Every day your customers ask for things tied to their account: what's on their plan, a copy of last month's invoice, or a new email address on file. The message from someone who isn't the customer looks exactly the same, and an AI agent will answer it unless someone has set it up to check who is asking first.

Verifying customer identity in support means confirming that the person asking is the account holder before the AI shares account details or changes anything, and handing over to a person when it can't be sure. It applies to existing customers writing in by chat or email, after they've signed up.

I'm Mike, co-founder of My AskAI. We help 200+ ecommerce and SaaS businesses run AI customer service inside the helpdesk they already use, and My AskAI is one of the seven tools below.

The tools differ most on the hardest case: someone writing in by email, or from a chat where they haven't logged in. On our side, an iGaming operator feeds live player data from its systems into the AI so it can see each player's account when it answers.

What does it take for an AI agent to verify customer identity?

Verifying the customer breaks down into six separate jobs, and most tools do some of them well and leave the rest to you. These are the six I scored every tool against.

Six jobs an AI agent has to cover to verify a customer, as six cards in a grid: knows which requests need a check, recognizes a signed-in customer, checks a customer who isn't signed in, shares or changes nothing until the check passes, hands over when a check fails or looks wrong, and lets your team see what happened.

1. It knows which requests need a check

A request such as "What card is on my account?" or "change my email" needs a check, because the answer discloses something private or changes the account. The AI can answer "What's your refund policy?" for anyone.

I want to be able to say in plain language which request types need a check, and have the AI follow it, so everyday questions get a quick answer and account details stay with the person they belong to. A subscription app might answer billing-date questions for anyone but ask for a check before it reads out the last four digits of a card.

2. It recognizes a customer who is already signed in

When a customer is logged in on your site or app, or writes from the email address on the account, the AI should know who they are without asking again. Asking a signed-in gamer to prove who they are before they can see their balance feels broken to them.

Getting this right usually means someone on your side connects your sign-in to the AI once. In the tools I looked at, it's a short piece of developer work. It removes the check entirely for every customer who is signed in.

3. It checks a customer who isn't signed in

Email and signed-out chat are where the AI has to run the check itself. Of the six jobs, this is the one where I saw the widest spread between tools. The AI needs a way to confirm the account holder inside the conversation, using a code sent to the email or phone already on file or details only the owner would know, checked against your records.

Security questions on their own are weak, because the answers are often easy to guess or find. A large study of account recovery found it next to impossible to write secret questions that are both secure and memorable, and US federal guidance on digital identity steers services away from them.

4. It shares or changes nothing until the check passes

Account details stay locked until the check passes, and after a pass the AI shares only what the request needs (a question about a withdrawal shouldn't reveal the customer's home address). Once a customer has passed, they shouldn't be asked again in the same conversation. The rule to hand your security reviewer is that nothing account-specific is disclosed or changed before a pass, and the AI never has more access than the customer does.

5. It stops and hands over when a check fails or looks wrong

Some checks should end with a person. Repeated wrong answers, a request to change the very email or phone number a code would go to, or a customer saying "I didn't ask for this" all belong with your team straight away.

The handover should include the conversation so far, so your agent doesn't start from scratch. And asking for a person should always work, whether or not the check passed.

6. Your team can see who was checked and what the AI did

Afterwards, your team should be able to see whether a check happened, how it happened and what the AI shared or changed. When a customer says "I never asked for that refund", that record is how you answer them.

Any job a tool leaves out becomes work your team does by hand, so test all six in a demo (our AI support agent vendor checklist covers the other questions to ask).

How did I score these tools for verifying customer identity?

The scores come from each vendor's help articles, product pages, pricing pages and customer stories. Six criteria are the jobs in the last section, and two more cover what it takes to go live and what it costs.

The criteria, in the order they matter for this job:

  • Knows when to check - whether you can say which requests need a check and the AI follows it
  • Recognizes signed-in customers - whether a customer signed in on your site, app or email is known without being asked again
  • Checks customers who aren't signed in - whether the AI can confirm the account holder inside an email or signed-out chat
  • Shares only after a pass - whether account details and changes stay locked until the check passes
  • Hands failed checks to a person - rules for failed or suspicious checks, an always-available route to a person, and a summary on handover
  • Shows what was checked - whether your team can see afterwards who was checked, how, and what the AI did
  • Setup effort - the time and people needed to go live, assuming your team uses AI to draft the rules and steps
  • Monthly cost at 1,000 account tickets - the lowest plan that covers the volume, monthly, on one declared basis

Each criterion is scored out of 10 and the eight scores add up to the Overall, out of 80. Phone support isn't scored. Where a tool's pages show it covers calls, I say so in its section.

