Wordsmith
A legal front door for in-house teams: a platform that captures legal requests from the rest of the business, triages and routes them, resolves the routine ones and records what happened. Requests arrive from Slack or email, are auto-triaged by priority, jurisdiction and commercial exposure, and are either answered by the system or routed to a named lawyer with context already assembled from connected systems such as Salesforce, Slack and a contract lifecycle tool. Around that sit contract review against the customer's own playbook, drafting through reusable Blueprints, an assistant, multi-step agents, repositories, reports, and a research module covering more than 130 jurisdictions across 60 countries that draws on curated government and regulatory sources with direct integrations to Companies House, EDGAR and Open Corporates for entity verification. Answers carry citations linking to the underlying legislation, and Reports links every extracted value to an exact clause and page so a reviewer can verify a cell and then approve it or flag it for review. The product is built to sit in the tools legal teams already use, through a Microsoft Word add-in distributed on Microsoft AppSource, a Slack app, Microsoft 365 and an MCP integration. Wordsmith runs on OpenAI, Anthropic and Google models through paid enterprise APIs, names all three publicly, and states that each is engaged under a zero data retention arrangement and that customer data is not used to train any model. Hosting is on AWS with EU data residency by default and US available. Wordsmith AI Ltd is an independent company registered in England and Wales, incorporated in October 2023 and backed by Index Ventures and General Catalyst. The platform is SOC 2 Type II certified, with ISO 27001 described as in progress.
Capability grades
All 15 axes, graded from public sources on the date shown. Hover a grade to see what the letter means on that axis.
AI Centrality
How much of the product is actually AI. Whether the machine learning is the mechanism the buyer is paying for or a feature layered onto conventional software, and whether the vendor is specific about which is which.
The machine learning is the mechanism the buyer pays for, and there is no conventional platform underneath it. The product is a four-step flow of receive, route, resolve and record, and the two steps carrying the value are both model-driven: auto-triage that classifies an incoming request by priority, jurisdiction and commercial exposure and routes it to the right lawyer, and resolution that answers or drafts without one. The vendor draws the distinction itself, arguing that technology which only makes a lawyer faster leaves the operating model unchanged. Review, drafting, the assistant, agents, research and intake are each a model capability; repositories and reports are records of what those produced rather than standalone systems. Company incorporated October 2023 with no legacy platform to layer onto.
Citation Accuracy and Hallucination Disclosure
Whether the vendor publishes measured accuracy on citations and assertions, grounds output to primary sources, and says plainly what its system does when it does not know. Legal has a documented public record of fabricated citations reaching filed briefs, so an untested claim of accuracy is not evidence.
Grounding is real and described in specifics, and no measured accuracy is published anywhere. On the evidence side: the research module states coverage of more than 130 jurisdictions across 60 countries drawn from curated government, regulatory and authoritative sources, with direct integrations to Companies House, EDGAR and Open Corporates for entity verification; the assistant returns a cited summary with links to the underlying legislation; the Word add-in returns answers described as cited and traceable; and Reports links every extracted value to an exact clause and page number with a click-through to verify. A reader can therefore open and check the source, which is the limb most of this market fails. What does not exist on any surface read on 31 Aug 2026 is a number: no accuracy figure, no hallucination rate, no benchmark, no described test set, no published evaluation of any kind. The vendor also lists smart web search as a source feeding legal answers without stating how a conflict between a curated source and a web result is resolved.
Autonomy and Oversight Model
What the system decides on its own, what a lawyer must approve, and whether the vendor documents where the review point sits. A tool that drafts under review and a tool that files without one are different products and different risks.
A real oversight posture with described review surfaces, short of the threshold that would make it a full control structure. The published model is that work reaches a lawyer only when it needs judgement, with routing to a named role shown as part of the flow. The review surfaces are concrete rather than asserted: in Reports a reviewer clicks a cell to verify it against the cited clause and page, then approves and locks the value or flags it for review; in the Word add-in every proposed redline can be accepted, rejected or improved individually. Against that, the company page describes chaining drafting into an end-to-end workflow that receives a request, drafts, routes for approval and sends with no manual step required, and nothing published states the threshold at which the system stops and escalates rather than resolving, or what happens after an output is found to be wrong. The proportion of requests resolved without a lawyer is presented as a headline statistic on the home page but the counter renders as zero, so the autonomy level is unquantified.
