Alexis Austin Litle

Resources

The Legal AI Buzzword Decoder

A field guide to the terms vendors use to sell AI to law firms. What each one sounds like it means, what the marketing wants you to hear, what it actually means, what that looks like once you are using the product, and a plain test for telling a legitimate use of the term from hype.

Where no outside buyer could really check a claim, the entry says so directly, instead of inventing a test that would not hold up. The glossary grows out of the vendor evaluation framework behind the talk How to Read the Matrix.

55 terms, A to Z

#

10x productivity

What people assume it means
That 10x productivity is a measured, typical result you personally can expect.
What the marketing says
Vendors claim their tool makes a lawyer, or a task, ten times faster or more productive.
What it actually means
Using the tool makes a lawyer, or a task, ten times faster or more productive.
What it looks like in practice
This sounds precise but is almost never based on a rigorous, controlled study of representative work. It is typically extrapolated from a best-case anecdote and generalized. Broader research into AI productivity gains finds real average improvements far more modest.
Legitimate or hype, the test
Ask which specific task was measured, compared to an unaided lawyer, across how many matters, and whether it’s a median or a cherry-picked outlier. Named case studies with time logs are verifiable. A bare 10x is not.

A

“ABA rules” boilerplate

What people assume it means
That a vendor citing ABA ethics rules in its marketing has actually built a compliant product.
What the marketing says
Vendors reference ABA rules and AI ethics guidance to signal that they’ve thought through legal compliance.
What it actually means
Citing ABA rules signals the vendor is aware of legal ethics. It does not mean the vendor’s own contract actually satisfies those rules.
What it looks like in practice
A workable three-part test for any AI vendor: no training on inputs or outputs, no retention beyond the session, and no third-party disclosure. The ABA is actively pushing firms to adopt AI policies and disclosures. The regulatory pressure behind the marketing language is real, even when a given vendor’s own compliance isn’t.
Legitimate or hype, the test
Ask the vendor to point to the specific contract clause matching each rule they cite. A rule citation with no matching clause is boilerplate.

Adaptive learning

What people assume it means
That the tool personalizes deeply to your firm’s specific style and preferences over time.
What the marketing says
“Adaptive learning” is borrowed from education technology marketing to suggest ongoing, meaningful personalization.
What it actually means
A system that adjusts to ongoing use, ideally personalizing to a user or firm over time, similar in spirit to self-improving.
What it looks like in practice
Where legal vendors make this claim, it’s usually account-level personalization, remembering preferred clauses or style, a narrower thing than the phrase implies.
Legitimate or hype, the test
Ask for a concrete example of something the system has adapted for a real customer, and whether it’s stored per firm. That is checkable, and something a vendor can put in a contract.

Agentic

What people assume it means
That the AI runs itself, making decisions and taking action on its own, like a digital employee.
What the marketing says
Every booth at a major legal tech conference claimed “agentic AI” for its product.
What it actually means
True agentic AI takes action on your behalf, makes decisions, and learns simultaneously. It is a genuinely autonomous system, not just a chat window.
What it looks like in practice
Most products that claim it store and retrieve information rather than learn and act on their own. The real test is two-part. Does it take action, and does it learn inside one firm’s own data, kept isolated to that tenant, rather than a shared pool?
Legitimate or hype, the test
A vendor that can’t describe both tenant isolation and real autonomous action, concretely, is using booth language, not describing a built feature.

Air-gapped

What people assume it means
That air-gapped guarantees the system has zero connection to the outside internet, ever.
What the marketing says
“Air-gapped” is marketed as the strongest possible isolation claim.
What it actually means
The system runs on computers with no network connection to the internet at all, physically or logically disconnected.
What it looks like in practice
True air-gapping is real but expensive, usually meaning firm-owned hardware limited to smaller, less capable models than the cloud giants. Vendors marketing air-gapped are often really selling on-premise deployment, which is a different and more common thing.
Legitimate or hype, the test
Ask directly whether this system has any network connectivity, ever, including for updates. “Yes, but we push monthly updates over the internet” means it isn’t really air-gapped.

Always up to date

What people assume it means
That the tool’s underlying law database is refreshed the moment new decisions come out.
What the marketing says
“Always up to date” is used to reassure buyers nothing they see is outdated.
What it actually means
A claim the tool’s law database is refreshed constantly, so nothing you see is outdated.
What it looks like in practice
This describes a data pipeline, not the AI’s thinking. Usually a real, valuable feature when true, but there’s normally a lag between a court publishing an opinion and it being indexed.
Legitimate or hype, the test
Ask for the exact update cadence and which sources it covers, then test it yourself against a recent real decision. A vendor with no concrete schedule is using the phrase as atmosphere.