The 7 AI customer service agents that verify customer identity at a glance

Here are the eight scores for each tool, ordered by total from left to right:

(scores out of 10)My AskAIIntercom FinFiniChatbaseAdaZendesk AI agentsDecagon
Knows when to check9887867
Recognizes signed-in customers8888886
Checks customers who aren't signed in6975546
Shares only after a pass8998977
Hands failed checks to a person9786768
Shows what was checked8697687
Setup effort7756564
Monthly cost at 1,000 account tickets8659242
Overall63605956504947

Same eight criteria, in plain words:

CriterionMy AskAIIntercom FinFiniChatbaseAdaZendesk AI agentsDecagon
Knows when to checkTask triggers, plain-language rulesCheck set per connectionStronger check, sensitive actionsSigned-in-only actionsSign-in where data's sensitiveProcedures hand to checkControlled operations
Recognizes signed-in customersEmail sender or signed-in chatSigned-in chat, developer setupSigned-in widget, email lookupWebsite sign-in, developer setupSign-in stepWeb and app chatNot shown for chat
Checks customers who aren't signed inTask checks your recordsEmailed code, chat and emailChecks before account changesGuest order lookupsAsks for the account emailRelies on sign-inVoice authentication
Shares only after a passProceeds only after confirmationData only after codeMismatches go to reviewPayments need signed-in customerSign-in before sensitive dataRestricted answers when signed-inStrict controls
Hands failed checks to a personHandover step plus summaryFailure route, no summaryEscalates with full contextAsk-for-a-person triggerError answer, some summariesEscalation flows you designTransfer with summary
Shows what was checkedEcho and InspectApprovals in conversationEvery action loggedLog by user and actionAudit logging namedVerified check markReasoning trace
Setup effortHelpdesk app, developer stepAdmin switch, developer sign-inLive in 14 daysSelf-serve, new widgetEnterprise buildAdmin plus developerBuilt with Decagon's team
Monthly cost at 1,000 account ticketsAbout $263 on Pro$693 to $990Growth plan fee$150 on StandardSales only$1,350 plus seatsSales only

My AskAI leads on two jobs. It puts the check only in front of the requests that need one, and it gets a failed or suspicious check to a person with the conversation attached. Its bill is also the lowest published one after Chatbase. Intercom Fin beats us on customers who aren't signed in, because its code check comes built in where ours is a Task your team sets up.

I put Intercom Fin second on the strength of the only ready-made check in this set that works by email as well as chat. Fini is the one tool whose pages show all six jobs, and it would score higher if its pricing page printed the Growth plan's monthly fee. Of the tools that publish prices, Chatbase has the lowest bill at this volume (about $150 a month on Standard), and it suits a small team that doesn't run a helpdesk yet.

Where does AI identity verification go wrong?

AI identity verification goes wrong in three ways you can test for in any demo, with what good looks like for each.

Three-card grid of identity check failure modes: takes the customer's word for it, checks everyone so genuine customers give up, lets someone change the details the check relies on.

Failure mode 1: It takes the customer's word for it

An AI that takes a single detail typed into the chat as proof, such as a name, an order number or an email address, will hand account details to the wrong person. Zendesk's authentication article warns that an email typed into a form or an AI agent's prompt could belong to someone else. People fall into the same trap: one analysis of bank call centers found agents almost always ask whichever check question is easiest for the customer.

In a demo, I ask for a balance or an address after giving nothing but an email address, and watch what comes back. Look for a real check against your records before anything is shared.

Failure mode 2: It checks everyone, so genuine customers give up

A code before a policy answer, or a second check after the customer already passed one, pushes people to ask for a person or leave. On one helpdesk's community forum, a support team asked for a way to let customers verify their email once instead of every day.

Check only where the request needs one, and let a pass last the whole conversation. Intercom, for example, doesn't ask again within 30 minutes. My view is that it should always be easy to reach a person, whether or not the customer passed.

Failure mode 3: It lets someone change the details the check relies on

Changing the email or phone number on file, then passing the next check sent to the new details, is the classic takeover route. A US consumer-protection alert describes scammers calling a phone provider's customer service and asking for the number to be moved to a phone they own. US rules for phone carriers bar checks that rely on readily available account or biographical details.

I keep these with a person at first. A change to contact details should always need a stronger check or a person, and a note should go to the old address when the change goes through. If the email on the account changes mid-conversation, the AI should check again.

Can My AskAI verify customer identity?

