and hand off to humans. Scripted chatbots follow fixed decision trees.
• Vendors bill in incompatible units: per outcome (Intercom Fin), per verified resolution (Zendesk), per session (Freshdesk), per conversation or per action (Salesforce Agentforce) and per automated interaction (Gorgias).
• Per-resolution pricing scales with success. Per-session and per-conversation pricing scales with volume, whether the AI helps or not.
• You are responsible for what your bot tells customers. A Canadian tribunal held an airline to an answer its chatbot invented.
• In the EU, people must be told they are talking to an AI system, a duty that applies from 2 August 2026 under the AI Act.
• Measure cost per resolved conversation, reopen rate and customer satisfaction on AI-handled conversations separately, with your own definitions.

1. What AI support tools do

“AI customer support” covers four distinct jobs. Vendors bundle them, but they are bought, priced and measured differently.

Resolution (sometimes called deflection). An AI agent talks to the customer directly, on chat, email, messaging apps or voice, and tries to close the issue without a human. “Deflection” is the older word, for steering a customer to a help article. Resolution goes further: the agent answers in its own words and can act, such as looking up an order.

Agent assist. The AI works beside a human agent inside the help desk. It drafts replies, summarizes long threads, suggests the relevant article or macro (a saved reply template) and fills in ticket fields. Vendors usually price it per seat. Intercom lists its Copilot at $29 per agent a month billed annually, and Freshdesk lists Freddy AI Copilot at $29 per agent a month billed annually as an add-on for its Pro and Enterprise plans.¹1 Source: Intercom, “Pricing”, accessed Oct 2026. Covers seat prices, Fin at $0.99 per outcome, the outcome definition, minimum monthly commitment, Copilot pricing and Fin on other help desks. Essential is shown at $19 for new customers, down from $29. ²2 Source: Freshworks, “Freshdesk pricing”, accessed Oct 2026. Prices billed annually.

Triage. The AI reads incoming tickets and sets intent, language, sentiment and priority, then routes the ticket to the right queue.

Quality assurance (QA). The AI scores conversations, human or AI, against a rubric: was the answer correct, was policy followed, was the tone right.

How this differs from the chatbot you already have

A classic chatbot is a decision tree. Someone wrote every branch: “Where is my order?” leads to “Enter your order number” leads to a status lookup. It is predictable, and it fails when a customer phrases things differently.

An AI agent uses a large language model (LLM) to understand the question and write the answer. That makes it flexible and introduces the main new risk: a fluent answer that is wrong (chapter 6).

These tools are not for every team. If you handle a few hundred tickets a month and most need judgment, a well-kept help center and a good human team will beat any agent. The economics in chapter 5 only start to matter at volume.

2. How an AI support agent works

Under the vendor names, nearly every product has the same three parts.

Retrieval over your knowledge

The agent does not “know” your refund policy. It searches your help center, internal docs and uploaded files, and gives the most relevant passages to the LLM with the question. The model writes an answer grounded in those passages. The technique is called retrieval-augmented generation (RAG).

Two things follow. First, answer quality is capped by content quality. If your help center contradicts itself or is out of date, the agent will repeat the contradiction confidently. Second, scope is set by what you connect.

Salesforce calls the same idea “grounding”, in which prompts are enriched with trusted company data, inside what it calls the Einstein Trust Layer.³3 Source: Salesforce Trailhead, “Explore Agentforce Guardrails and Trust Patterns”, accessed Oct 2026.

Actions and tools

Many tickets need data or a change in another system: order status, a subscription change. The agent does this by calling APIs you configure.

Intercom splits this in two. A data connector is a configured API call that lets Fin answer a customer-specific question in one step, such as order status, and the company says connectors are available to all customers. A Procedure is a multi-step process defined in natural language that combines connectors with business logic, waits and identity checks such as one-time passcodes. Intercom says Procedures are under managed availability.⁴4 Source: Fin Help, “Fin Procedures and Data connectors explained”, accessed Oct 2026. Intercom’s quick-start guide also advises excluding payment details and personal data from connector responses. Salesforce structures Agentforce around topics, instructions and actions: topics group what the agent may do, and instructions use patterns like “Always…”, “Never…” and “If x, then y”.³

Every action is a place where a wrong answer becomes a wrong refund. Give the agent the narrowest set of actions it needs, with identity checks before anything touches an account.