Attorney-client privilege protected

What people assume it means
That a vendor can guarantee your privilege is safe just by making that promise.
What the marketing says
Some vendors market their tools as privilege protected, as if it were a built-in feature.
What it actually means
A claim that using the tool won’t cause a court to find privilege was waived by letting a third-party vendor see the communication.
What it looks like in practice
This is the most legally loaded claim in this cluster. Whether privilege survives is a legal determination courts are actively deciding, not something a vendor can unilaterally guarantee. A vendor’s contract can support a privilege argument but can’t promise a judge will agree.
Legitimate or hype, the test
Ask what specifically backs the claim, whether confidentiality terms, no-training guarantees, or access restrictions, all of which are checkable in the contract. But genuine verification requires your own ethics counsel’s judgment on current case law. No vendor page settles this.

Autonomous agent

What people assume it means
That an autonomous agent can be trusted to handle a task from start to finish without supervision.
What the marketing says
Vendors pitch autonomous agents as a leap beyond simple chatbots, software that gets things done on its own.
What it actually means
Software that completes a multi-step task on its own, deciding, acting, and adjusting, rather than answering one question at a time.
What it looks like in practice
Law firm risk memos treat this as a real, consequential capability, not pure hype. But autonomy is a spectrum. Drafting an email for approval is very different from filing something with a court.
Legitimate or hype, the test
Ask specifically what actions the system can take without a human clicking approve, and in what systems. This is concrete and contractually specifiable.

B

Bank-grade or military-grade encryption

What people assume it means
That bank-grade or military-grade encryption is a distinct, superior technical standard.
What the marketing says
Vendors use “bank-grade” or “military-grade” encryption to sound maximally secure.
What it actually means
Not a real technical category, because there’s no such standard. Vendors almost always mean AES-256, a free, publicly published standard anyone can implement.
What it looks like in practice
It’s a sales adjective, not a spec, and nearly any vendor, good or bad, can truthfully claim it. It tells you nothing about implementation quality or who holds the encryption keys.
Legitimate or hype, the test
Ask for the actual algorithm and version, AES-256-GCM at rest and TLS 1.3 in transit, and who holds the decryption keys. A vague or defensive answer is a real, checkable red flag.

Benchmark accuracy

What people assume it means
That a cited accuracy percentage, like 95% accurate, came from a rigorous, independent test.
What the marketing says
Vendors cite benchmark accuracy percentages as proof of reliability.
What it actually means
The vendor ran its AI against a set of test questions it wrote itself, then reported the percentage it got right.
What it looks like in practice
A cited figure like 95% accurate almost never comes from an independent test. It is usually a small, self-selected question set rather than a broad sample of real legal work.
Legitimate or hype, the test
Genuinely checkable. Ask for the actual test questions, the model’s answers, and the sources cited. A vendor unwilling to share this, even under NDA, is likely hiding cherry-picked, easy questions.

C

Co-pilot

What people assume it means
That “co-pilot” products are inherently safer or more conservative than “agent” products.
What the marketing says
“Co-pilot” branding, borrowed from Microsoft, signals that the human stays in control.
What it actually means
An AI assistant meant to work alongside a human, suggesting or drafting for review, as opposed to acting fully on its own.
What it looks like in practice
Two products both called co-pilots can have very different real autonomy underneath the same branding.
Legitimate or hype, the test
Ask exactly where the human review checkpoint sits before anything leaves the system, whether sent, filed, or submitted. A vendor should show that checkpoint live, not just describe it.

Confidential by design

What people assume it means
That confidential by design names one specific, defined technical mechanism.
What the marketing says
“Confidential by design” is a marketing riff on the real GDPR concept of privacy by design.
What it actually means
Confidentiality built into the architecture from the start rather than added later.
What it looks like in practice
This is an umbrella phrase different vendors fill with different substance, whether encryption during processing, a no-training guarantee, or just “we take confidentiality seriously” with no defined mechanism at all.
Legitimate or hype, the test
Ask the vendor to name the specific mechanism behind the phrase and get it in writing. A vendor that can only restate the phrase when pressed is using decoration, not a documented principle.

Context-aware

What people assume it means
That “context-aware” means the tool deeply understands your specific matter and history.
What the marketing says
Nearly every product connected to your files calls itself context-aware.
What it actually means
A system that factors in surrounding information, your specific matter, prior documents, and firm preferences, rather than answering in a vacuum.
What it looks like in practice
This is one of the vaguer capability claims. Almost any product connected to your existing files can plausibly call itself context-aware, regardless of how deep that integration actually goes.
Legitimate or hype, the test
Ask a demo follow-up question that depends on something mentioned three questions earlier, or on an uploaded document. Directly testable, no inside access needed.