We built My AskAI to work inside the helpdesk your team already runs: Zendesk, Intercom, Freshdesk, Freshchat, Gorgias or HubSpot. It installs as an app from each helpdesk's marketplace, so your team keeps its inbox and routing, and account requests are handled next to every other ticket.

The My AskAI homepage hero, pitching an AI customer service agent within your helpdesk, with a red 'Create AI Agent' button.

Most account questions need no check, and the AI answers those from your help content, Custom Answers and historic tickets. The check applies only to the requests you name.

How does My AskAI check who it's talking to?

My AskAI runs checks inside a Task, a multi-step workflow you describe in plain language. You give each Task example messages that should start it and examples that shouldn't, so "change my email" starts the check and "what's your refund policy?" gets an answer straight away. Guidance adds plain-language rules for topics that go straight to a person. You pick every request type that gets a check, so knowing when to check scores 9, the top mark for that job.

We use User Data to give the AI the customer's account details, such as their plan, subscription or recent orders, when your helpdesk has already identified them. That covers an email from the address on the account and a chat where your helpdesk knows the customer is signed in. Signed-in chat depends on your helpdesk's sign-in being connected by a developer. Recognizing signed-in customers scores 8, level with five others.

For a customer who isn't signed in, your team sets up a Task that asks for details your process uses, such as an order number and its postal code. The Task checks both against your records through a Tool. Checking customers who aren't signed in scores 6 for us and 9 for Intercom Fin, because its code step is built in and ours is a check your team builds. Each Task step can say "only proceed after confirmation", and each action can run automatically or wait for a teammate to approve it, which puts sharing only after a pass at 8.

We send a failed check to a person. A Task can hand over at any step, for example when a customer keeps struggling to confirm their contact details, and customers can ask to speak to a person at any time. Before the handover, the AI writes a summary of the conversation, and it stays out of the ticket until your team hands it back. No tool here scores higher than our 9 for handing failed checks to a person.

Afterwards, your team can ask Echo, the assistant in our dashboard, why the AI answered or acted as it did, and Inspect & Logs shows the user data and guidance it used for any conversation. I give that an 8 for showing what was checked, since Fini's per-step log is the one clearer record in this set.

Setup scores 7, tied for top: the Task is written in plain language, and a developer connects the Tool to your records once. Writing the Task is quick if you paste your current check process into an AI tool and ask it for steps. You can then run the whole thing in internal-note mode first, so your team reads every proposed reply before a customer does.

Capabilities shipped (out of 6)

JobMy AskAI
Knows which requests need a check✅ Task triggers and Guidance rules in plain language
Recognizes a customer already signed in✅ email sender or helpdesk-verified chat
Checks a customer who isn't signed in⚠️ a Task that checks your records
Shares or changes nothing until the check passes✅ proceeds only after confirmation, automatic or approved
Hands failed or suspicious checks to a person✅ handover inside the Task, summary, always reachable
Team can see who was checked✅ Echo and Inspect & Logs

Who's using My AskAI for account requests?

Three of our customer stories show parts of this setup in use. An iGaming operator on Intercom connected live player data from its systems through User Data, which feeds its payments Tasks, and rolled the AI out progressively.

A digital-goods marketplace, also on Intercom, uses a named Task for each common ticket type, including accounts blocked by a verification check. The Tasks work out which case the customer is in and give them the right answer. A creator-monetization platform on Zendesk keeps its trust and safety tickets with people.

What does My AskAI cost at 1,000 account tickets a month?

We charge $199 a month for Pro, with 1,000 credits included, then $0.12 a credit, and it covers up to 5,000 tickets a month. I've assumed 600 account tickets arrive by chat and 400 by email, each with two AI replies, which is a typical ticket for us. A chat ticket like that uses 1 credit and an email ticket 1.5, so 1,000 tickets use 1,200 credits. That's $199 plus the extra credits at $0.12, or $223.

We bill Tasks at $0.02 a reply on top, and both replies are part of the check Task, which adds $40. That comes to about $260 a month, or roughly $0.26 a ticket, and the price is the same whether the AI finishes the ticket or hands it over. Cost scores 8, behind only Chatbase's lower bill. The 30-day trial unlocks every feature with unlimited tickets and no card, so you can prove it on real account requests before you pay.

For more on My AskAI, see our Tasks and Tools page, our User Data page and our pricing page.

Can Intercom Fin verify customer identity?

Fin is Intercom's AI agent. It runs inside Intercom, or on another helpdesk with a minimum monthly commitment, and it also covers the phone through Fin Voice.

The Intercom homepage hero, titled 'A complete system for human and AI customer service', with 'Start free trial' and 'View demo' buttons.

