Antidote vs Legal Decoder: how they compare in 2026

Remove the models and there is no product. Antidote ships no document system, no workflow engine and no time recording of its own; it is explicitly an overlay on the firm's existing time recording and practice management software. Every function it performs is a model task: extracting billing rules from a client's outside counsel guideline documents and the firm's historical billing data to build Individual Client Standards, analyzing time entries continuously through the billing cycle against those standards, and generating a corrected narrative for a fee earner to accept. The vendor's own naming makes the same point, describing the analysis layer as Antidote AI. There is no rules engine or template library underneath that would survive the models being taken away, and no conventional software product the AI is a feature on.

The models turn billing prose into data, and a rules engine does the judging. Legal Decoder's own framing is that UTBMS codes tell you what was billed but not what happened, so its technology reads the narrative itself: entries are parsed, categorized, enriched, standardized and then evaluated against more than 45 proprietary compliance flags. Aperture adds a natural language interface on top, and the vendor is explicit that the AI agents decide what data and calculations a question needs while a separate deterministic engine does the arithmetic. Underneath the language work sits an analytics platform of flags, benchmarks and indices that would still run on structured invoice feeds without any model reading prose.

Grounding is described at the level of its inputs, and accuracy is asserted without measurement. The source material is named concretely: outside counsel guideline documents, internal firm guidelines and historical billing data, combined into Individual Client Standards for each client against which entries are checked. That is a real account of what the output is grounded in. Everything testable is missing. No accuracy figure, test set, evaluation, error rate or description of retrieval method appears, and nothing addresses what happens when the model misreads a guideline or rewrites a narrative into something the fee earner did not do. Accuracy is instead asserted, with corrected recommendations said to ensure compliance. There is no published statement about hallucination in either direction.

The method is described in detail and the performance is not measured. Aperture's published answers show an Analysis Request Breakdown setting out the data used for the response, and every figure is computed by a deterministic engine rather than the model. The underlying flags are documented down to the individual rule. Worked examples in the published bankruptcy retrospective show which entries triggered which flag and why. What is asserted rather than tested is quality. The site describes the methodology as validated by courts and says Legal Decoder is the only invoice review technology accepted at the federal court level, without citing a decision or naming a court. No accuracy, false positive or agreement rate for the flags is published.

A human approval step is part of the published design, and the threshold structure around it is absent. The workflow is stated plainly: Antidote sends suggested fixes directly to email so fee earners can review, approve and move on, and corrections are described as recommendations accepted with one click. Nothing is presented as writing to the billing system unattended, which matters because the output is text that reaches a client invoice. What is not published is the control structure around that review. Nothing says what happens to entries nobody approves, sets a threshold at which the system acts alone, surfaces a confidence signal to the reviewer, or describes what a supervising partner sees as against a fee earner. No escalation path is described.

The product flags and a person decides, and the review surface is the flag itself. Every flag attaches to an individual line item with the rule that fired, and users drill from an index category to a timekeeper to the entry, so each conclusion can be checked against the billing record behind it. Aperture is described as sticking to observations and suggested next steps by default and producing a specific action plan only when asked. Users hold read only permissions, so nothing in the analysis can alter the underlying billing data. What is not published is any threshold or confidence level at which a flag fires, or guidance on what a flagged entry may not be used for when it becomes the basis of a fee objection or a write down.

No production evidence is published on the vendor's own pages. There are no named customers, case studies, testimonials or customer logos; the only logo strip is headed Backed by and shows investors rather than users. The home page figures are modeled rather than observed, and the vendor says so in a footnote. The 8 to 10 percent write off rate, the 30 extra days to payment, the 17 million dollars of working capital tied up and the 800,000 dollars of partner time are based on a firm of 100 partners billing 200 million dollars annually. That basis is disclosed, but a hypothetical firm is not a deployment. Trade coverage of the January 2026 funding round refers to adoption by firms in the United States, United Kingdom and Australia without naming any of them, and funding coverage is not deployment evidence.