Custom model

What people assume it means
That the vendor trained its own model from the ground up, specifically for legal work.
What the marketing says
“Custom model” is used broadly, covering everything from a full retrain to a good prompt template.
What it actually means
Covers an enormous range in practice, from a genuine retrain to light prompt customization on a stock model.
What it looks like in practice
A vendor stuck with heavy custom infrastructure can actually be slower to adopt a better base model later. “Custom” can become a liability, not just a selling point.
Legitimate or hype, the test
Ask exactly what was customized, and how fast they can switch to a new base model. Not verifiable from outside: the real scale or quality of any retraining.

Customer data ownership

What people assume it means
That a vendor’s “you own your data” clause means no copies of your documents exist outside your own systems.
What the marketing says
Vendors highlight data-ownership clauses as a confidentiality safeguard.
What it actually means
“You own your data” clauses usually address intellectual property rights, meaning who has the legal right to the content, not whether copies of those documents were made, retained, or exposed to the vendor’s systems along the way.
What it looks like in practice
A sound operating rule is to never give an AI tool original documents. Make copies, isolate them in a separate file, and don’t let the tool into your core systems. There’s no exception worth the risk for entering client information.
Legitimate or hype, the test
Ask two separate questions. Who legally owns this data, and what copies of it now exist outside my firm’s systems, and where.

Cutting-edge

What people assume it means
That cutting-edge means the vendor is using techniques no one else has access to.
What the marketing says
“Cutting-edge” is one of the most common phrases in AI press releases and vendor marketing.
What it actually means
The product uses the newest, most advanced techniques currently available.
What it looks like in practice
It’s an atmosphere word, not a spec. Usually it signals the vendor licenses a well-known model from OpenAI, Anthropic, or Google, commercially available to any vendor with a budget.
Legitimate or hype, the test
Ask which specific models they use and what legal-specific adaptation was added. A concrete answer is substance. The bare phrase is decoration.

D

Data sovereignty

What people assume it means
That where a server physically sits fully settles what laws and protections apply to your data.
What the marketing says
“Data sovereignty” is marketed as a strong, geography-based privacy guarantee.
What it actually means
Data is subject to the laws of the country where it’s physically stored.
What it looks like in practice
Real, and it matters for cross-border compliance, but often overstated. At least one legal AI vendor’s own general counsel has publicly pushed back on the framing, arguing protection is about contractual commitments and practices, not just server location.
Legitimate or hype, the test
You can verify where data is stored and the vendor’s policy on government data requests. How a cross-border legal dispute would actually play out is not something a buyer can confirm alone. That’s a question for your own privacy counsel.

Dedicated instance

What people assume it means
That a dedicated instance is a meaningfully different, more secure setup than the standard product.
What the marketing says
“Dedicated instance” is marketed as an enterprise-tier privacy and performance upgrade.
What it actually means
A separate running copy of the software application itself for your firm, not just separate data storage.
What it looks like in practice
Usually an enterprise-pricing feature that overlaps heavily with single-tenant in vendor marketing. The two terms are often used almost interchangeably.
Legitimate or hype, the test
Ask the vendor to explain exactly how this differs from their standard offering. Compute, storage, a separate service commitment? An unclear answer is a real gap between label and substance.

Disruptive

What people assume it means
That disruptive means the product is actively upending the market right now.
What the marketing says
“Disruptive,” borrowed from business strategy language, implies a cheaper or simpler product that will eventually displace market leaders.
What it actually means
A cheaper or simpler product that will eventually displace market leaders, a claim about the future borrowed from business strategy language.
What it looks like in practice
Legal tech commentary flags this as overused relative to reality, since genuine disruption happens rarely in a conservative, relationship-driven field. Most disruptive AI tools make existing tasks faster rather than replacing the legal process itself.
Legitimate or hype, the test
Not checkable at the time of marketing, because genuine disruption is usually confirmed only after the fact. Ask instead what specific existing tool or workflow this replaces, and whether any reference customer has actually switched away from it.

Domain-specific model

What people assume it means
That the AI itself has specialized legal knowledge baked into its core.
What the marketing says
“Domain-specific” is pitched as proof the model itself, not just the software wrapper, understands law.
What it actually means
A real motivation exists. A Stanford study found general models had 69 to 88% error rates on specific legal recall questions, versus 1 to 3% on easier tasks like summarizing.
What it looks like in practice
But the actual fix in most “domain-specific” products is often a retrieval and citation wrapper, not a retrained model. Specialization matters. The question is how a given vendor achieved it.
Legitimate or hype, the test
Ask for a citation-accuracy demo on real legal questions, not summarization. That test is replicable yourself.