How does Intercom Fin check who it's talking to?

Before Fin looks up a customer's data, it can email a code to the address on file. The customer replies on the same email thread or types the code into chat, and gets three attempts (each code lasts 10 minutes). Nobody else here offers a ready-made check that works in both channels, so Fin gets the top score of 9 for checking customers who aren't signed in.

It stops short of 10 only because Intercom labels the feature a beta, and a customer with no email on file can't use it. The data lookup runs only on a correct code, the pass lasts 30 minutes, and a changed email triggers a fresh code. Those three rules earn a 9 for sharing only after a pass.

You turn the code on for each data lookup, and Intercom pitches its Procedures for identity verification. I find a setting per data lookup a little coarser than choosing by request type, so knowing when to check scores 8. Signed-in chat customers can be recognized once a developer secures Intercom's chat, and I give that an 8 as well.

When the code fails three times, expires or has no email to go to, a workflow can send the customer down a failure route. Fin hands over when a customer clearly asks for a person, but Procedures can't write an automatic summary note on handover, and with no human routing target set up Fin offers no handover at all. Those gaps hold handing failed checks to a person at 7. In a trial, I check the routing target first.

Intercom's articles describe approval answers recorded in the conversation and logs for each data lookup, but not a record of whether each code check passed. That puts showing what was checked at 6. Setup scores 7: the code check is an admin setting and Procedures are written in plain language, but the signed-in path needs a developer.

Capabilities shipped (out of 6)

JobIntercom Fin
Knows which requests need a check✅ code switched on per data lookup
Recognizes a customer already signed in✅ signed-in chat, developer setup
Checks a customer who isn't signed in✅ emailed code in email and chat, three tries
Shares or changes nothing until the check passes✅ data only after a correct code
Hands failed or suspicious checks to a person✅ failure route, escalation rules
Team can see who was checked⚠️ approvals recorded, no check record shown

Who's using Intercom Fin for account requests?

Of the Intercom stories I read, Rocket Money comes closest to account work. It started Fin on 10% of conversations and widened it to carefully scoped jobs, including account access requests. Fin now resolves 68% of the conversations it's involved in. At Fundrise, Fin resolves 50% of support volume.

What does Intercom Fin cost at 1,000 account tickets a month?

Intercom prices Fin at $0.99 per outcome, and a completed Procedure counts as an outcome even when it ends in a handover. By my reading, that means a failed check that ends with a person can still be billed. If Fin finishes 700 of the 1,000 tickets, that's $693 a month, rising to $990 if every handed-over ticket also completed a Procedure. Intercom seats come on top, as they would for any AI you run in Intercom.

Fin's bill goes up as it closes more tickets, so cost scores 6. It's published and self-serve, but it's the highest published bill here after Zendesk's.

For more on Intercom Fin, read our complete guide to Fin, browse the Intercom Fin alternatives roundup or see how it compares with My AskAI.

Can Fini verify customer identity?

Fini is built for fintech support teams and works alongside helpdesks including Zendesk, Intercom, Gorgias and HubSpot. Its finance page covers voice, chat and email.

The Fini homepage hero, pitching a self-learning AI agent that Fini says resolves 90% of support tickets, with a Book a demo button and resolution-rate, accuracy, language and monthly-resolution figures.

How does Fini check who it's talking to?

Fini's finance page promises a stronger identity check before any sensitive action, every time. I like how clear that rule is, and it scores 8 on knowing when to check, a point behind My AskAI because Fini's public pages don't show how you pick which requests need one.

Signed-in customers can be recognized in Fini's chat widget, and Fini can look a customer up in your helpdesk by email, so recognizing signed-in customers earns an 8. For customers who aren't signed in, the Atlas story lists identity checks before account changes, and Fini says it can fully automate a phone-number change once the customer passes. The public pages show the checks happening, though not what the customer is asked, which puts checking customers who aren't signed in at 7. A phone-number change is the one I keep with a person longest, so ask Fini's team how it decides a pass is strong enough.

I score sharing only after a pass at 9, because in the Atlas story any mismatch goes to human review. Fini also hands over with the full context, so handing failed checks to a person scores 8. Its public pages describe escalation on a mismatch but not a route for a customer who simply asks for a person.

Its record is the most detailed in the set. Fini says every action is logged, including the data pulled, the policy applied and the step taken, so its 9 for showing what was checked is the top score for that job. In a dispute, the log lets your team show the customer which data the AI pulled and which step it took.

Setup scores 5. Fini says customers go live in 14 days and are fully autonomous in 30, but I marked it down because the rollout is demo-led and signed-in widget users need a developer.