The numbers are real and the customers are not named. The published Allied Nevada Gold retrospective analyzes a named Chapter 11 case: over fifty invoices and 2,000 pages covering 10.7 million dollars in fees, 14,755 line items and 135 timekeepers across eight named law firms, with over 21,000 flags and roughly 2.4 million dollars identified as unrealized savings, and findings for each firm set out with percentages. That is a demonstration on public court filings rather than a customer deployment; the firms analyzed are the subjects, not the buyers. The other success stories, including an AmLaw 100 firm's risk management program and the Toys "R" Us fee review, describe the customer only by segment, and the homepage quotes are attributed to roles rather than people.

Confidentiality is asserted in general terms, and every specific commitment sits behind an access request or in an unpublished contract. Antidote's trust center states that the commitment to data privacy and security is embedded in every part of the business. It lists a Data Protection Policy, an Information Security Policy and an Encryption Policy, all behind a Get access flow. The published privacy notice expressly excludes the material that matters. It records that users may upload confidential company information, which may include personal data, and that its collection and protection are subject to a separate agreement and not covered by that policy. Nothing published addresses training on customer content, separation by matter or client, tenancy, or what the named model provider may retain. The content ingested is time entry narratives, which describe what a lawyer did on a matter and are routinely redacted in fee disputes for that reason.

The confidentiality commitments are specific to the one thing this product handles, and they sit outside any agreement a buyer can read. Invoice narratives describe matters, clients and strategy, and Aperture's published answers address that directly. Identifying fields such as matter, client and attorney names are tokenized before any data reaches the model, so the model sees strings like Matter_ABCDEF. Tokens cannot be reverse engineered, new identifiers are generated each session rather than held in a persistent vault, and results are remapped inside Legal Decoder's own environment. The vendor also states that data is never used to train models and that users hold read only permissions. Nothing published addresses separation between customers, or privilege or work product in the narratives. Both the Terms of Use and the Privacy Policy state that client processing is governed by separate master services and data processing agreements that are not published.

The advice line and jurisdiction limits matter little here, and professional responsibility, which does apply, is unaddressed. The product gives no legal advice and has no surface for consumers. The terms restrict access to Law Firm Users authorized by a law firm, require use in a commercial capacity, and state that the services are not intended for personal use. Professional responsibility applies directly. Antidote generates the narrative that describes a lawyer's work and then travels to the client on an invoice, which engages the rules on fee reasonableness and candor to a client. Nothing published addresses it: no rule of professional conduct, bar or regulator, no ethics guidance, no statement that a fee earner remains responsible for the accuracy of a narrative the model rewrote, and nothing on whether an entry corrected by AI must still reflect the work actually done. The nearest provision runs the other way, with the terms disclaiming any liability for a law firm's delivery of legal services to its clients.

The line between analysis and judgment is drawn in the vendor's own words. The published methodology states that the technology neither provides legal advice nor operates as a referendum on the quality of legal services provided or the outcome delivered by outside counsel. That is the relevant disclaimer for a product whose output is used to challenge another lawyer's fees. The Terms of Use dated 21 January 2026 add that website content is not legal advice and creates no attorney client relationship. What is missing is the other half. Nothing addresses the duty of the person acting on a flag, whether a fee examiner, a general counsel or a billing partner, or what review a flagged entry should get before it becomes an objection or a write off.

No governance position is published for a system that rewrites the record of billable work. There is no responsible AI page, no principles statement, no named owner accountable for model behavior, no evaluation before release, no testing regime and nothing on bias. The navigation and footer list How it works, Resources, About, Contact, Terms and Conditions, Privacy Notice, Cookie Policy and the trust center, and none of them carries governance material. The trust center's documents all concern security: an Acceptable Use Policy, Backup Policy, Data Protection Policy, Encryption Policy and Information Security Policy. Security controls are a different subject from AI governance. The trust center does carry one candid line about an adjacent gap, recording that an application security monitoring program is being put together.

How the system works is explained; how it is governed is not. The Aperture material is open about mechanism. It says the model plans the analysis and never calculates, that tokenization happens before processing, and that the underlying data is parsed, categorized, enriched, standardized and evaluated in sequence. Nobody inside Legal Decoder is named as accountable for the flags or the models, nothing describes what is tested before a rule or release ships, and no evaluation results are published. Nothing addresses whether flags fall unevenly across practice areas, firm sizes or timekeeper seniority, which matters for a product whose output is used to question individual lawyers' bills.