E

End-to-end encrypted

What people assume it means
That end-to-end encrypted means literally no one, including the vendor, can ever read your document.
What the marketing says
“End-to-end encrypted” is one of the highest-trust security phrases a vendor can use.
What it actually means
Strictly, this means data scrambled on your device that only the recipient’s device can unscramble. Not even the provider can read it in the middle.
What it looks like in practice
This is the most misapplied term in this cluster. An AI has to see your unencrypted document to process it, so this almost always really means encrypted in transit and at rest, decrypted during AI processing, which is a different and weaker claim. A rare, genuine version, using confidential computing or secure enclaves, does exist.
Legitimate or hype, the test
Ask exactly where in the pipeline the document is decrypted and who can see it then. “Our servers decrypt it to feed the AI” is normal security, not true end-to-end encryption.

Enterprise-grade, and free tiers

What people assume it means
That paying for the enterprise tier guarantees stronger technical protection for your data.
What the marketing says
Vendors position “enterprise-grade” as a technical guarantee of confidentiality and security.
What it actually means
“Enterprise-grade” is often a pricing-tier name, not a technical spec. What actually changes tier to tier, meaning training defaults, retention windows, and support commitments, is the real question.
What it looks like in practice
On free tiers, the blunter rule holds. If it’s free, you’re the product, and by extension, so are your clients.
Legitimate or hype, the test
Ask what specific contractual terms differ, tier by tier, in writing. “Enterprise” with no listed differences is a label, not a feature.

F

Fine-tuned

What people assume it means
That fine-tuning is a straightforward improvement that reliably makes a model better at legal work.
What the marketing says
Vendors cite fine-tuning as evidence of specialized training and improved accuracy.
What it actually means
Fine-tuning means taking an already-built model and giving it extra training on a narrower set of examples, without building a new model from scratch.
What it looks like in practice
Done carelessly, fine-tuning can actually make a model worse at general reasoning. One legal tech vendor’s own blog warns it’s “often completely unnecessary or impractical.”
Legitimate or hype, the test
Ask what data was used, how much, and whether independent testing that the vendor did not run, or a live trial on your own documents, backs the claimed improvement.

G

GDPR compliant

What people assume it means
That GDPR compliant is a verified status, similar to a certification.
What the marketing says
Vendors state “GDPR compliant” on their homepage as a blanket assurance.
What it actually means
The vendor claims to follow the EU’s data privacy rules, covering consent, deletion rights, 72-hour breach reporting, and purpose limitation.
What it looks like in practice
No government body issues a GDPR compliant certificate, unlike ISO 27001. A bare homepage claim is a self-assessment, and often really means “we have a DPA available” or “we have EU data centers.”
Legitimate or hype, the test
Ask for the actual Data Processing Addendum and check for named safeguards, such as Standard Contractual Clauses, named sub-processors, and breach timelines. Whether internal practices fully meet every obligation isn’t verifiable by an outside buyer without an audit, and that limit is real.

“Generic AI” as a differentiator

What people assume it means
That “not generic AI” means the vendor built something specifically for legal work at a technical level.
What the marketing says
Vendors distance themselves from “generic AI” language to sound specialized.
What it actually means
“Generic AI” is usually an admission dressed as a pitch. The vendor is reselling access to a general model with little legal-specific work layered on top.
What it looks like in practice
Generic AI marketing tends to get called out publicly once buyers dig in. The term now works against a vendor about as often as for one.
Legitimate or hype, the test
Ask what, specifically, is legal-specific about the product beyond the underlying model. No concrete answer means it’s generic AI, regardless of the label on the box.

H

Hallucination-free, or near-zero hallucination

What people assume it means
That the product essentially never invents facts or fake citations.
What the marketing says
Some vendors have marketed their products as near-zero or 100% hallucination-free.
What it actually means
Zero is not achievable with current AI. “Hallucination” is itself a slightly misleading word. It anthropomorphizes a technology that is working exactly as designed when it produces the most probable answer instead of the correct one.
What it looks like in practice
Narrowing what a model is allowed to read, down to a smaller vetted document set, is real engineering that measurably cuts error rates. That is a reduction, not an elimination. Independent academic testing has repeatedly found real, non-zero error rates in exactly the products making this claim, and several vendors have quietly softened their language since.
Legitimate or hype, the test
Any claim of zero hallucinations is a red flag on its own. Ask for the measured error rate on a real test set and how the vendor restricts what the model draws on. Then check whether an independent, published study, not the vendor’s own test, has measured this specific product.