Capabilities shipped (out of 6)

JobFini
Knows which requests need a check✅ stronger check before any sensitive action
Recognizes a customer already signed in✅ signed-in widget, helpdesk lookup by email
Checks a customer who isn't signed in✅ identity checks before account changes
Shares or changes nothing until the check passes✅ mismatches go to human review
Hands failed or suspicious checks to a person✅ escalates with the full context
Team can see who was checked✅ every action logged

Who's using Fini for account requests?

Atlas, a fintech, went from 15% to 70% automation with Fini across its main support journeys, including identity checks and account changes. Wefunder uses Fini for account work such as password resets and two-factor fixes.

What does Fini cost at 1,000 account tickets a month?

Fini's pricing page says each plan includes a monthly allowance of tickets, well above 700 on Growth, and Growth charges $0.89 for each resolved ticket past it.

So at this volume, what you'd pay is the Growth plan's monthly fee. Cost scores 5 because the page doesn't print that fee, so ask Fini for it before you compare.

For more on Fini, read our complete guide to Fini or browse the Fini alternatives roundup.

Can Chatbase verify customer identity?

Chatbase is an all-in-one AI chat tool with a chat widget and helpdesk of its own, so adopting it means running chat in Chatbase. For a small team with no helpdesk yet, that's often a plus, and it can hand tickets to Zendesk, Freshdesk, Gorgias, Intercom and HubSpot.

The Chatbase homepage, pitching conversational AI agents that resolve issues across chat, email and voice, with a Procedures panel showing Order return and refund, Order status lookup and Upgrade plan switched on and Cancel subscription switched off.

How does Chatbase check who it's talking to?

Chatbase's Identity Verification feature lets the AI know a customer who is logged in to your website, once your developer connects it. That scores 8, level with most of the set. I think it's the simplest version of that job here.

For built-in shop actions you mark each one as signed-in only or open to anyone, which scores 7 on knowing when to check (a point below Fini and Intercom Fin, because anything outside those actions needs a custom procedure). A customer who isn't signed in can look up a shop order with their checkout email or phone and order number. Any other check is a procedure you build, so checking customers who aren't signed in scores 5.

Payment actions require a signed-in customer, so sharing only after a pass scores 8. Those actions only work in its website chat, which keeps it off 9.

A customer can ask for a person and the AI adds a summary to the ticket (a teammate can also take over and stop the AI). A handover on a failed check is something you build into the procedure, so handing failed checks to a person scores 6. Showing what was checked gets a 7: the activity log filters by customer and by action result (full audit logs are Enterprise-only).

Setup scores 6: it's self-serve and quick to start, but the signed-in check needs a developer and you're moving chat into a new widget. You can also show the AI to a percentage of website visitors, so you can roll it out gradually.

Capabilities shipped (out of 6)

JobChatbase
Knows which requests need a check✅ each shop action marked signed-in only or anyone
Recognizes a customer already signed in✅ website sign-in, developer setup
Checks a customer who isn't signed in⚠️ guest order lookups, custom procedures
Shares or changes nothing until the check passes✅ payment actions need a signed-in customer
Hands failed or suspicious checks to a person⚠️ ask-for-a-person trigger and summary
Team can see who was checked✅ activity log by customer and action

Who's using Chatbase for account requests?

Aplazo, a buy-now-pay-later provider in Mexico, uses Chatbase to sign up merchants, and half of the inbound merchants it wins now come through it. Jumia used it to scale support for its network of independent sales agents across eight African markets. Both stories are about sales and agent support, and neither shows the checks I scored above.

What does Chatbase cost at 1,000 account tickets a month?

Chatbase's monthly plans are Hobby at $40 for 700 credits, Standard at $150 for 4,000 and Pro at $500 for 15,000. Standard is the lowest plan with Chatbase's helpdesk and the payment actions scored above. I've priced it on a model that uses one credit per reply, so 1,000 tickets at two replies each use 2,000 credits, well inside Standard's 4,000.

A pricier model uses two credits per reply, which takes the same volume to 4,000 credits, still $150. Every extra call the AI makes to run an action costs more credits, and past 4,000 each 1,000 costs $40, so the bill can climb. It's the lowest published bill here, so cost scores 9, and that climb keeps it off 10.

For more on Chatbase, read our complete guide to Chatbase, browse the Chatbase alternatives roundup or see how it compares with My AskAI.

Can Ada verify customer identity?

Ada is an enterprise AI agent that covers chat, messaging apps, email and voice, and hands over to helpdesks including Zendesk, Gorgias and Freshchat. That breadth makes it a strong fit if your customers reach you on the phone and in messaging apps as well as in chat.