Operational and Outcome Evidence
Named, dated evidence that the product works in production at real firms or legal departments. Case studies with figures and identified customers count. Unattributed testimonials and launch announcements do not.
Named customers, named individuals and at least one figure, short of dates and method. Trustpilot appears with Anoop Joshi, Chief Trust Officer, alongside a stated 85 percent reduction in contract review time; the Financial Times with Dan Guildford, General Counsel; Belron with Janet McCarthy, General Counsel; and Multiverse with Tara Haig, General Counsel. A logo strip carries 25 named organisations including BT, Canva, Deliveroo, Nikon, Rakuten, Selfridges, Skyscanner and Starling, and dedicated customer story pages exist. Two things hold this at B beyond the missing dates and method. The home page outcome counters, covering share resolved without a lawyer, speed, outside counsel reduction and hours saved per lawyer, all render as zeroes rather than figures, so the platform's own headline results are not actually stated. And the same sentence about lawyers focusing on strategic work is attributed on that page to both Anoop Joshi and Janet McCarthy. Individual case study pages were not opened on 31 Aug 2026, so the date and method limbs are rebuttable.
Privilege and Confidentiality Posture
How client confidences are handled: attorney client privilege and work product treatment, segregation of one client matter from another, whether client data trains any model, and what the vendor commits to in writing rather than in marketing.
Substantive commitments across most of this ground, with the training promise sitting in policy rather than in a located clause. Privilege is addressed directly rather than skirted: the security page states the product is designed to maintain attorney-client privilege and that its architecture aligns with ABA Formal Opinion 512. The no-training statement is unusually absolute and repeated, extending on a dedicated article to customer data never being used to train any underlying model or included in any pre-training corpus. The third-party model position is the strongest read in this pull: OpenAI, Anthropic and Google are each named with the specific contractual instrument under which zero data retention applies, and prompts and outputs are stated not to be stored, logged or used for training by any provider. Clause 11.2 of the terms binds each party to hold the other's Confidential Information and to use it only for implementing the contract. Two gaps hold it at B. No training prohibition was located in the retrievable portions of the customer agreement, so the commitment a buyer can point to is a policy statement rather than a term, and the data processing agreement that the terms incorporate by reference sits on a trust portal that refused automated access. Separation between customers is asserted as complete with no co-mingling, but no mechanism is documented.
UPL and Professional Responsibility Posture
Whether the vendor is clear that it supplies a tool rather than legal advice, who its audience is, and how it addresses unauthorized practice of law, competence and supervision duties, and jurisdiction limits. ABA Formal Opinion 512 is the reference point.
A published position that engages with named guidance rather than reciting a disclaimer, short of the competence and supervision limb. The security page names ABA Formal Opinion 512 twice, once in stating that the architecture aligns with it and once in a FAQ answering whether legal professionals must disclose AI use, which tells the reader that obligations vary by jurisdiction and bar association, that some recommend disclosure for substantive use while others do not require it for internal tools, and that teams should review Opinion 512 and applicable local bar guidance. The same page states that the product supports professional legal judgment rather than replacing it, and a published article engages with SRA and bar association duties to prevent unauthorised disclosure of client information. What is missing: nothing published addresses a lawyer's own competence and supervision duties in operating the tool, no jurisdiction limits are placed on the advice line despite research coverage spanning 130 jurisdictions, and no statement that Wordsmith is not a law firm was located on any surface read on 31 Aug 2026, which matters for a product whose stated purpose is resolving legal requests without a lawyer.
AI Governance and Bias Disclosure
Published governance over model behaviour: who owns it inside the vendor, what is tested before release, and what is disclosed about disparate output across matter types, parties, or populations.
No governance position for model behaviour was located. Searched the home page, the security page and its full FAQ, the downloads page, the company page, the research, assistant, reports and Word integration pages, and the published articles on data privacy and contract review, on 31 Aug 2026. Nothing names an individual or role accountable for model behaviour, describes what is tested before a release ships, sets out a responsible AI framework or principles, or discloses anything about uneven output across matter types, counterparties or populations. The security disclosure is substantial but it is information security, which is a different subject and does not carry on this axis, and the ISO 27001 certification described as in progress is likewise a security standard rather than an AI management one. One limitation worth stating: the Vanta trust portal linked from the footer refused automated access, and if it carries an AI governance policy this grade is rebuttable on that document alone.