HIPAA-ready versus HIPAA compliant

What people assume it means
That HIPAA-ready and HIPAA compliant mean roughly the same thing.
What the marketing says
Vendors advertise HIPAA-ready, letting buyers assume the product is compliant out of the box.
What it actually means
“Ready” tends to mean the capability exists somewhere in the product line, usually gated behind the most expensive tier and a signed Business Associate Agreement, not that what you’re buying today is compliant.
What it looks like in practice
One real vendor evaluation found SOC 2 Type 2 and enforced HIPAA compliance only at the enterprise tier, despite HIPAA-ready language on the general marketing page.
Legitimate or hype, the test
The compliance gate is the Business Associate Agreement, not the tool. Ask for the BAA by name and which tier it applies to. No BAA offered means not compliant, regardless of what the page says.

I

Instant

What people assume it means
That a fast answer is also necessarily a correct and complete one.
What the marketing says
“Instant” is used to emphasize speed as the headline benefit.
What it actually means
A claim of essentially no perceptible wait for an answer.
What it looks like in practice
Often literally true for the response itself, but speed says nothing about accuracy or completeness. Not visible from a website: whether speed came from skipping slower verification steps.
Legitimate or hype, the test
The speed itself is trivially checkable by using the product. Whether verification was skipped to stay fast generally isn’t checkable from outside.

ISO 27001

What people assume it means
That any mention of ISO 27001 means the whole company and product are certified.
What the marketing says
Vendors cite ISO 27001 as proof of internationally recognized security.
What it actually means
An internationally recognized certification that an independent auditor reviewed a company’s security processes against a defined checklist, at a point in time.
What it looks like in practice
One of the more legitimate, verifiable claims in this cluster when it is real. The catch is scope. A certificate can cover only part of a company, the office rather than the product infrastructure, and fabricated or lapsed certifications have been documented in legal tech.
Legitimate or hype, the test
Fully checkable. Ask for the certificate number and certifying body, and verify against a public registry such as IAF CertSearch. Separately confirm what scope the certificate actually covers.

“It learns,” and tenant isolation

What people assume it means
That the AI is quietly getting smarter about your firm specifically, the more you use it.
What the marketing says
Vendors say the tool learns from your data, implying private, ongoing personalization.
What it actually means
Learning often means writing to a shared database that other customers’ data also flows through, not a model adapting privately to one firm.
What it looks like in practice
If it’s pulling from many firms’ data and learning from all of it, that raises a real security question the vendor needs to answer concretely, not wave away.
Legitimate or hype, the test
Ask how it’s secure if it’s learning from everyone’s data. A vendor who can describe tenant isolation in specific terms is describing something real. One who just says it’s secure is not.

M

Multi-agent

What people assume it means
That “multi-agent” means several genuinely different, separately built AI systems are collaborating.
What the marketing says
“Multi-agent architecture” is used to sound sophisticated and modular.
What it actually means
Several specialized AI agents that each handle part of a task and hand off to one another, rather than one model doing everything.
What it looks like in practice
This is a real, describable engineering pattern generally, but the individual agents are very often the same general-purpose model called multiple times with different instructions, not separately built models.
Legitimate or hype, the test
Ask the vendor to walk through what each agent does and where a human can inspect or intervene at each handoff.

Multimodal

What people assume it means
That the AI can genuinely handle scans, handwriting, and audio, not just typed text.
What the marketing says
“Multimodal” is marketed as a distinguishing, advanced capability.
What it actually means
An AI system that works with more than typed text, including images, scanned documents, handwriting, and audio.
What it looks like in practice
This is real and meaningful where true, since a text-only tool genuinely can’t read a scanned, handwritten amendment. But most current major models are multimodal by default, so the claim is easy to make without legal-specific work behind it.
Legitimate or hype, the test
Hand the vendor a messy real-world document in a demo, a scanned fax or a handwritten margin note, and see what it actually does with it.

N

Next-generation

What people assume it means
That next-generation describes a genuine leap forward specific to that product.
What the marketing says
“Next-generation” implies the product represents a meaningfully newer era of technology than what came before.
What it actually means
This product represents a meaningfully newer generation of technology than what came before.
What it looks like in practice
It carries almost no specific technical content alone. In legal AI it typically distinguishes generative AI from older keyword-search software, a real distinction, but the phrase alone says nothing about this particular tool.
Legitimate or hype, the test
Not independently checkable as a standalone claim. Ask for a specific, named comparison instead, meaning what this does that the old tool couldn’t, and why.