Handoff

When the agent cannot or should not continue, it passes the conversation to a person, ideally with a summary and the data it already gathered. Chapter 6 covers the rules.

For how these agents compare with AI used elsewhere in operations, see the complete guide to AI agents in enterprise operations.

3. The main tools compared

Five help desk vendors and two enterprise specialists cover most shortlists. The table uses what each vendor publishes about itself, at list price as of October 2026. Where a vendor publishes no rates, we say so.

6 Source: Zendesk, “Pricing”, accessed Oct 2026. Covers per-agent plan prices, included automated resolutions and overage rates. 7 Source: Salesforce, “Agentforce pricing”, accessed Oct 2026. Salesforce notes the pricing is for information only and may change. The $0.10 per action figure is our arithmetic from $500 per 100,000 credits. 8 Source: Salesforce, “Service Cloud pricing”, accessed Oct 2026. Starting prices per user per month; Core is billed annually. 9 Source: Gorgias, “Pricing”, accessed Oct 2026. Parts of the page showed conflicting plan prices; we used the plan table. 10 Source: Decagon, “Pricing AI agents”, accessed Oct 2026. 11 Source: Sierra, “Outcome-based pricing for AI agents”, accessed Oct 2026.

Tool	Sold as	AI billing unit	Published AI rate	Seat or base price (list)
Intercom Fin	Help desk plus AI agent; Fin also works on other help desks	Outcome (once per conversation)	$0.99 per outcome; minimum monthly commitment applies1	$19 to $132 per seat a month; Fin alone needs no seats1
Zendesk AI agents	Included in Suite and Support plans	Automated (verified) resolution	5 or 10 included per agent a month; then $1.50 committed or $2.00 pay-as-you-go6	$19 to $115 per agent a month, annual; Enterprise via sales6
Salesforce Agentforce for Service	Add-on or edition on Service Cloud	Conversation, or Flex Credits per action, or per-user add-on	$2 per conversation; $500 per 100,000 Flex Credits, 20 credits per action7	Service Cloud from $25 to $550 per user a month8
Freshdesk Freddy AI	Freshdesk add-on	Session	500 sessions included; then $49 per 100 sessions2	$19 to $89 per agent a month, annual2
Gorgias AI Agent	Ecommerce help desk	Automated interaction (also counts as a ticket)	$0.90 annual or $1.00 monthly; $1.50 above plan allowance9	$10 to $750 a month by ticket volume; Enterprise custom9
Decagon	Enterprise AI agent platform	Per conversation or per resolution	Not published10	Custom contract
Sierra	Enterprise AI agent platform	Outcome-based	Not published11	Custom contract

A few notes the table cannot carry.

Intercom and Salesforce are becoming one company. On 15 June 2026 Salesforce announced a definitive agreement to acquire Fin, the company formerly known as Intercom, for about $3.6 billion, with closing expected in the fourth quarter of Salesforce’s fiscal year 2027, subject to regulatory clearance.¹²12 Source: Salesforce Investor Relations, “Salesforce Signs Definitive Agreement to Acquire Fin”, accessed Oct 2026. Until it closes they remain separate vendors. If you sign a multi-year contract with either, ask how its terms survive the deal.

Gorgias is for ecommerce. Its plans are priced by ticket volume, from 50 tickets on Starter to 5,000 on Advanced, with per-ticket overage.⁹ It fits online stores, not B2B software support.

Decagon and Sierra sell to large companies. Neither publishes rates. Decagon describes two models, a fixed rate per conversation or a higher rate per resolved conversation with escalations free, and says most of its customers choose per-conversation pricing.¹⁰ Sierra says it charges for outcomes such as a resolved conversation or a saved cancellation, and that in most cases an unresolved conversation is not charged.¹¹ Expect a sales process and an implementation project, not a self-serve trial.

Agentforce has more than one meter. Salesforce lists per-conversation pricing, Flex Credits consumed per action and per-user add-ons at $125 per user a month.⁷ Which one you are on changes the cost curve completely, as chapter 5 shows.

4. What “resolution” means, vendor by vendor

Every vendor reports a resolution or automation rate, and most bill on something close to it. They do not count the same thing.