AI Safety and Data Stewardship
Retention, deletion, access control, and what happens to prompts and documents after they are processed. Whether the vendor states its subprocessors and its incident practice, or leaves the buyer to assume.
Substantive published policy across most of the ground, short of an openly reachable subprocessor list. Published on the security page: AES-256 at rest and TLS 1.2 or higher in transit; access restricted on a least privilege basis; SSO through Okta, Azure Active Directory and Google Workspace over OIDC and SAML 2.0 with MFA and role-based access control; annual independent penetration testing; and a formal incident response plan with confirmed breaches notified to affected customers within 72 hours together with details and remediation steps. The privacy policy adds a commitment to notify the customer and any applicable regulator of a suspected breach where legally required, and states that access is limited to those with a genuine business need under a duty of confidentiality. Retention is customer-controlled: data can be permanently deleted at any time, all customer data is permanently deleted on termination, and custom retention policies can be configured. A published article is candid that a platform must retain some information to function, which is more useful than an unqualified zero-retention claim. The gap is that no subprocessor list is published on the open site and no default retention period is stated; both the list and the incident response plan sit on the trust portal.
AI Liability and Recourse
What the vendor stands behind contractually when its output is wrong. Indemnities, caps, carve outs, insurance, and whether any of it is published or only reachable through a negotiated agreement.
Liability is addressed through a limitation clause that disclaims the exposure the product creates, with the indemnity running the other way. Clause 8 requires the customer to indemnify Wordsmith, its officers, directors, employees, agents and consultants against liabilities, costs, damages and losses arising from the customer's use of the product and from any breach of the terms, carved back only where the claim arises from Wordsmith's own failure to meet its obligations. Clause 9.1.1 then states that Wordsmith shall not in any circumstances have liability for losses or damages suffered by the customer, subject to exceptions at clauses 9.2 and 9.6. No indemnity running to the customer, no warranty on output and no insurance position was located. A retrieval limit to state plainly: the terms page returns no body text to automated fetching, so the agreement was read through the search index, and the text of clauses 9.2 and 9.6, which carry the exceptions and would ordinarily contain any cap, could not be retrieved on 31 Aug 2026. If either contains a vendor indemnity or an output warranty this grade moves.
Practice Systems Integration Depth
How deeply the product reaches into the systems legal work already lives in: document management such as iManage and NetDocuments, Word and Outlook, contract lifecycle management, matter management, e-billing, and court filing systems.
Real integrations documented, with depth described in one place and asserted in the rest. The Word add-in is the strongest evidence: it is distributed through Microsoft AppSource under product identifier WA200008590, and its integration page describes what it actually does rather than that it exists, covering running a playbook against a contract in the sidebar, accepting, rejecting or improving each redline in place, querying the repository by at-mention to insert approved clause language, and running research with jurisdiction and language selection. A Slack app has its own install route, and dedicated pages exist for Microsoft 365 and MCP. Research integrates directly with Companies House, EDGAR and Open Corporates. The illustrated request flow shows context pulled from Salesforce, Slack and Ironclad. What is missing: no document management integration such as iManage or NetDocuments was located, the Salesforce and contract lifecycle connections appear in an illustrative workflow rather than in documentation stating what moves and in which direction, and no developer or API reference index was located on the open site on 31 Aug 2026.
Deployment Model and Data Residency
Where the software runs and where the data sits. Multi tenant cloud, single tenant, private deployment, on premises, and whether region of residence is a published option or an enterprise conversation.
