ChatBeacon

Without the models, a complete live chat, cobrowse, video and text messaging platform remains. That is what SmartMax sold as SightMax from 2005 and as ChatBeacon from 2017, when its first bots ran through a third party bot service rather than its own models.

The AI line is real and shipped: a visitor facing assistant, a recall feature that answers from past resolved chats and escalation routing, plus a staff drafting assistant and a multi agent builder, released together as the 2026 AIX platform. It is also priced as an optional annual add on to the seat license, and the vendor's own pricing page states that fully air gapped installs run the platform without the AI layer. Models were added to a communications product rather than being the product.

The oversight toolkit is well described, and the question is which path gets it. The trust pages name the escalation triggers (a direct request for a person, a low confidence response, a conversation leaving its approved scope, defined topics and workflow conditions), carry the transcript and customer context to the live agent, and restrict who may change AI configuration. The staff assistant drafts replies that a person sends or edits, so the representative stays the author of record.

The visitor facing path is looser. The entry AI tier is sold as an agent that answers, retrieves and resolves on its own, drawing on past chats, a vector store and the live web, and the pricing page lists workflow input and output guardrails only in the higher multi agent tier. The vendor's own pricing and AI facts pages also disagree on whether escalation is included in the entry tier. Every threshold is set by the institution with no stated default, and testing responsibility is assigned to the customer.

The recall feature carries the most useful model risk property in the product: every answer it gives carries a confidence match score and cites the past conversations it drew from, so a reviewer can trace a response to its source. That is auditable in a way a free generating assistant is not, and the confidence score also drives handoff to a person. The same design raises the central model risk question.

Recall treats resolved chats as proven answers and is described as always learning, so a rate, fee or policy answered correctly last year is recalled as verified this year. The citation shows where an answer came from, not whether it is still true, and nothing describes expiry, review or retirement of recalled answers. Performance evidence is thin: the resolution figure appears four ways on the vendor's own pages, and no accuracy or false resolution rate is published.

ChatBeacon publishes plenty of numbers and little evidence behind them. Its figure for the share of conversations the AI resolves without a person appears four ways on its own live pages. A banking customer story tile that links nowhere says 43 percent. The credit union page says 68 percent and a banking dashboard 73 percent, both labeled illustrative, and a page addressed to AI systems as current and authoritative says 87 percent.

The one named institution, Blue Ridge Bank, is quoted through a job title with no named person. That name is shared by at least two unrelated United States banks, and an earlier search indexed version of the same page attributed a different quote under the same title to an unnamed regional credit union. The credit union testimonial is labeled on the page as an illustrative composite.

The vendor claims more than 1,200 customer teams and a 4.6 rating on G2, where the visible reviews date from 2020 and describe the live chat product before the AI line existed. No institution names ChatBeacon in its own announcement, and no named individual reports a measured outcome.

The pooled data question gets a direct answer. SmartMax states it does not use customer conversations to train ChatBeacon AI models, that only the content needed to generate a response goes to the configured provider when AI is enabled, and that disabling AI stops that flow. Answers are grounded in institution approved instructions, documents and a vector store, and knowledge can be scoped per workflow. Provider side retention is left to the provider's own terms.

Two inconsistencies matter. The training promise is worded three ways across the site. The binding security page carries the full version, while a credit union page promises only not to train public models and an AI facts page promises not to train models for other companies. The security page wording is the one to contract to.

And the recall feature page tells visitors to expect no hallucinations and no guessing, while the vendor's own AI governance page states that AI generated content may be incomplete, outdated or incorrect. The trust page is the candid one.

A public service privacy notice divides responsibilities between SmartMax and the institution. Retention is configurable, transcripts can be exported and deleted, a data processing addendum is offered, and the vendor states it does not sell conversation data. The cobrowse answers are specific: sessions are started or approved by the visitor and limited to the active session, and designated fields such as Social Security numbers and passwords are masked from staff. Two gaps remain.

The subprocessor list is available only on request. And the credit union page answers the question of whether ChatBeacon meets NCUA and GLBA requirements with an unqualified yes, which no vendor can say, since those obligations sit with the institution and the credit union regulator does not certify vendors. The vendor's own AI governance page makes exactly that point, stating that using ChatBeacon does not by itself satisfy an organization's compliance obligations.

ChatBeacon's own security page gets the attribution right. Its certification table carries a column naming who holds each credential, lists SOC reports, ISO 27001 and others as held by Microsoft Azure, and states that these do not represent a separate certification of ChatBeacon. SmartMax publishes a substantial control description behind that, covering encryption in transit and at rest, enterprise sign on under the institution's own multifactor policies, and role based permissions.

Audit logging, peer reviewed releases, separated environments and incident notification are covered too, and it supports standard security questionnaires. No attestation held by SmartMax itself is public, though its security team offers documentation on request. The marketing pages have not caught up.

The credit union page places SOC 2 Type II via Microsoft Azure over the words independently audited security controls, the about page lists SOC 2 Type II and HIPAA with no holder at all, and the pricing table lists SOC 2 and HIPAA compliance as a plan feature. An earlier search indexed version of the banking page read SOC 2 Type II and HIPAA compliant, so the correction is recent and partial.