Intercom Fin. Intercom counts an outcome when a customer confirms the issue is resolved, when they do not ask for more help after Fin responds, or when Fin completes a Procedure, including handoffs.¹ Its help center calls the second case an “assumed resolution”. Fin must have given a real answer, so a reply to a greeting does not count, and if the customer comes back to a resolved conversation the charge is reversed, even across billing periods. You pay once per conversation however many questions Fin answers.¹³

Zendesk. Since 18 May 2026 Zendesk uses resolution tiers. After a conversation ends, an LLM reads it to judge whether the request was satisfactorily resolved. A verified resolution passes that check and counts against your allowance. A contained resolution is one where the customer asked for nothing more but the check failed, and it is not billed. An assisted escalation, where the AI helped but a human finished, is not billed either. Email conversations end 72 hours after the first email; messaging conversations end two hours after the first message by default, extendable to 72 hours.¹⁴

Freshdesk. A Freddy AI Agent session covers all of one customer’s interactions with the AI agent within a 24-hour window.¹⁵ The unit measures usage, not success. When sessions run out, conversations go to human agents.¹⁵

Gorgias. An interaction counts as automated when the issue is fully handled by AI Agent or another automation feature and the customer does not need a human within 72 hours. If AI Agent hands over to a person, you pay only the ticket fee. If it resolves the ticket alone, you pay both a ticket fee and an automation fee.¹⁶16 Source: Gorgias Docs, “Automate pricing”, accessed Oct 2026.

Salesforce Agentforce. On the conversations model, Salesforce charges $2 per conversation, and on Flex Credits it charges per action.⁷ Neither unit depends on whether the customer’s problem was solved.

What this means in practice

First, vendor-reported resolution rates are not comparable. An “assumed” resolution counts a customer who gave up and closed the window. A verified resolution excludes some genuine successes the checking model was unsure of. Rebuild any two vendors’ numbers on one definition, using your own data, before comparing them.

Second, pricing models carry different incentives. Under per-resolution pricing, check what counts as closed. Under per-conversation and per-session pricing, check what counts as an engagement and keep the agent away from conversations it cannot help with.

5. A worked cost example

Here is one support team priced on four platforms. Prices are list prices as of October 2026 from each vendor’s pricing page. The arithmetic and assumptions are ours; real bills will differ with discounts, taxes, commitments and configuration.

The team

• 10 human agents.
• 8,000 customer conversations a month, all of which reach the AI agent first.
• The AI resolves 35% of them, or 2,800 conversations, by each vendor’s own definition. We chose 35% for illustration. It is not a benchmark, and your rate will differ.
• One session or one conversation per customer conversation (a simplification).

The arithmetic

Intercom (Advanced seats plus Fin). 10 seats at $85 is $850. 2,800 outcomes at $0.99 is $2,772. Total: $3,622 a month.¹

Zendesk (Suite Professional, annual). 10 agents at $115 is $1,150. The plan includes 10 automated resolutions per agent, so 100 are included. The other 2,700 at the committed rate of $1.50 cost $4,050. Total: $5,200 a month. At the pay-as-you-go rate of $2.00 the AI part is $5,400 and the total $6,550.⁶

Freshdesk (Pro, annual). 10 agents at $55 is $550. The first 500 sessions are included, leaving 7,500 sessions, or 75 packs of 100 at $49: $3,675. Total: $4,225 a month.²

Salesforce (Service Cloud Core plus Agentforce). 10 users at $195 is $1,950.⁸ On the conversations model, 8,000 conversations at $2 cost $16,000, for a total of $17,950 a month. On Flex Credits, at $500 per 100,000 credits each credit costs half a cent, so a 20-credit action costs $0.10. If the agent averages three actions per conversation, 24,000 actions cost $2,400, for a total of $4,350 a month.⁷ Your action count per conversation is the number to estimate before you sign.

Platform	Seats (list)	AI usage at 35% resolved	Total a month	AI cost per AI-resolved conversation
Intercom, Advanced + Fin	$850	$2,772	$3,622	$0.99
Zendesk, Suite Professional (committed)	$1,150	$4,050	$5,200	$1.45
Freshdesk, Pro + Freddy sessions	$550	$3,675	$4,225	$1.31
Salesforce Core + Agentforce conversations	$1,950	$16,000	$17,950	$5.71
Salesforce Core + Agentforce Flex Credits (3 actions)	$1,950	$2,400	$4,350	$0.86

What changes the answer

A better resolution rate. At 60% resolved (4,800 conversations), Intercom’s AI cost rises to $4,752 and Zendesk’s to $7,050, because they bill on success. Freshdesk’s and Agentforce’s stay where they were, because they bill on volume.