Residency is stated clearly and as a customer choice, and the tenancy model is not stated at all. The security page publishes hosting on AWS with EU data residency by default and US available, and adds a commitment that data is not transferred outside the customer's chosen region without explicit consent, which is a stronger statement than most vendors make because it binds the vendor to the customer's selection rather than merely offering one. The underlying AWS certifications are named as ISO 27001, SOC 1, SOC 2, SOC 3 and PCI DSS Level 1. Three things hold this at B, all checked 31 Aug 2026. Nothing published states whether the platform is multi-tenant or single-tenant, and separation between customers is asserted without a tenancy model behind it. No private, single-tenant or on-premises option was located. And where processing happens as distinct from where data is stored is not addressed: three US-headquartered model providers process prompts, and the relationship between that processing and the customer's chosen region is not explained.
Security Certifications and Trust Center
Independent attestation a buyer can pull without a sales call: SOC 2, ISO 27001, penetration test summaries, a trust center with current reports and named scope rather than a badge image.
Certification is real and stated and a trust portal is openly linked, but no scope, date or auditor is available on any readable surface. The security page states SOC 2 Type II certification described as independent, GDPR compliance, annual independent penetration testing, and ISO 27001 in progress. A Vanta-hosted trust portal is linked from the site footer without a gate, and the vendor states that provider policies, the data processing agreement and the incident response plan sit there. That portal counts as a genuine access route and is credited here as reachable: it refused automated access through its robots file, which is a retrieval limit on the index's side and not a gap on the vendor's, so nothing is scored against Wordsmith for it. What could not be established from any surface read on 31 Aug 2026 is the substance an attestation is judged on: no auditor is named, no coverage period is given, no report scope is described, and no penetration test partner or summary appears. A badge stating SOC 2 Type II with no scope and no date is what the open site offers, and the portal is where the answer presumably sits.
Model Supply Chain Disclosure
Which models sit underneath, whose they are, where they run, and whether the vendor commits to telling customers when that changes. A legal buyer inherits every dependency it cannot see.
The providers are named with unusual precision and the change commitment is absent. The security page identifies all three model providers, OpenAI, Anthropic and Google, states that each is used through a paid enterprise API, and names the specific contractual instrument under which zero data retention applies to each: OpenAI's API Data Usage Policy, Anthropic's Commercial Terms, and the paid API Gemini Terms. Naming the instrument rather than asserting a commitment in the abstract is more than almost any vendor does, and it lets a buyer check the upstream terms directly. Two gaps, checked 31 Aug 2026. No specific model or version is identified for any of the three, so a buyer knows whose models but not which, which matters for a product whose research module also feeds in live web results. And nothing published commits to notifying customers when the provider set or the models change, which for a multi-provider architecture is the disclosure a customer most needs.
Commercial Transparency
Whether a buyer can learn what this costs without entering a sales process: published rates, the unit being charged, what sits behind an enterprise tier, and what implementation adds.
No pricing information is published at any level, including the unit of charge. Searched on 31 Aug 2026: the main navigation, which carries no pricing entry; the footer Product section, which lists downloads, comparisons, ROI and security and no pricing; the home page, the security page, the downloads page and the company page. Every call to action across the property is to book a demo. Neither a figure, a tier structure, a per-seat or per-matter unit, nor a statement of what implementation adds was located. The agreement confirms that a priced self-service route exists, since the self sign-up terms authorise charging fees by credit card, debit card or automated clearing house on online purchase, but no number is published anywhere a prospective buyer can reach without entering the signup flow or a sales process. The ROI page in the footer was not opened, so this is rebuttable if it carries a rate rather than a calculator.
Firm and Practice Coverage
Who the product is actually built for. AmLaw, midlaw, small firm and solo, in house departments, government and courts, and which practice areas are supported rather than merely claimed.
Segment coverage is described with real substance and the boundaries are left open. The property segments explicitly and in dedicated pages: by role for General Counsel and Legal Operations, by requesting team for procurement, sales, security and human resources, and by size across small business, mid-market and enterprise. Geographic and legal coverage is stated concretely as more than 130 jurisdictions across 60 countries, and the site publishes locale variants for eleven countries in five languages. What is absent is any statement of where the product stops. Law firm segments are not addressed, which is consistent with a product built for in-house teams, though external counsel appear only as recipients of shared reports and the position is never stated outright. Government and court use is not addressed. And practice coverage is expressed as the requesting business function rather than as areas of law, so a reader learns which internal team a request comes from but not which areas of law the product is competent in, on a platform whose research module claims 130 jurisdictions. Checked 31 Aug 2026.