Real material exists around the edges, and the published policy does not reach the content the product ingests. Published: a retention table giving 180 days from the end of the relationship for identity and account records, 12 months for technical and usage data, and 7 years for inquiry records. International transfers rest on UK adequacy regulations or the International Data Transfer Agreement, and an ICO registration number, ZB821053, can be checked on a public register. A named subprocessor list covers Amazon Web Services, Anthropic, Google, Linear and WorkOS, and storage is offered in the EU, US or Australia. The rules for customer content are absent, and the privacy notice says so expressly: uploaded confidential company information is subject to a separate agreement and is not covered by that policy. So no retention period, deletion route or incident notification commitment applies to time entries, narratives or ingested guideline documents. Encryption and Backup Policies sit on the trust center behind an access request.

One question is answered precisely and the rest are not. The Aperture material says the product is not a system of record and does not store, own or act as the repository for billing information, which should remain in the customer's electronic billing or matter management system. It also says users have read only access. Beyond that, no retention period for what Legal Decoder does hold, no deletion commitment, no encryption statement, no subprocessor list and no incident practice is published. There is no security or trust page anywhere on the site. The Privacy Policy dated 21 January 2026 covers website visitors and marketing only, and states that client processing sits under separate agreements that are not published.

Nothing published allocates loss for the platform, and the published document says so itself. The only language allocating loss is in the website terms of use. They exclude indirect and consequential loss and loss of profit, revenue, contracts, data and goodwill, and preserve liability for death, personal injury and fraudulent misrepresentation. They then state that any liability for losses arising from use of the services is limited as set out in the separate agreement between Antidote and the user's employer. That agreement is not published, and website terms do not govern the platform, so the published terms point to a document no one can read before signing. No indemnity of any kind, no warranty on output and no cap figure appears anywhere. Antidote's trust center lists a Cyber Insurance document under Legal, behind the Get access flow, so an insurance position exists, with its scope and limits not published.

Nothing published says who bears the loss if a flag is wrong. The Terms of Use dated 21 January 2026 govern the website: they disclaim warranties, exclude indirect and consequential damages and require the user to indemnify Legal Decoder, and they address the website rather than the analysis a customer buys. The Privacy Policy points client processing to master services and data processing agreements which are not published, and no warranty, indemnity, cap or insurance position covering the product is published. For a product whose output is used to reduce another party's fees, nothing addresses the consequences of a mistaken flag.

Integration is the product's whole delivery model, and not one system is named. The claim is central and repeated: the product works inside existing workflows, promises no workflow disruption, and sits on top of a firm's existing time recording and practice management software. That is true of the architecture, since Antidote reads and corrects entries held elsewhere. But there is no integrations page. No practice management, time recording, ebilling or document management system is named on the vendor's own pages, and there is no API reference, connector list or statement of what syncs in which direction or what a firm must configure. A real capability is claimed with nothing an implementer could use.

No integration into the systems legal billing already runs through is published. Aperture is described as a complementary analytics layer rather than a replacement for an electronic billing or enterprise legal management system, and the customer's system is expected to remain the system of record, but nothing names a single one of those systems, describes how invoice data reaches Legal Decoder, or documents an API, file format or connector. There is no integrations page and no developer documentation; the Partners page describes commercial partnerships with alternative legal service providers and consultancies rather than technical connections.

Residency is offered as a choice, and processing is not addressed separately from storage. Antidote's home page states that data is stored in region and names three options: the European Union, the United States or Australia. Amazon Web Services is identified on the trust center as the hosting subprocessor, so the infrastructure provider is named. Tenancy is never addressed: nothing states whether the platform is single or multi tenant, and no dedicated or isolated option is offered. Where processing happens is not stated as distinct from where data is stored. That matters because Anthropic and Google are named subprocessors and no location is given for model processing, so choosing EU storage does not establish that the model call stays in the region.