O

On-premise

What people assume it means
That on-premise always means the software runs on hardware physically inside your own office.
What the marketing says
“On-premise” is marketed as the maximum-privacy deployment option.
What it actually means
The AI runs on computers physically owned by your firm, not the vendor’s servers on the internet.
What it looks like in practice
A real, meaningful architectural choice, usually a higher up-front cost in exchange for physical custody of the data. Often a smaller, less capable model than the cloud giants, and “available” sometimes really means an enterprise tier most customers don’t actually buy.
Legitimate or hype, the test
Ask whether it’s genuinely on the firm’s own hardware or a rebranded single-tenant cloud instance, which is a different and more common thing, and whether vendor engineers ever remotely access it.

P

Patent-pending

What people assume it means
That the technology has already been reviewed and approved as genuinely novel.
What the marketing says
“Patent-pending” is used to imply validated, protected, cutting-edge technology.
What it actually means
A patent application has been filed but not yet approved. It says nothing about the significance of the technology, whether it will ultimately be granted, or how the technology actually works.
What it looks like in practice
Rare in mainstream generative AI legal marketing specifically, and more common in older legal tech subsegments like document automation or IP tools.
Legitimate or hype, the test
Fully checkable, because patent applications are public record. Look up the actual filing. A vendor unwilling to give the application number is a clear yellow flag.

Plug-and-play

What people assume it means
That plug-and-play means the product will work perfectly with your case management system the moment you connect it.
What the marketing says
“Plug-and-play,” borrowed from hardware, promises zero integration work with your existing systems.
What it actually means
Zero integration work with your existing systems, borrowed from hardware language, where you plug in a device and it works.
What it looks like in practice
Some genuinely lightweight, self-contained tools, like a browser extension or a standalone chat tool, really can be near-instant. The phrase is most inflated for anything connecting into case or document management systems, where real value comes from the integration and data preparation work vendors claim to skip.
Legitimate or hype, the test
Ask specifically what plug-and-play means for your tech stack. Which systems, what setup, how many IT hours.

Proprietary AI

What people assume it means
That the vendor built its own unique AI model from scratch, something no one else has.
What the marketing says
Marketing uses “proprietary AI” to imply a custom-built brain unlike anything else on the market.
What it actually means
Most of the time it means the vendor built its own prompts and guardrails on top of an existing model like ChatGPT, Gemini, or Claude. It is not a separately built model.
What it looks like in practice
A wrapper with well-engineered prompts can still be useful, but it isn’t “proprietary AI” in the sense buyers hear it. One marketing firm reportedly charged $7,000 for what amounted to figuring out the prompts.
Legitimate or hype, the test
Ask directly whether anything besides prompts and guardrails was built. If the answer only describes prompt engineering, the term is doing more work than the product is.

Proprietary algorithm

What people assume it means
That there’s a technical whitepaper or explanation somewhere backing up the claim.
What the marketing says
“Proprietary algorithm” is presented as evidence of unique, superior technology.
What it actually means
A method the company owns and keeps confidential as a trade secret, not a published, publicly known process.
What it looks like in practice
Because it’s kept secret by design, this is one of the least checkable claims in vendor marketing. There’s no whitepaper coming, and there isn’t meant to be one.
Legitimate or hype, the test
Genuinely not verifiable from the outside. “Proprietary” is a promise of non-disclosure, not a lie you can catch. Test outputs on your own documents instead of trying to verify what’s inside.

Purpose-built

What people assume it means
That the whole product, model and all, was designed from the ground up for legal work.
What the marketing says
“Purpose-built for legal” is a common headline claim across legal AI vendors.
What it actually means
It usually means retrieval, citation checking, and workflow guardrails were added around a general model. The underlying brain is very often still the same GPT, Claude, or Gemini everyone else uses.
What it looks like in practice
The purpose-built layer can still be genuinely valuable, and a good citation-checking layer is real engineering, even when the core model isn’t unique to that vendor.
Legitimate or hype, the test
Ask what specifically is purpose-built, the model or the retrieval and verification layer around it. A vendor pointing to a named legal database and a citation check is showing something real.

R

Real-time

What people assume it means
That “real-time” means the tool is plugged into a live, continuously streaming data feed.
What the marketing says
“Real-time” is used loosely in marketing to mean fast, or as things happen.
What it actually means
Technically, real-time means a guaranteed, near-instant response window. In marketing, it’s used loosely to mean fast, or as things happen.
What it looks like in practice
Usually it just means the tool answers quickly, not a claim about a live, continuously streaming feed. That is the separate “always up to date” claim, often bundled in for effect.
Legitimate or hype, the test
Ask what it is real-time relative to. A vendor pointing to a specific, named data feed is describing something real. One meaning only fast chat is using it as an adjective.