Legal Signals
What each signal meansA signal records what public sources say on the date shown. It is not a grade and it is not a recommendation. Where a signal reads Not addressed, it means the index did not locate the material in public sources on that date, which is a statement about disclosure rather than about the product.
Client Data in Training
Can material a lawyer puts into this product be used to train a model?
A public policy or trust page states no training on customer content, with no matching term located in the published agreement.
The security page states that customer data is not used to train Wordsmith's own models or to improve the platform for other customers, and a published article extends that to customer data never being used to train any underlying model or included in any pre-training corpus. Separately, zero data retention is stated to be contractual with the three model providers, each named with its instrument. That contract is with OpenAI, Anthropic and Google rather than with the customer. No training prohibition was located in the retrievable portions of the customer agreement on 31 Aug 2026; the terms page returns no body text to automated fetching and was read through the search index, and the data processing agreement the terms incorporate by reference sits on a trust portal that refused automated access. The value is bounded by that and is rebuttable on the DPA.
Prompt and Output Retention
How long does the product keep what a lawyer typed, and can that be set to zero?
The customer controls the retention window, by product configuration or by contractual instruction, but zero retention is not stated as available.
The security page states that customers retain full control of their data, that it can be permanently deleted at any time, that all customer data is permanently deleted from Wordsmith's systems on termination, and that custom retention policies can be configured to match internal data governance requirements. No default retention window is published for prompts or outputs. A published article distinguishes zero training from zero retention and states plainly that a platform must retain some information in order to function, so no-retention is not offered as a setting at the Wordsmith layer; the zero retention that is stated applies to the model providers, where inputs and outputs are said not to be retained after processing.
Ethical Walls and Matter Segregation
Does retrieval respect the firm’s ethical walls, or can the model read across them?
Segregation is asserted in public materials with no published detail on how it is enforced.
Separation is asserted at customer level on the security page, which states that data is kept completely separate from other customers with no co-mingling and no cross-contamination, supported by SSO through Okta, Azure Active Directory and Google Workspace and role-based access control. Within a customer, a published article states that playbooks and templates are shared across the organisation only where the user chooses to share them. How either boundary is enforced is not published: searched the home page, the security page, the company page and the product and integration pages on 31 Aug 2026 and located no tenancy model, no description of how retrieval applies access controls at query time, and no statement of whether the repository can be read across internal teams. The buyer here is an in-house department, so the relevant test is tenant-level separation rather than matter-level walls.
Third Party Request and Subpoena Notice
If someone subpoenas the vendor for a firm’s data, does the firm hear about it first?
This signal has not been recorded for this vendor yet. It is not a finding either way.
Primary Law Corpus Provenance
Where does the law in this product come from, and does the vendor have the right to use it?
Coverage is described by jurisdiction with no identification of the underlying corpus.
Coverage is described by jurisdiction without identifying the underlying legal corpus. The research module states more than 130 jurisdictions across 60 countries with access to primary legislation, databases and guidance from hundreds of trusted legal sources, described as curated government, regulatory and authoritative sources that are continuously updated, supplemented by smart web search. No individual legislation or case law source is named and no licence or public domain basis is stated for any of it. Three sources are named, but they are corporate registries used for entity verification rather than legal authority: Companies House, EDGAR and Open Corporates. No update cadence or lag is published beyond the word continuously.
Good Law Verification
Does the product tell you when the authority it just cited has been overruled?
No located public material addresses whether authority is checked for subsequent history.
Searched the home page, the security page, the company page, and the research, assistant, reports and Word integration pages on 31 Aug 2026. Nothing addresses whether authority returned to the user is checked for subsequent history, and no citator, treatment signal or good law indicator was located. This signal applies squarely to this product rather than being out of scope, because the research module retrieves primary legislation and returns cited answers linked to it. The nearest published material concerns currency of sources rather than status of authority: sources are described as continuously updated and smart web search is said to ensure answers reflect the most recent developments, neither of which tells a reader whether a cited provision or decision still stands.
Refusal and Uncertainty Behaviour
What does the product do when the answer is not in the corpus?
No located public material addresses what the product does when it cannot ground an answer.
