There is no published ranking of leading AI agents for customer support. The label means nothing until you know what the vendor actually does with your customer data. That is what this page is for.
Before an AI agent reads a single customer message, there are five questions your procurement checklist must clear. Each one has a clear answer from a trustworthy vendor and a vague answer from one that does not want you to ask.
Who sets the transcript retention period, and who can shorten it
Every AI agent that reads customer conversations produces a transcript. That transcript is a record of personal data. Under most privacy regimes, you are responsible for what you keep and how long you keep it.
The question to ask is simple: who decides when a transcript is deleted?
A vague answer sounds like this: "We retain data securely." That tells you nothing. It does not say who decides, how long, or what happens when a customer asks to be forgotten.
An acceptable answer names a specific control. At Asyntai, chat retention is configurable, and there is an auto-delete option. That means your team sets the window, and it runs automatically without anyone remembering to act. The deletion applies to the transcript, not to whatever your CRM recorded during the same conversation. Both need to be addressed.
What a vague answer hides: a vendor that keeps transcripts indefinitely because they are useful for training future models. That is not a feature. That is your customer data being used without a clear disclosure or consent mechanism.
Your procurement checklist before an AI agent touches a customer record
Ask each question. A vague answer means you are not meant to ask it again.
Who inside your company can read a transcript
The transcript of a customer support chat often contains information the customer did not mean for your whole organisation to see. A password reset request. A complaint about a colleague. A health question routed to the wrong department. The transcript is sensitive, and access to it needs to follow the same rules as access to any other customer record.
Ask: does the platform let you restrict who can read a transcript, and can those restrictions be set per website or per tab?
A vague answer: "We take security seriously."
An acceptable answer describes role-based access or per-tab permissions. At Asyntai, access control is per-tab and scoped per website. That means a support agent working on your orders system can read transcripts from your orders tab without gaining access to transcripts from your returns tab. A manager can see across all of them. Those permissions are set in the platform, not left to whatever your internal systems allow.
If the answer does not distinguish between what different people can see, the platform treats all your team members as equally trusted. That is a data governance problem, not just a UX issue.
What a deletion request actually deletes
When a customer asks for their data to be removed, the clock starts. The question is what the vendor actually removes, and how you can verify it happened.
Ask: when a deletion request is made, does the vendor remove the transcript, and is there an audit entry recording that removal?
A vague answer: "We comply with all deletion requests."
An acceptable answer involves a deletion log, a confirmation that the transcript is gone from the agent's own storage, and an acknowledgement that this is separate from what your CRM or helpdesk system retained during the same session. No single vendor controls all the places customer data lands. The vendor should be clear about what it controls and what it does not.
Asyntai's audit log is append-only. Entries cannot be edited or deleted. That means if a transcript was created, read, and removed, there is a permanent record of each step. You can show an auditor exactly what happened and when. That is what a deletion request should produce: not just a promise, but a traceable outcome.
Where does the model run
An AI agent that processes customer messages does so by sending that data to a language model. The model runs somewhere. That location determines which data sovereignty rules apply, and who has practical access to your customer data during processing.
Ask: where does the model run, and is the data processed in a specific jurisdiction?
A vague answer: "In the cloud."
What counts as acceptable depends on your regulatory context. If your customers are European, you need to know whether data leaves the European Economic Area. If you are in a regulated industry, you need to know whether the processing location satisfies your contractual and legal obligations. A vendor that cannot name the region is a vendor that has not thought about this, or does not want you to.
Beyond the location itself, ask what happens to the data after the model produces a response. Does the vendor retain the input for any purpose? Can it be used to improve future models? These are the questions that deletion and retention policies answer. They only matter if you ask them.
What the audit trail records
Every interaction between a customer and an AI agent needs to be accountable. When something goes wrong, you need to be able to reconstruct what was said, what the agent did, and whether the right human was looped in at the right time.
Ask: is the audit log append-only, and can entries be edited or deleted after the fact?
An append-only log is one where every event is timestamped and permanently recorded. Nothing is removed, nothing is changed. That is the only kind of audit log that can be trusted, because a log that can be edited is a log that can be sanitised.
Asyntai's audit log is append-only. Entries cannot be edited or deleted. Every chat session, every access event, every configuration change is recorded permanently. When you need to demonstrate what happened, the log has not been optimised for that presentation. It is a raw record.
A vague answer: "We maintain detailed logs of all activity." That sounds reassuring until you ask whether those logs can be altered. If the answer is anything other than no, the log is not an audit trail. It is a convenience feature.
PII redaction and output controls
Two more controls sit alongside the audit trail. Both matter if your support agents are working alongside an AI, or if the AI is answering customers directly.
PII redaction removes sensitive personal data from what the agent can see. Names, phone numbers, email addresses, national insurance numbers: all of these can be stripped from a transcript before the model processes it. That reduces what the model sees, what could theoretically be retained, and what appears in any logs.
Ask: can you redact PII at the input stage, and is that configurable per website or per conversation type?
At Asyntai, PII redaction is configurable. You decide what gets redacted and where. That is not a one-size-fits-all toggle. A customer support conversation about an order does not need the same redaction rules as a complaint about a billing error.
Output controls govern what the AI agent is allowed to say. This is distinct from a prompt rule. A prompt rule lives in the instruction you give the model. An output classifier runs after the model has produced its response and checks whether it contains anything it should not: a refund amount that was not approved, a customer email address, a policy statement that is out of date.
At Asyntai, output controls use a separate output classifier rather than a prompt rule. That means the check is structural and consistent, not dependent on how carefully the prompt was written on any given day.