A worse one. At 15% resolved, the volume-billed tools still charge for all 8,000 conversations. That is the case for keeping the agent away from conversation types it cannot handle.

What we left out. Copilot or agent-assist add-ons, voice, implementation work, the people who maintain content and Procedures, and any minimum commitment. Gorgias is missing because its ticket-based plans and dual ticket-plus-automation fee need a store’s own ticket mix to price fairly.⁹¹⁶

Change the resolution rate and the ranking changes. Run the arithmetic on your own volumes and definition of resolved.

6. Accuracy, guardrails and escalation

An LLM can produce an answer that sounds right and is not supported by your content. This is hallucination. In support it looks like invented policies, wrong eligibility and steps for features you do not have.

The award was small, about C$812 including interest and fees.¹⁷ The principle is not.

Guardrails that work

• Narrow the scope. Decide which topics the agent may handle and refuse the rest. Salesforce recommends mapping user inputs to specific topics to reduce hallucinations.³
• Ground answers in approved sources only. Turn off general model knowledge where the tool allows it.
• Gate actions. Read-only lookups first. Changes to orders, money or account access only after identity verification, with limits (for example, refunds above a set amount go to a human).
• Protect sensitive fields. Keep payment details and personal data out of what connectors return to the model. Intercom’s own setup guidance makes the same point.⁴
• Log everything. You need the full conversation, the sources retrieved and the actions taken to investigate a complaint. Salesforce lists audit trails among its Trust Layer features.³

Escalation design

Write escalation rules before you write anything else. Typical triggers:

• The customer asks for a person. Honour it on the first request.
• The topic is out of scope, regulated, or involves a complaint, legal threat, safety issue or cancellation you want a human to save.
• The agent fails to find a grounded answer or a verification step fails.
• Sentiment turns sharply negative, or the same question comes back twice.

The handoff should carry the summary and collected data so the customer does not repeat themselves. Gartner’s analyst advises against making generative AI a mandatory first step, noting that customers pushed through several failed AI exchanges before reaching a person are less likely to use the tool again.⁵

7. Data privacy and the EU AI Act

This chapter compares vendor terms and summarizes regulation. It is not legal advice; involve counsel for your own situation.

Questions to ask every vendor

• Is there a data processing agreement (DPA)? A DPA is the contract that governs how a vendor processes personal data on your behalf. Zendesk says that where a DPA is in place, its terms continue to apply to AI features.¹⁸18 Source: Zendesk Help, “Zendesk AI Data Use Information”, accessed Oct 2026.
• Do you train on our data? Answers differ. Zendesk says its generative AI features are not trained on customer data and that no third party will use your inputs to train its models; its own non-generative models may be trained on sanitized customer data.¹⁸ Intercom says it uses anonymized customer data to train and fine-tune some of the models behind Fin, and that customers can opt out through workspace security settings, with requests processed within 30 days.¹⁹19 Source: Intercom Help, “AI model training and fine-tuning”, accessed Oct 2026.
• Which model providers process our data, where and for how long? Ask for the sub-processor list and the retention terms with each LLM provider. Salesforce says data sent through its Trust Layer is never retained by a third-party LLM provider.³
• What reaches the model? Know which fields your connectors and Procedures expose, and whether masking applies to agent traffic.

The EU AI Act transparency duty

Regulation (EU) 2024/1689, the AI Act, requires that AI systems intended to interact directly with people be designed so those people are informed they are interacting with an AI system. The exception is where that is obvious to a reasonably well-informed, observant person in context. The information must be clear and given at the latest at the first interaction. Article 113 sets 2 August 2026 as the date Article 50 applies.²⁰20 Source: EU AI Act, “Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems”, accessed Oct 2026. Reproduces the text of Regulation (EU) 2024/1689; application date per Article 113.

The 2026 Digital Omnibus agreement moved the deadline for marking AI-generated content, a different paragraph of the same article, to 2 December 2026.²¹21 Source: CMS, “EU AI Act developments: key political agreement on the Digital Omnibus on AI”, accessed Oct 2026. In practice: label the agent as AI in the widget and the first message, in every EU market you serve, and do not give it a human name and photo that suggest otherwise.

8. Measurement, rollout and common mistakes

What to measure

Define each metric yourself and calculate it from raw conversation data, not only from the vendor dashboard.