The Ada homepage hero, titled 'The agentic customer experience platform', with a 'Speak to an expert' button.

How does Ada check who it's talking to?

Ada's sign-in docs say that sign-in confirms the customer is who they claim to be, and that it should happen before the AI shares any sensitive data. Sharing only after a pass scores 9 for that rule. The same docs tell you to ask for sign-in where data is sensitive or the account changes, which scores 8 on knowing when to check.

Once a customer signs in, Ada's tools act as that customer, so the AI can reach nothing in the account that the customer couldn't reach themselves. I give that an 8 on recognizing signed-in customers.

For a customer who isn't signed in, Ada's Playbook docs show a verify-identity step that asks for the account email, passes it to a check you connect to your own system and branches on the result. I like the branching, but you build that check yourself, and the docs example asks for nothing beyond an email address, which on its own is weak proof. That puts checking customers who aren't signed in at 5, level with Chatbase and one below My AskAI, whose example asks for two details.

When sign-in fails, Ada gives an error answer, and a failed tool step hands the conversation off. Ask to see that failure path in a demo before you sign.

Handover summaries only exist on some helpdesks, so handing failed checks to a person scores 7. Ada's trust page names audit logging, but for a dispute I want a record of the check inside the conversation, and its docs describe none, so showing what was checked scores 6.

Setup scores 5, because it's an enterprise implementation and the sign-in step needs a developer on your side.

Capabilities shipped (out of 6)

JobAda
Knows which requests need a check✅ sign-in where data is sensitive or the account changes
Recognizes a customer already signed in✅ sign-in step, tools act as the customer
Checks a customer who isn't signed in⚠️ Playbook step you build, email-only example
Shares or changes nothing until the check passes✅ sign-in before any sensitive data
Hands failed or suspicious checks to a person✅ error answer and handoff, summaries on some helpdesks
Team can see who was checked⚠️ audit logging named, no check record shown

Who's using Ada for account requests?

Brigit, a financial health app, reports 40% ticket resolution with Ada. Its case study describes how the team built trust in the AI before letting it answer customers.

What does Ada cost at 1,000 account tickets a month?

Ada's pricing page publishes no prices and leads to a demo booking. One procurement marketplace puts the median Ada contract at $72,000 a year, which is well above anything else priced here. Sales-only pricing with no monthly figure scores 2 on cost.

For more on Ada, read our complete guide to Ada, browse the Ada alternatives roundup or see how it compares with My AskAI.

Can Zendesk AI agents verify customer identity?

Zendesk AI agents run inside Zendesk and are included on every Suite plan from Suite Team, with voice available too. If your team already runs Zendesk, they're the AI built into the tool you know.

The Zendesk AI agents page hero, titled 'Self-improving AI agents built for resolution', with a 'Start your free trial' button and a 14-day free trial note.

How do Zendesk AI agents check who they're talking to?

Once a customer is signed in through your site or app, Zendesk's AI agents don't ask for their name or email again, and your agents see a green check mark beside each message they send. With that check mark, an agent picking up the ticket can see at a glance which customers are signed in (and which still need a check). It earns the 8 on recognizing signed-in customers and the 8 on showing what was checked, both held there because the sign-in only covers Zendesk's web chat and mobile app messaging.

For a customer who isn't signed in, Zendesk's answer is to get customers signed in. Checking customers who aren't signed in scores 4, the lowest in the set. If a lot of your account requests arrive by email, I think that leaves your team checking most of them by hand. AI answers from restricted help articles are shown only to signed-in customers, which scores 7 on sharing only after a pass.

Generative procedures can hand off to a shared identity verification step. Your admins decide which requests need a check by building flows (a job for someone comfortable in Zendesk's builder), so knowing when to check gets a 6. Escalation flows are yours to design and can be removed entirely, and Zendesk's escalation articles don't mention a handover summary, so handing failed checks to a person also scores 6.

Setup scores 6, because it needs both an admin in Zendesk and a developer to connect your sign-in to Zendesk's messaging.

Capabilities shipped (out of 6)

JobZendesk AI agents
Knows which requests need a check⚠️ procedures hand off to a shared check step
Recognizes a customer already signed in✅ web and app chat, not asked again
Checks a customer who isn't signed in❌ relies on the customer being signed in
Shares or changes nothing until the check passes✅ restricted answers only for signed-in customers
Hands failed or suspicious checks to a person⚠️ escalation flows you design, no summary shown
Team can see who was checked✅ green check mark beside signed-in messages

Who's using Zendesk AI agents for account requests?

