P

Perspective AI

The model is the entire product. Adaptive conversation, follow up on vague answers and extraction that turns unstructured answers into structured fields are what the vendor sells. Remove the model and nothing remains, because the thing being replaced is the form. Perspective AI is a horizontal intake platform serving healthcare, financial services, insurance, education, real estate and internal operations as well as legal.

Its legal offering is specific: a dedicated legal use case names personal injury, family law, estate planning and immigration qualification, the vendor publishes an analysis of legal intake software that places itself in that market, and it describes capturing the parties and matter details needed to flag conflicts. Legal is one of several verticals rather than the company's center of gravity.

Nothing published, and the extraction claim is the thing that needed measuring. The vendor's central technical proposition is schema aligned data extraction, structuring unstructured conversational answers into defined fields, and it markets the output as validated structured outputs on which downstream automation triggers rather than guesswork from dropdowns. Extraction accuracy is directly measurable and no figure is published: no field level accuracy, no error rate, no evaluation, no statement of what happens when a conversational answer is ambiguous and the schema requires a value.

The word validated is doing significant work in the marketing and nothing states what validation means or who performs it. The consequence is specific: a misextracted field in legal intake produces a structured summary that reads as authoritative and routes a matter incorrectly, and the conversation it came from is the only place the error is visible.

Two published mechanisms put a person at the right points. First, configuration can be reviewed before deployment: the firm sets routing rules for strong and weak fits and can preview the full conversation before it goes live, so a lawyer can inspect exactly what the system will say to a prospective client rather than finding out afterward. Second, on conflicts, the vendor states that the platform captures the parties and matter details during the conversation and routes the summary to staff who confirm against the firm's own records.

That is an explicit statement that the machine gathers and a person decides on the one question where an automated determination would be indefensible. The qualification decision itself is autonomous. Routing strong and weak fits happens without a person, and nothing states what a weak fit outcome does to the inquiry or whether a borderline case is escalated.

Nothing located. No customer is named in any vertical, and no case study, usage figure, funding announcement, independent analyst placement, award or third party review of this vendor was found. The evidence is entirely self authored, though large in volume: an extensive blog publishing category analysis, buyer frameworks and comparative rankings across legal intake, insurance, education, real estate and internal request management, in several of which the vendor ranks itself first. Rankings a vendor publishes about itself are marketing rather than evidence.

Nothing located. No confidentiality statement, no encryption reference, no data handling description and no treatment of privilege or prospective client confidences was found. The exposure is heightened by the product's own design proposition: the vendor argues that forms lose the context that matters and that its conversations capture the why behind each inquiry, which means the system is built to elicit more sensitive detail from a prospective client than a form would, in named practice areas including family law and immigration.

It also operates across healthcare, where the same conversations capture presenting concerns and care needs. A platform designed to extract fuller disclosure publishes nothing about how that disclosure is protected.

Not located. The system holds an adaptive conversation with a prospective client, asks follow up questions about their legal situation, assesses fit against the firm's criteria and routes on that basis, all before any lawyer is involved. The vendor's own framing is that this happens before the paralegal picks up the phone. That is an automated system holding a substantive discussion about a person's legal problem and reaching a conclusion about it.

Nothing published states that the conversation is not legal advice, describes what the agent may say when a prospective client asks a substantive legal question during the conversation, or discloses to the person that they are speaking to an AI. The interaction is conversational and adaptive rather than form based.

Nothing is published on governance. The platform qualifies inquiries against fit criteria the firm defines and routes strong fits differently from weak ones, so it decides which people reach a professional. That includes immigration and family law, where the people being assessed are often in difficulty and have few alternatives. The adaptive design adds to the risk. Because the conversation branches on responses, two people in similar situations can be asked different questions and produce different summaries depending on how articulately they answer early, and nothing indicates whether performance across different groups of respondents has been examined.