Searched the home page, the security page, the company page and the product pages for research, assistant, agents, reports and the Word add-in on 31 Aug 2026. No explicit no-answer or abstention path is documented and no confidence or grounding score was located. Two features come closest and neither is the same thing: the triage flow is described as surfacing work to a lawyer only when it needs their judgement, which is routing rather than abstention, and Reports allows a reviewer to flag a value that needs review, which is a human action on an answer already given rather than the system declining to give one.
Fabricated Citation Record
Does a public court record exist involving output from this product?
No court order, opinion or disciplinary record naming this product has been located as of the date shown. This is a statement about the public record, not a finding about the product.
No court order, opinion or disciplinary record naming this product has been located as of 31 Aug 2026. Instrument searched: the AI Hallucination Cases database maintained by Damien Charlotin, which tracks decisions worldwide where a court addressed hallucinated AI content and records the tool implicated where known, searched on both the product name and the company name Wordsmith AI Ltd, alongside 2026 sanctions summaries in the trade press. This is a statement about the public record on the date shown rather than a clearance, and it is bounded by what that database covers. The product is sold to in-house departments rather than to litigators, so its output does not commonly reach a filed brief.
Bar Guidance Alignment
Has the vendor engaged in public with the ethics opinions its buyers are bound by?
Public materials engage with at least one named ethics opinion.
Public materials engage with a named ethics opinion. The security page cites ABA Formal Opinion 512 twice: once in stating that the product's architecture aligns with it, and once in a FAQ answering whether legal professionals must disclose AI use to clients, which tells the reader that obligations vary by jurisdiction and bar association, that some recommend disclosure for substantive AI use while others do not require it for internal tools, and that teams should review Opinion 512 together with applicable local bar guidance. A published article separately engages with SRA and bar association duties to take reasonable steps against unauthorised disclosure of client information. No mapping of the product to named opinions across more than one jurisdiction was located, so local guidance is referred to generically rather than identified.
Billing and Fee Posture
Does the vendor address what happens to the bill when the work takes an hour instead of six?
Public materials claim time savings without addressing billing or disclosure.
Public materials are built around time and cost saved. A customer story states an 85 percent reduction in contract review time, a published article claims contract processing times reduced by 50 to 60 percent, Reports is described as auditing more than 2,800 agreements in days rather than weeks, and the home page presents hours saved per lawyer each week and reduced outside counsel spend as headline outcomes, though those counters render as zeroes. Searched the home page, security page, company page, product pages and articles on 31 Aug 2026 and located no per matter record of AI-assisted work intended for fee or disclosure purposes and no published guidance on billing treatment. The buyer is an in-house department rather than a firm billing a client, so the question lands on outside counsel spend rather than on the hourly bill, but nothing addresses either.
Outside Counsel Guideline Readiness
Can a firm get this vendor through a client’s AI clause without a bespoke negotiation?
A current subprocessor or model provider list is published.
A model provider list is published without a gate. The security page names all three providers that see customer content, OpenAI, Anthropic and Google, states that each is accessed through a paid enterprise API, and identifies the specific contractual instrument under which zero data retention applies to each. That is the disclosure most useful to a legal team answering a client AI clause, and it is readable before any agreement is in place. What could not be confirmed on 31 Aug 2026 is anything beyond it: no full subprocessor list appears on the open site, and the data processing agreement, the provider policies and the incident response plan are all stated to sit on the Vanta trust portal, which is linked openly from the footer but refused automated access, so its contents were not verified and no client-facing consent or notification pack could be confirmed to exist.
Court Disclosure Support
If a judge’s standing order requires an AI disclosure, can the product produce one?
Some elements of the record are available, short of a document level export.
Some elements of a record exist, short of a document-level export. Reports links every extracted value to its source with an exact clause and page number, and a reviewer clicks through to verify it and then approves and locks the value or flags it for review, which produces a per-value record of both the source relied on and the human check. The Word add-in similarly records each proposed redline as accepted, rejected or improved. The missing element is the model: which of the three providers produced a given passage is not disclosed anywhere on the property, and no export designed to cover model used, sources retrieved and human verification together was located on 31 Aug 2026. The product serves in-house departments rather than producing court filings, so a judicial standing order is not its usual context.