Reasoning model

What people assume it means
That “reasoning” means the AI thinks through a problem the way a person would.
What the marketing says
“Reasoning model” is used to imply deeper, more careful analysis than a standard chatbot.
What it actually means
A model that works through a problem in explicit intermediate steps before answering, rather than one-shot. Slower, and often more accurate on complex questions.
What it looks like in practice
This is a real, technically grounded distinction, and reasoning-style models genuinely exist and behave differently, but vendors often use “reasoning” loosely for any multi-step or agentic workflow.
Legitimate or hype, the test
Ask whether the vendor is using an actual reasoning-model API or just chaining regular calls together. Test it yourself on a complex question and a simple one.

Revolutionary

What people assume it means
That revolutionary describes a proven, already-happened transformation.
What the marketing says
“Revolutionary” is used to imply the product will fundamentally transform, not just improve, how legal work is done.
What it actually means
The product will fundamentally transform, not just improve, how legal work is done.
What it looks like in practice
Legal tech commentary notes true industry-wide revolutions are rare in a conservative, relationship-driven field, and change tends to be incremental. This is almost always aspirational, not a description of documented change.
Legitimate or hype, the test
There’s no outside way to verify this in the present tense, because whether something was revolutionary is usually only clear in hindsight. Ask for the specific, measurable change claimed instead.

S

Seamless integration

What people assume it means
That seamless means zero setup, zero friction, and full coverage of your workflow.
What the marketing says
“Seamless integration” is a near-universal phrase in software marketing generally, legal AI included.
What it actually means
The tool connects to your other software without friction or manual work.
What it looks like in practice
Real, vendor-built integrations exist and are valuable, but seamless is often applied even when the integration is really an export and import step, or covers only part of your workflow.
Legitimate or hype, the test
Ask for the specific, named list of systems it integrates with, native or through a third-party connector, and see it working live against your actual systems in a demo.

“Secure”

What people assume it means
That “secure” covers everyone who might see your data, including the vendor’s own staff.
What the marketing says
Vendors describe their platforms simply as secure, without specifying secure from what.
What it actually means
SaaS security usually means security from outside hackers, not from the vendor’s own staff or subprocessors reading your files as part of normal operations.
What it looks like in practice
A product can be genuinely well defended against attackers and still have a wide circle of people internally who can access client files routinely.
Legitimate or hype, the test
Ask secure from whom. A vendor that names a specific threat model, whether external attackers, insider access, or subpoena, is describing something real. One that just repeats “secure” is not.

Self-improving

What people assume it means
That the tool keeps getting smarter on its own, without anyone doing anything.
What the marketing says
“Self-improving” is a leftover phrase from an earlier era of machine learning marketing, still occasionally used to sound futuristic.
What it actually means
A system that gets better over time without a human explicitly retraining it.
What it looks like in practice
Rare in current legal AI marketing specifically. Where the idea appears today it’s usually the softer “learns from attorney edits,” where corrections get logged and used to adjust future suggestions, a narrower and more concrete thing.
Legitimate or hype, the test
Ask what concretely improves, and over what time frame. A vendor that can’t name a specific feedback mechanism is using aspirational marketing.

Single-tenant

What people assume it means
That single-tenant means your firm’s data is on physically separate hardware from everyone else’s.
What the marketing says
“Single-tenant” is marketed as maximum data separation from other customers.
What it actually means
Your firm gets its own separate server or database, instead of sharing infrastructure with other customers that’s logically separated by software rules.
What it looks like in practice
This is a real spectrum, not a binary, running from shared infrastructure with software-level separation, to isolated virtual environments, to fully separate physical hardware. Usually a premium tier reserved for the largest clients.
Legitimate or hype, the test
Ask exactly what is and isn’t shared, meaning compute, storage, the database, and encryption keys, and get it in the contract, not just the sales deck.

SLM versus LLM

What people assume it means
That “SLM,” meaning small language model, is simply the smaller, more advanced, or more premium option, maybe even running on your own servers.
What the marketing says
Vendors pitch hybrid setups, LLMs as generalists and SLMs as specialists, and use “small” to imply a leaner, more private, more exclusive setup, often alongside a higher price.
What it actually means
An LLM is a very large, general-purpose model trained on huge amounts of internet text. An SLM is a smaller, more narrowly trained model, cheaper and faster, sometimes more accurate on a narrow task, worse at open-ended reasoning. “Small” refers to the training data, not where the model lives or runs.
What it looks like in practice
One legal vendor pitches LLMs as “versatile legal interns” and SLMs as “expert paralegals.” Often this signals a cost and control play, cheaper to run and easier to keep data contained, more than a proven accuracy edge. Paying more for a vendor’s product doesn’t by itself mean you’re getting a smaller, more contained model.
Legitimate or hype, the test
Ask directly what “small” refers to in their pitch, the training data, the deployment, or just the price tier, and ask for accuracy comparisons on your own sample documents. This is an architecture choice a vendor can show without difficulty.