Nothing published says where the software runs or where invoice data is held. The product is reached through a web portal, and no tenancy model, hosting provider, region or residency option appears anywhere; the Privacy Policy addresses website data and says nothing about where client processing happens. Buyers in this market routinely send an entire outside counsel spend file, which makes the absence a live question for a procurement review.

Named certifications and a real trust center, without evidence that can be read openly. The SafeBase portal at trust.antidotelegal.com renders in full and carries three compliance entries: GDPR, ISO/IEC 27001:2022 with the standard version named, and SOC 2 Type 1. Its document set lists an Antidote Legal Security Summary 2026, a Pentest Report, the certification documents, Cyber Insurance and five policies. The certifications are stated rather than implied, and a route to the underlying reports visibly exists. No report can be read openly. Every document sits behind a Get access flow that offers to start a security review, view and download sensitive information or ask for information, without saying whether fulfillment is self serve or runs through a sales conversation. No auditor is named, no audit period or report date beyond the year appears, and no scope statement is published. The attestation is SOC 2 Type 1, an assessment of control design at a point in time, not Type 2.

No independent security attestation is published. There is no security page and no trust center on the estate, and no SOC 2, ISO 27001, penetration test or named auditor is mentioned on any page; the only security language is the Privacy Policy's statement that reasonable administrative, technical and physical safeguards are maintained and that no system is completely secure.

The provider is named and the rest of the chain is not. The trust center publishes a subprocessor list without an access request, naming Amazon Web Services, Anthropic, Google, Linear and WorkOS. Anthropic's listing identifies whose models process customer content. Google appears alongside it, though the list does not say whether that entry is infrastructure or a second model provider. No model is named, only the provider, so which model or version reads a time entry is not stated. Where model processing runs is not stated; naming a cloud host says where the platform sits, not whose model reads the content. No commitment to notify customers when a subprocessor, provider or model changes is published.

The model is referred to and never identified. The Aperture material describes data being sent to an LLM after tokenization and says potentially sensitive fields are protected before any data reaches the AI model, which tells a buyer that a third party model is in use without saying whose. No model, version or provider is named, nothing states where inference runs, and no commitment to notify customers when the model changes is published. The vendor's own distinction, that the model plans and explains while a deterministic engine computes, limits what the unnamed model actually decides.

No pricing information is published at any level, including the unit of charge. The navigation and footer include no pricing page. Every call to action on every page is Book a demo, routing to a contact form. Nothing states whether the product is charged per fee earner, per timekeeper, per entry checked, per matter or as a platform fee, and no tier, minimum, term or implementation cost appears. The terms of use refer to placing an order and to a separate agreement with the customer's employer without describing any commercial term. The product's own pitch is quantified in pounds and dollars of recovered revenue, so the site gives a detailed model of the benefit and nothing about the cost.

No price, rate or unit of charge is published: there is no pricing page, tier, rate per invoice or per dollar reviewed, or trial, and every route ends at a scheduled call. Pricing Decoder is a product for pricing legal work, not a statement of what Legal Decoder costs.

The buyer is claimed broadly and the boundary is drawn only in the contract. Marketing addresses leading law firms and leading global firms without defining either. The only quantitative indicator of segment is the modeled example of a firm of 100 partners billing 200 million dollars a year, which is an illustration rather than a statement of who the product is for. The real boundary sits in the terms, which restrict access to Law Firm Users authorized by a law firm and exclude personal use. In house and government legal teams are therefore outside the product by contract rather than by any published statement of scope. Practice area matters little here, since billing compliance works on time entries whatever the matter type. Size and geography are not stated plainly. No firm size range, minimum timekeeper count or jurisdictional coverage appears, and the only geographic signal is the choice of storage regions.

Three buyers are addressed with separate pages and different arguments: corporate legal teams justifying outside counsel spend, law firms defending rates and realization, and bankruptcy fee examiners reviewing fee applications, with the bankruptcy work documented in detail against the US Trustee Program's 2013 guidelines. The analytical coverage is specific too, with the flag set organized into staffing efficiency, workflow efficiency and billing hygiene and individual rules named. What is not stated is the boundary: nothing says which billing formats, jurisdictions, languages or alternative fee arrangements the analysis handles poorly, or what happens with billing that is not hourly.