No AI policy, model card, bias or fairness testing, evaluation, monitoring, governance body, ISO 42001 or EU AI Act positioning is published.

No stewardship position is published. Nothing states whether intake conversations, extracted data or routing outcomes are used to train or improve models, no retention period is published, and no deletion right is described. The horizontal footprint widens the omission. The same platform captures presenting concerns and care needs in healthcare intake and financial position in wealth management qualification, so one undisclosed training and retention posture covers health information, financial information and legal matter details together.

The free tier with no code setup adds to this, since anyone can deploy an agent collecting such material with no published statement about where it goes.

No published position is published on liability for AI output, warranty, service levels or remedy. Two exposures are specific to this product. A misextraction produces a structured summary that a firm acts on without seeing the conversation behind it, and the vendor markets automation triggered by those outputs, so an error spreads into routing and workflow before anyone reads it. And a person qualified as a weak fit is routed away without anyone learning it happened.

The vendor's own marketing frames the problem it solves as teams moving forward on incomplete information and making weaker decisions, which describes the risk its extraction would create if wrong, and no position addresses it.

Destinations are named by type and no system is named individually. Structured summaries go to an inbox, a CRM or Slack, and the agent can run as a website widget, popup, email link or standalone page, so both the output path and the input surface are described. Slack is the only product named. What a law firm would check is absent. No practice management system is named, there is no Clio, MyCase or Filevine integration or case management connector, and no API documentation is published.

A firm cannot confirm that its own CRM is a supported destination or that intake will reach its matter records. A generic CRM destination describes a category rather than an integration.

Nothing located. No hosting provider is named, no region or residency commitment is published, and no deployment or tenancy options are described beyond the places the intake agent can be embedded, which is a placement question rather than an infrastructure one. The gap is wider because the same platform is marketed for healthcare patient intake and financial services qualification, so health and financial data flow through infrastructure whose location is unstated, and both carry obligations that turn on location.

No certification, attestation, trust center or security page is published. No SOC 2 of either type, ISO 27001, HIPAA statement, named auditor, penetration testing partner or encryption statement was found. The material is absent rather than gated. The HIPAA omission is the notable one. The vendor markets healthcare patient intake that captures history, urgency and care needs, which is protected health information, and publishes no HIPAA posture, no offer of a business associate agreement and no security credential of any kind.

Nothing located. No foundation model provider, model family or version is named, no distinction is drawn between proprietary and third party models, and no subprocessor list was found. The absence is conspicuous against the vendor's own published expertise: it writes detailed comparative analysis of the conversational AI category, classifying competitors by architecture and distinguishing AI native platforms from chat skins on form builders, so it demonstrably understands that the underlying architecture matters to a buyer, and it discloses nothing about its own.

A platform conducting adaptive conversations and performing schema aligned extraction is running one or more language models supplied by someone, and no party in that chain can be identified from public material.

A free entry point is published and no price is. The vendor offers a free tier with no code setup and lead capture agent templates, so a buyer can deploy and evaluate the product without a sales conversation. Nothing above it is published. No price, range, tier structure or unit of charge is published, and nothing says whether pricing runs per conversation, per completed intake, per agent, per seat or by volume, which are very different economics as inquiry volume grows.

No figures from other sources are published either. The barrier to trying is low and the cost of adopting is undisclosed.

Practice areas are named and nothing bounds them. Legal coverage names four practice areas: personal injury, family law, estate planning and immigration. Coverage extends well beyond legal into healthcare patient intake, financial advisory and wealth management, insurance first notice of loss and quoting, internal request management, education and real estate, with use case material for each. Firm coverage is not described.

There is no firm size, no jurisdiction and no statement of who the product is not for, and nothing says whether the legal configuration differs materially from the healthcare or insurance one or is the same engine pointed at different criteria. The breadth cuts both ways: a platform serving seven verticals is unlikely to carry deep legal knowledge in any one of them, and nothing published shows what the legal version knows that the generic one does not.