• Resolution rate, your definition. For example: the AI answered, no human touched the conversation, and the customer did not return on the same issue within seven days. Compare it with the vendor’s number every month.
• Reopen and recontact rate. The share of AI-resolved conversations followed by another contact on the same issue. This catches “assumed” resolutions that were abandonment.
• Customer satisfaction (CSAT) on AI conversations, separately. CSAT is the post-conversation rating; track response rate too.
• Escalation quality. Time to a human after escalation, and whether the human had to re-ask for information.
• Cost per resolved conversation. Total support cost (seats, AI usage, add-ons, the people who maintain content) divided by resolved conversations, with AI and human resolutions shown separately.
• Accuracy audit. A weekly sample of AI answers checked by a senior agent against policy. QA tools help; a spreadsheet works to start.

Budgets are moving toward this work. A Gartner survey of 199 service leaders in April and May 2026 found AI spending up 38% while overall service budgets grew 2%.²²22 Source: Gartner, “Gartner Survey Finds AI Spending by Customer Service Leaders Has Surged by 38 Percent”, 26 Aug 2026, accessed Oct 2026.

A rollout plan

1. Clean the content (weeks 1 to 3). Pull your top 50 contact reasons. Make sure each has one current, unambiguous help article. Archive the rest.
2. Shadow mode (weeks 3 to 5). Run the agent on past or live conversations without sending answers, or as agent assist only. Have senior agents grade the answers.
3. Narrow launch (weeks 5 to 8). Turn the agent on for one channel and a short list of topics, with read-only actions and generous escalation. Label it as AI.
4. Measure and tune (weeks 8 to 12). Review failed and escalated conversations weekly. Fix content first, instructions second, model settings last.
5. Add actions. Introduce write actions one at a time, each with verification and limits.
6. Expand channels and topics only when your own resolution, reopen and CSAT numbers hold.

Common mistakes

• Comparing vendor resolution rates as if they were the same metric. They are not (chapter 4).
• Launching on a stale help center. The agent will quote it faithfully.
• Hiding the human. It annoys customers and, given the Gartner findings, pushes them away from the channel entirely.⁵
• Pointing a per-conversation agent at everything. You pay for conversations it was never going to resolve.
• Giving write actions too early. Start read-only. A wrong answer is a complaint; a wrong refund is a loss.
• No owner. Someone must own content, instructions and the weekly review. Without one, quality drifts.
• Ignoring the bill until it arrives. Set usage alerts and caps where the vendor allows them, and check the minimum commitment.

AI agents often sit next to rule-based automation that moves data between systems; for that layer, see the complete guide to workflow automation tools. For where support tooling fits among the rest of your systems, see the complete guide to building a B2B revenue stack.

9. The verdict

For routine questions the main tools are close in capability. What separates them is how they count. Per-outcome pricing (Intercom Fin, Gorgias, Zendesk’s verified resolutions) ties the bill to success, so audit what “success” means. Per-session and per-conversation pricing (Freshdesk, Agentforce conversations) ties it to volume, so control what reaches the agent. Pick the counting rule that matches your volume and expected resolution rate, then measure results on your own definition.

Frequently asked questions

What is the difference between an AI agent and a chatbot?

A chatbot follows a decision tree someone wrote in advance. An AI agent uses a language model to understand free-text questions and write answers from your content, and it can call APIs to take actions. The agent is more flexible and can also be confidently wrong, which a scripted bot rarely is.

Which pricing model is cheapest?

It depends on your resolution rate. Per-resolution pricing is cheaper when the agent resolves a small share of what it sees, because unresolved conversations cost nothing. Per-session or per-conversation pricing gets cheaper per resolution as the agent improves. Run chapter 5’s arithmetic on your own numbers.

Can I use Intercom Fin without moving off my current help desk?

Intercom says Fin can be used with your current help desk, naming Salesforce among others, at $0.99 per outcome with no seat costs, as of October 2026.¹ Check integration depth for your help desk before assuming parity with Intercom’s own inbox.

Do these vendors train AI models on our support conversations?

Some do, under conditions. Zendesk says its generative features are not trained on customer data, while Intercom says it trains some Fin models on anonymized customer data unless you opt out.¹⁸¹⁹ Read the vendor’s AI data-use terms and your DPA before launch.

Do we have to tell customers they are talking to AI?