How AI agents help platforms deliver faster and more consistent customer support
The reason support teams buy AI agents is not to tick a data governance checkbox. It is to help their platforms deliver faster and more consistent support outcomes for every customer. Those two goals are distinct and they pull in different directions.
Faster support means reducing the time between a customer asking a question and receiving a useful answer. More consistent support means that the second customer with the same question receives the same answer, or a better one, rather than whatever the available agent had time to type. Faster is a speed metric. Consistent is a quality metric. Most platforms can improve one at the expense of the other. The best AI agents improve both.
They do this by removing the bottlenecks that slow human agents down. When a customer asks about a policy, an order status, or a return procedure, the answer exists somewhere on your website. The AI agent retrieves it, formats it, and delivers it without routing the question to a human queue. That is where faster comes from: questions answered in seconds instead of minutes, without a human in the loop at all.
Consistency comes from the same mechanism. A human agent answers the same question differently on a busy afternoon than on a quiet morning. They interpret a policy slightly differently when they are tired. They miss an update that came out last week. An AI agent trained on your current content answers every instance of the same question from the same source. That is the structural difference between a human queue and an automated one, and it is why consistency improvements from AI agents tend to be large rather than marginal.
The governance controls described on this page are what make those consistency improvements auditable. When the AI agent delivers an answer, the platform records who asked, what was asked, and what was answered. When a customer raises a complaint about an AI response, you can reconstruct exactly what happened. That is the accountability layer that turns an AI agent from a cost-cutting tool into a reliable part of your support operation.
Volume, cost per message, and what you are actually buying
When support leads evaluate AI agents, the conversation about cost often obscures the conversation about what you are actually paying for. Here is what the arithmetic looks like on Asyntai's self-serve tiers.
The message count is what the AI agent processes, not every chat your platform receives. A customer who gets an answer from your FAQ does not consume a message. A customer who types a question and gets an AI-generated response does. That distinction matters when you model your volume.
What you are buying is not just responses. You are buying the infrastructure that makes those responses accountable: the audit log, the access controls, the PII redaction, the output classifier. Those controls are not add-ons on the lower tiers. On Asyntai, they are available on every paid plan, starting with Starter. That means a small team running two agents on the Starter plan has the same data governance foundations as a larger team on Pro.
The free plan exists for evaluation. It carries no retention controls, no per-tab permissions, and no append-only audit log. Those are the controls you need in production. Evaluation environments and production environments have different requirements, and a vendor that bundles controls into an enterprise tier only is a vendor that expects you to pay enterprise prices to get basic governance.
What each plan includes
The controls described above are not all available on every tier. Here is where they sit in Asyntai's published plan structure.
| Control | Free | Starter / Standard / Pro | Enterprise | What it means in practice |
|---|---|---|---|---|
| Configurable retention and auto-delete | No | Yes | Yes | You set the window; it runs without manual action. |
| Per-tab permissions and per-website scoping | No | Yes | Yes | Different people see different transcripts, configured in the platform. |
| Append-only audit log | No | Yes | Yes | Every event is permanent and tamper-proof. |
| Configurable PII redaction | No | Yes | Yes | You decide what the model sees and where. |
| Separate output classifier | No | Yes | Yes | Output checked structurally, not relying on prompt wording alone. |
| OIDC single sign-on (Entra, Google) | No | Yes | Yes | Staff log in through your existing identity provider. |
The free plan is for evaluation. Production use with customer data requires Starter or above. Enterprise includes a custom contract with a negotiated message allowance and additional controls negotiated per engagement.
These controls are the minimum viable set for any organisation that processes customer data through an AI agent. They are not optional extras. Without them, you cannot demonstrate what the agent did, who saw it, when it was deleted, or what the customer was asked to consent to.
The question you should be asking instead
Forget "leading." The question that actually decides whether an AI agent is right for your support operation is: can you answer every one of the five questions above with a specific, verifiable answer?
If the vendor will not tell you where the model runs, do not buy it.
If the audit log can be edited, it is not an audit log.
If you cannot set your own retention period and have it run automatically, you are relying on the vendor's interest in your data to determine how long it is kept.
If per-tab permissions are not available, every user in your organisation has access to every transcript. That includes the ones containing complaints about colleagues, health queries, and financial disputes. Support teams do not need that access, and your customers did not grant it.
A vendor that answers all five questions clearly has thought about data governance. A vendor that gives vague answers has not, or is avoiding the question deliberately. Neither is a foundation for a product that handles your customer relationships.
Your procurement checklist
Before you sign a contract with any AI agent vendor, get the following in writing. If they will not provide it, do not interpret that as a negotiating position. Interpret it as the answer.
- Who sets the transcript retention period, and what happens when it expires?
- Can you restrict which team members can read which transcripts, and how is that enforced?
- When a deletion request arrives, what specifically is deleted, and what confirmation can the vendor provide?
- Where does the model run, and does the vendor's data processing agreement match your regulatory requirements?
- Is the audit log append-only, and can any entry be edited or deleted after the fact?
If you have those answers in writing before you sign anything, you have done more due diligence than most procurement processes manage. If the vendor resists providing them, treat that resistance as the answer.
Asyntai publishes its controls and which plan carries each. Start at asyntai.com or talk to the team about the Enterprise plan if your procurement process requires a specific data processing agreement.
Sources and further reading, September 2025.
- Asyntai published capabilities and plan structure, asyntai.com pricing page.
- Asyntai audit log, access control, PII redaction, and output classifier specifications, asyntai.com features page.
- OIDC single sign-on documentation, asyntai.com integrations page.