Babbel uses Zendesk AI agents and reports an AI automation rate above 50% across its general support volume.

What do Zendesk AI agents cost at 1,000 account tickets a month?

Zendesk bills per automated resolution on top of Suite seats, at $2.00 pay-as-you-go or $1.50 on a commitment, with 5 included per agent each month on Suite Team. I've modeled a five-agent team: 700 resolutions minus 25 included is 675 billable, or $1,350 a month pay-as-you-go. That's the highest published bill here, and it grows as the AI resolves more, so cost scores 4.

For more on Zendesk AI agents, read our complete guide, browse the Zendesk AI alternatives roundup or see how it compares with My AskAI.

Can Decagon verify customer identity?

Decagon is an enterprise AI agent for chat, email and voice that works with helpdesks including Zendesk and Intercom. Its product overview lists identity verification among the critical operations it runs securely and with strict controls.

The Decagon homepage hero, titled 'The AI concierge for every customer', with a work email field and a 'Get a demo' button.

How does Decagon check who it's talking to?

As I read its pages, Decagon runs checks through rules you set with its team, including rules for escalation across chat, voice and email. That scores 7 on knowing when to check, below Fini and Ada, because the public pages describe the controls in general terms.

Its security page describes flexible authentication methods for voice, and short-lived access to your systems for each session. I look for that second part in any tool with account access, because it means the AI can't keep reaching into an account once the conversation is over. The public pages don't show how a signed-in chat customer is recognized or how a chat or email customer is checked, so Decagon scores 6 on both. Its finance page stresses validation and auditability, which puts sharing only after a pass at 7.

Decagon hands over to a person, and on voice it passes the call across with a short summary. That's worth an 8 for handing failed checks to a person. You can trace the agent's reasoning and review its behavior afterwards, so showing what was checked gets a 7. I think that trace is useful when a customer disputes what the AI did.

Setup scores 4, the lowest here, because implementation is demo-led and built with Decagon's team.

Capabilities shipped (out of 6)

JobDecagon
Knows which requests need a check✅ identity verification under strict controls
Recognizes a customer already signed in⚠️ per-session access, chat sign-in not shown
Checks a customer who isn't signed in⚠️ voice authentication, chat and email not shown
Shares or changes nothing until the check passes✅ strict controls, validation
Hands failed or suspicious checks to a person✅ escalation, voice transfer with summary
Team can see who was checked✅ reasoning trace and review

Who's using Decagon for account requests?

Chime, a fintech, reports 70% resolution across chat and voice with Decagon. Its finance page lists jobs from password resets and balance questions to fraud alerts and disputes.

What does Decagon cost at 1,000 account tickets a month?

Decagon's own post on pricing describes charging per conversation or per resolution without giving a rate. Like Ada, it scores 2 on cost.

For more on Decagon, read our complete guide to Decagon, browse the Decagon alternatives roundup or see how it compares with My AskAI.

What does automating identity checks save you? (worked example)

The basis for every row is 1,000 account tickets a month, meaning requests where the account holder has to be confirmed before the answer or the change. I've assumed 600 arrive by chat and 400 by email, each with two AI replies, a typical ticket. The AI finishes 700 and 300 go to a person, such as failed checks and contact-detail changes.

For the no-AI row, I've assumed six minutes of team time per ticket for a team that already uses AI to help where a person does the work. That's priced at $0.40 a minute, which allows for overheads on top of the median US wage for customer service representatives of $21.53 an hour.

ScenarioPlan modeledMonthly costEffective $/ticketNotes
No AIYour team handles every ticket$2,400$2.401,000 tickets × 6 minutes × $0.40
My AskAIPro$263$0.26$199 + 200 extra credits at $0.12, plus $40 of Task replies; 30-day free trial
Intercom FinPer outcome$693 to $990$0.69 to $0.99700 outcomes at $0.99, up to 1,000 if every handover completes a Procedure; seats extra
Zendesk AI agentsSuite Team, pay-as-you-go$1,350$1.35675 resolutions at $2.00 after 25 included for five agents; seats extra
FiniGrowthGrowth plan fee-700 resolved tickets fall inside Growth's monthly allowance; $0.89 per resolved ticket past it; fee not printed on the pricing page
DecagonSales onlyNot published-Pricing by quote
AdaSales onlyNot published-Pricing by quote
ChatbaseStandard$150$0.152,000 credits on a one-credit model, 4,000 on a two-credit model; lowest plan with its helpdesk and payment actions

On these assumptions, every AI row also leaves your team 300 tickets to handle, about $720 a month of team time at the same rate, so it doesn't change the order. Without AI, part of the six minutes each ticket takes goes on confirming who is asking before any real work starts.