SOC 2 Type 2

What people assume it means
That SOC 2 Type 2 certification means the vendor is genuinely secure and trustworthy with client data.
What the marketing says
It’s treated in sales conversations as table-stakes proof of security, almost a magic word.
What it actually means
SOC 2 Type 2 audits a vendor’s internal security process over time. It says nothing about whether your specific documents are ever seen by a human at that company.
What it looks like in practice
It’s real and checkable, and you can ask for the report, but it’s a baseline, not a confidentiality guarantee.
Legitimate or hype, the test
It’s legitimate as far as it goes. The hype is in treating it as the whole answer. Ask separately who can access your data, and when.

T

Turnkey

What people assume it means
That turnkey means you can start using the product immediately with no setup at all.
What the marketing says
“Turnkey” promises the product works immediately with no setup or custom work needed.
What it actually means
The product works immediately with no setup or custom work needed.
What it looks like in practice
One of the more misleading superlatives. Deploying AI into a real firm almost always requires connecting to document systems, configuring permissions, and training staff. It often really means “we hosted it with a simple interface,” not “no implementation work.”
Legitimate or hype, the test
Ask for a written implementation timeline and a concrete list of what your staff or IT must do. A short, specific list with a realistic go-live date is real. Vague repetition of “turnkey” is not.

V

Vertical AI

What people assume it means
That “vertical AI” describes something specific and technical about how the product works.
What the marketing says
“Vertical AI” is used to signal deep industry focus, as opposed to general-purpose AI.
What it actually means
It’s more an investor and positioning term than a technical claim. It describes a competitive bet, not what’s actually inside the product.
What it looks like in practice
Two “vertical AI” companies in the same space can have wildly different amounts of real, legal-specific engineering underneath the same positioning language.
Legitimate or hype, the test
This term makes no specific technical claim to verify. Ask for concrete legal-specific features instead of accepting the label as proof of anything.

W

“We don’t train on your data”

What people assume it means
That turning off training means your data is safe and isn’t being kept anywhere.
What the marketing says
Vendors offer “we don’t train on your data” as reassurance that client confidentiality is protected.
What it actually means
Not training is a promise about future model versions. It says nothing about whether your document is still sitting on the vendor’s servers, unencrypted to their own staff, right now.
What it looks like in practice
Retention windows on tools observed in the wild have run anywhere from two weeks to two years, regardless of the training toggle. Training and storage are two separate questions.
Legitimate or hype, the test
Ask training and retention as two separate questions. A vendor collapsing them into one “we don’t train” answer is dodging the retention question.

“We integrate with everything”

What people assume it means
That broad integration is purely a convenience feature, with no downside.
What the marketing says
“We integrate with everything” is pitched as a selling point, maximum flexibility and minimum friction.
What it actually means
Every system a tool connects to is a door into your data. Broad integration claims are a risk signal, not just a convenience one. More doors, not fewer.
What it looks like in practice
Each additional integration is a place client data can flow to a third party you never vetted. A tool can genuinely integrate with dozens of systems and still only be safely usable through a few of them in your actual workflow.
Legitimate or hype, the test
Ask for the specific integrations used in your actual workflow, and who has access through each. “Everything” isn’t an answerable security question.

Z

Zero data retention

What people assume it means
That zero data retention guarantees nothing about your input is ever stored, full stop.
What the marketing says
Vendors market zero data retention as a strong privacy promise, sometimes as a plain feature, sometimes as a premium add-on.
What it actually means
The vendor, or its underlying model provider, processes a request, generates a response, then deletes the input, so it is not stored afterward and not used to train. When it’s a specific, named contractual term, distinct from the vaguer “we don’t train” line, it’s a genuine differentiator worth shopping for.
What it looks like in practice
Some legal AI vendors have negotiated true zero-retention terms with their underlying model providers. But watch for the tiering trick: nothing, some, everything, but 10x the price. Real zero-retention terms are often reserved for the most expensive tier, and the phrase is also sometimes used loosely to mean only “we don’t train,” while the vendor’s own platform still logs prompts for debugging or analytics.
Legitimate or hype, the test
Ask which tier includes a written zero-retention clause, and get the clause itself, not just the phrase. Also ask separately whether the vendor’s own platform, not just the model API, logs anything. A vendor with no contract language is likely offering marketing gloss.

This is not legal advice. For anything touching privilege, confidentiality obligations, or a specific vendor contract, run it by your own counsel.