SmartMax is a software supplier and holds no financial license, which fits its shape. There is no public sandbox participation, program admission or supervisory relationship. The one regulatory position the marketing states is overstated: the credit union page answers yes to meeting NCUA and GLBA requirements. The trust center is more careful, assigning AI disclosure, prohibited use and legal compliance to the institution. One product behavior stands out.

The banking page's own demonstration shows the assistant quoting a thirty year fixed mortgage rate and then routing the borrower to a loan officer, which places an automated agent inside a credit conversation before any application is taken. With a recall feature that answers from past chats, a quoted rate is only as current as the conversation it was recalled from, and nothing describes how rate and fee answers are kept current or reviewed.

ChatBeacon publishes no fairness testing, outcome monitoring or accessibility evaluation for its agents, and its governance page assigns the duty to avoid discriminatory uses to the customer. The exposure is specific in banking. The vendor's own banking use cases include mortgage prequalification, where the assistant collects borrower information and routes qualified prospects to a loan officer.

Fair lending rules protect prospective applicants as well as applicants, and a screen that decides who counts as qualified before a person is involved is where discouragement on a prohibited basis can occur without anyone seeing it. The institution sets the criteria, but nothing public describes how agent behavior in those conversations is tested for consistent treatment across customers.

The trust center is candid about where responsibility sits, and it sits with the institution. The AI governance page states that AI generated content may be incomplete, outdated or incorrect, that financial and other high impact uses need human review, and that the customer owns configuration, knowledge accuracy, notices and legal compliance. That is a clear allocation, not recourse. At least one competitor writes a guarantee against hallucinations into its contract.

ChatBeacon's recall page promises visitors no hallucinations in marketing copy, while its governance page disclaims the same risk. The promise sits in the marketing and the disclaimer sits in the governance page, which is the reverse of where a buyer needs them. No correction path for a member given a wrong answer, error rate commitment or liability term is published.

ChatBeacon names its model provider. The pricing page states every AI tier runs on the institution's own OpenAI account and key, billed by OpenAI directly, and the about page names the model generation the platform integrates, GPT-5.4. The account structure matters more than the name: the institution holds the provider contract itself, so the data use terms governing its conversations are its own rather than inherited through the vendor. The disclosure stops at the provider.

The security page also describes a ChatBeacon managed provider configuration without naming what it uses, and the live web source in the entry AI tier is not identified. No model is mapped to a workflow, and the subprocessor list is available only on request.

Integration with the systems a bank runs on is described in a single sentence. ChatBeacon connects through secure interfaces and webhooks, runs alongside common cores and digital banking platforms, and scopes the exact connection with each institution before launch. Not one core processor, digital banking platform or loan origination system is named. The 3,000 integration figure comes largely from a general automation connector marketplace and says nothing about banking depth.

What is specific is infrastructure fit rather than integration: identity through Microsoft Entra ID and Okta, text messaging through the institution's own Twilio number, and a self hosted build on Windows Server and SQL Server, a stack many credit union technology teams already run. There is no published list of supported cores or of the depth of each connection.

Deployment is disclosed in detail. The same product ships three ways with a published comparison of who controls what. The options are a multi tenant cloud on Microsoft Azure in a United States hosted environment, a single tenant dedicated cloud with an isolated database per customer in an agreed Azure region, and a perpetual license installed on the institution's own Windows servers and SQL Server database.

The vendor states a customer can move between cloud and self hosted by database restore rather than rebuild. Responsibility boundaries are written per model, separating what Microsoft secures, what SmartMax operates and what the institution controls. The pricing page discloses the AI data path plainly: model calls leave the network while transcripts stay in the institution's database, and fully air gapped installs run without the AI layer. That qualifies the self hosted card's line that nothing leaves the network. A region list is not published.

Price disclosure is close to complete. ChatBeacon publishes its cloud seat prices at $39 and $69 per operator per month with an annual discount, and its visitor facing and staff facing AI packages with setup fees from $1,599 a year. Every add on is priced, down to $7 a month for single sign on through Microsoft Entra ID.

The full perpetual license curve for installs on the institution's own servers is published too, from $1,599 for five operators to $18,999 for unlimited operators, with maintenance stated at 30 percent a year and the first year included. The billing basis is stated plainly: seats are counted by concurrent operators, and AI is a flat annual fee with no charge per conversation or per resolution.

One cost sits outside the price and the vendor says so: every AI tier runs on the institution's own OpenAI account, billed by OpenAI directly. Only the single tenant enterprise tier is quoted rather than listed.

Banks and credit unions are separately addressed segments with their own maintained pages. The credit union material is written for credit unions, referring to members and to examiner review rather than reusing banking copy, and insurers get a third page. The vendor's own history puts financial institutions first: its 2005 product counted banks and credit unions among its first customers, and its self hosted model was built for them. Two limits matter.

Financial services is one of nine vertical pages, beside healthcare, education, retail, car dealerships and recreational vehicles, and the same product is sold to all of them without a separately packaged banking application. And the number of financial institutions on the platform is not disclosed, so the 1,200 customer teams figure cannot be read as depth in this segment. The coverage described is community institutions in the United States, with no large bank, wealth or capital markets use. ChatBeacon asked to be included in this category.