On the per-outcome and per-resolution rows, the bill goes up with every account ticket the AI closes. My AskAI charges per ticket whether the AI finishes it or hands it over, and add-ons such as Tasks bill on top on every plan, which is why our row includes the $40 of Task replies. Our ROI calculator runs the same sums on your own volume.

So which AI customer service agent is best for verifying customer identity?

We top the scoreboard on 63. Our AI answers the questions that need no check, puts the check in front of the requests you name, and sends every failure to a person with a summary, all inside Zendesk, Intercom, Freshdesk, Freshchat, Gorgias or HubSpot.

Intercom Fin, on 60, is the pick if you want a code check today and many of your account requests arrive by email. Fini, on 59, suits a fintech team that wants the vendor to run a stronger check before every sensitive action and log every step. Chatbase is a sound, low-cost start for a small team with no helpdesk yet. If most of your account requests come by phone, Decagon and Ada are the ones to look at.

Whichever tool you choose, I'd roll it out in this order:

  1. Let the AI answer the account questions that need no check, such as policies and order status.
  2. Let it answer signed-in customers from their account details.
  3. Test the check out of customers' sight first, so your team reads every proposed step before a customer does.
  4. Keep contact-detail changes with a person until you trust the checks.
  5. Send every failed check to a person, always.

The iGaming operator we work with followed the same idea and rolled the AI out progressively on top of live player data. The digital-goods marketplace gives each common ticket type its own Task. To see how a Task checks a customer before it acts, try My AskAI free for 30 days on your own account requests.

FAQs

How do you verify a customer's identity?

Check what the person tells you against details already on your records, or send a code to the email or phone already on the account. Security questions on their own are weak, because the answers are often easy to guess or find. In my view, you should only check where the request discloses or changes something, so everyday questions stay quick.

Can an AI agent verify a customer's identity by itself, or does the customer need to be logged in?

Both are possible. Most tools here can recognize a signed-in customer once your developer connects your sign-in, and we also treat an email from the address on the account as verified. For customers who aren't signed in, Intercom Fin emails a code, Ada's Playbooks can ask for the account email and pass it to a check you build, Fini runs identity checks before account changes, and in My AskAI we run a Task that checks the details against your records.

Which requests actually need a check?

Order status, opening hours and policy questions need no check. Anything that reveals private account details or changes the account does, such as billing or card details, a withdrawal, or a new email, phone number or address. Changes to contact details need the strongest check or a person (I treat them as the riskiest request of all), because they're the usual route into a takeover.

What does the AI do when a check fails, and when does a person take over?

A good setup stops and hands the conversation to a person with a summary, without making the customer start again. Intercom Fin sends the customer down a failure route after three wrong codes, Ada gives an error answer on a failed sign-in, and Fini sends mismatches to human review. In My AskAI, we let a Task hand over at any step, and customers can ask to speak to a person at any time.

Which AI agents let staff approve a sensitive account change before it goes through?

In this set, My AskAI does it per action: each account action in a Task can run automatically or wait for a teammate to approve it, and you choose which. Fini sends any mismatch in a check to human review before the account changes.

Whichever tool you pick, keep changes to contact details with a person until you trust the checks, because they're the usual route into a takeover.

Can an AI support agent handle KYC and identity verification questions?

Yes, for the questions around a check, such as what's still needed and why an account is blocked. A digital-goods marketplace on My AskAI uses a named Task for accounts blocked by a verification check, and the Task gives each customer the right answer for their case.

I'd keep the decision on the check itself with your KYC provider or compliance team. Our roundup of fraud detection AI agents scores six tools on how they handle stuck verification checks.

Can it verify someone over email as well as chat?

Yes, with the right tool. Intercom Fin's code works by replying on the same email thread, and we treat an email from the address on the account as verified. Zendesk's sign-in covers its web chat and mobile app messaging, and Chatbase's payment actions only work in its website chat.

How to verify customer identity over the phone?

The same six jobs apply on the phone, and one contact-center glossary describes passive checks that confirm a caller without asking them anything. In this set, Decagon, Ada, Fini, Intercom Fin, Zendesk and Chatbase all show a voice channel on their pages.

What are the best practices for customer authentication?

Seven that cover most support teams:

  • Check only where the request discloses or changes something.
  • Confirm every name, email address or order number against your records.
  • Don't rely on security questions alone.
  • Let one pass last the whole conversation.
  • Treat contact-detail changes as the riskiest request.
  • Send failed checks to a person with a summary.
  • Keep a record of every check and what the AI did.