In the EU, yes, unless it is obvious from context: Article 50 of the AI Act has required this since 2 August 2026.²⁰ Elsewhere rules vary, but clear labelling is good practice.

What resolution rate should we expect?

We do not publish a benchmark, because vendors define resolution differently and rates depend on ticket mix and content quality. Run a shadow-mode test on your own conversations and calculate the rate on your own definition before you commit to a volume.

Should the AI agent handle refunds and cancellations?

Only after it handles questions well, and only with identity verification and limits. Some teams keep cancellations with humans on purpose. Sierra, for example, describes a saved cancellation as one of the outcomes it charges for.¹¹

Sources and notes

1. Source: Intercom, “Pricing”, accessed Oct 2026. Covers seat prices, Fin at $0.99 per outcome, the outcome definition, minimum monthly commitment, Copilot pricing and Fin on other help desks. Essential is shown at $19 for new customers, down from $29. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

2. Source: Freshworks, “Freshdesk pricing”, accessed Oct 2026. Prices billed annually. ↩ ↩² ↩³ ↩⁴ ↩⁵

3. Source: Salesforce Trailhead, “Explore Agentforce Guardrails and Trust Patterns”, accessed Oct 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵

4. Source: Fin Help, “Fin Procedures and Data connectors explained”, accessed Oct 2026. Intercom’s quick-start guide also advises excluding payment details and personal data from connector responses. ↩ ↩²

5. Source: Gartner, “Gartner Survey Finds 87 Percent of Customers Say Companies Using GenAI for Customer Service Must Provide Access to a Human Agent”, 4 Aug 2026, accessed Oct 2026. Survey of 3,566 B2B and B2C customers, February to March 2026. ↩ ↩² ↩³

6. Source: Zendesk, “Pricing”, accessed Oct 2026. Covers per-agent plan prices, included automated resolutions and overage rates. ↩ ↩² ↩³ ↩⁴

7. Source: Salesforce, “Agentforce pricing”, accessed Oct 2026. Salesforce notes the pricing is for information only and may change. The $0.10 per action figure is our arithmetic from $500 per 100,000 credits. ↩ ↩² ↩³ ↩⁴ ↩⁵

8. Source: Salesforce, “Service Cloud pricing”, accessed Oct 2026. Starting prices per user per month; Core is billed annually. ↩ ↩²

9. Source: Gorgias, “Pricing”, accessed Oct 2026. Parts of the page showed conflicting plan prices; we used the plan table. ↩ ↩² ↩³ ↩⁴

10. Source: Decagon, “Pricing AI agents”, accessed Oct 2026. ↩ ↩²

11. Source: Sierra, “Outcome-based pricing for AI agents”, accessed Oct 2026. ↩ ↩² ↩³

12. Source: Salesforce Investor Relations, “Salesforce Signs Definitive Agreement to Acquire Fin”, accessed Oct 2026. ↩

13. Source: Intercom Help, “Fin AI Agent resolutions”, accessed Oct 2026. ↩ ↩²

14. Source: Zendesk Help, “About automated resolution tiers”, accessed Oct 2026. Accounts still on the pre-May 2026 platform follow an older article with similar time windows. ↩ ↩²

15. Source: Freshworks Support, “Freddy AI Agent sessions FAQs”, accessed Oct 2026. ↩ ↩² ↩³

16. Source: Gorgias Docs, “Automate pricing”, accessed Oct 2026. ↩ ↩²

17. Source: Dentons Canada via Mondaq, “Airline ordered to compensate a BC man because its chatbot provided inaccurate information”, accessed Oct 2026. Award: C$650.88 damages, C$36.14 interest and C$125 fees. ↩ ↩²

18. Source: Zendesk Help, “Zendesk AI Data Use Information”, accessed Oct 2026. ↩ ↩² ↩³

19. Source: Intercom Help, “AI model training and fine-tuning”, accessed Oct 2026. ↩ ↩²

20. Source: EU AI Act, “Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems”, accessed Oct 2026. Reproduces the text of Regulation (EU) 2024/1689; application date per Article 113. ↩ ↩²

21. Source: CMS, “EU AI Act developments: key political agreement on the Digital Omnibus on AI”, accessed Oct 2026. ↩

22. Source: Gartner, “Gartner Survey Finds AI Spending by Customer Service Leaders Has Surged by 38 Percent”, 26 Aug 2026, accessed Oct 2026. ↩