Where Lawyers Are Working

Legal AI raised billions on the premise that lawyers need purpose-built legal software. This paper tests that premise against evidence from a place nobody is watching: what small and mid-sized firm lawyers actually say to each other, all day, for twenty-nine months.

We conducted a full-corpus census of 48,696 messages across 54 channels in the Best Era community Slack, a private network of solo, small, and mid-sized law firm owners and their teams. Every message was tested against fixed term sets for four AI ecosystems. No sampling was used.

The result is a single, steep curve. Anthropic-family mentions rose from 16 messages in 2024 to 1,189 in the first eight months of 2026 — a 74-fold raw increase, and a 37× increase after normalizing for the community's own growth in message volume. In the community's dedicated #ai channel, the relationship between Anthropic and OpenAI inverted completely: from 4.4% versus 19.1% of messages in 2024 to 19.2% versus 3.1% in 2026.

The rise is not concentrated in a handful of enthusiasts. The share of active members who discussed Claude in a given half-year grew from 1.7% to 32.6%. The share of active days containing at least one Anthropic mention grew from 5% to 87%. The conversation spread from 5 channels to 16.

Most consequentially, the character of the conversation is different from the character of ordinary conversation about software. Messages mentioning Claude are 49% longer than the corpus mean, 4.1 times more likely to run past 1,000 characters, draw 34% more thread replies, and are 5.5 times more likely to contain the phrase “I built.” Lawyers are not evaluating a vendor. They are reporting construction.

Against this, the two most highly valued legal AI companies in the world — Harvey, at an $11 billion valuation, and Legora, at $5.6 billion — together account for 82 messages from 15 members over the same period. Claude accounts for 1,393 messages from 71 members. The mindshare ratio is 17 to 1.

Our thesis is not that Harvey and Legora are overvalued. It is that the market is measuring the wrong end of the profession. What starts small grows big. The general-purpose tool, adopted from below by lawyers with no procurement committee and no pilot program, is where the work is actually moving.

Almost everything published about AI adoption in law measures one of two things: what large firms buy, or what lawyers say on a survey. Both are lagging indicators. Procurement follows a budget cycle. Surveys ask lawyers to summarize their own behavior in retrospect, to a stranger, in a single sitting.

This corpus is different in three ways that matter.

It is behavioral, not declarative. Nobody in these channels was answering a question about AI adoption. They were asking each other for help, showing each other work, and complaining. The record is a byproduct of doing the job.

It is longitudinal and continuous. Twenty-nine months, 860 days with traffic, no gaps. The same people, in the same rooms, across the entire period in which frontier models became practically useful for legal work.

It is the segment nobody instruments. The community's 504 member accounts are overwhelmingly solos, small firms, and mid-sized firms — the part of the profession that has no CIO, no innovation officer, and no six-figure software budget, and which therefore adopts tools the way consumers do: individually, immediately, and without a pilot program. Posting members span 12 time zones on three continents, and their stated practices range from personal injury, criminal defense, family law, immigration, employment, trademark, intellectual property, and trusts and estates to real estate closings. Roles include founding and managing partners, solo practitioners, firm administrators, intake managers, marketing staff, and legal nurse consultants.

That segment is where most American lawyers work. It is also where nearly all American legal consumers are served. If AI is changing legal practice, the change is visible here first and most honestly.

The source is a standard Slack workspace export covering 15 April 2024 through 27 August 2026. Every message object of type message across all 54 public channel directories was parsed — 48,696 in total. For each message we retained channel, date, author, text, attachment titles and link previews, reaction counts, and thread reply counts.

Each message's text was concatenated with its attachment and link-preview text, lowercased, and tested against four fixed regular-expression term sets representing product families: Anthropic, OpenAI, Google AI, and legal-specific AI vendors. The exact patterns are reproduced verbatim in Appendix B so that any reader can reproduce the counts.

A message counts once for a family regardless of how many times a term appears in it. All figures are message counts, not term counts. This is a deliberately conservative choice: it means a member who writes four paragraphs about Claude scores the same as one who writes “Claude?”

Because the community's own message volume more than doubled over the period, raw counts overstate growth. Wherever a trend claim is made, we report the normalized rate — mentions per 1,000 messages in the same interval — alongside the raw count.

Reaction counts are reported as observed in the export and are a floor, not a ceiling: reactions added after the export date are absent, and the export does not distinguish reactions on thread replies from reactions on parents.

For twenty months, Claude was a rounding error in this community. Then, over roughly six months, it became the default subject.

In calendar 2024, sixteen messages out of 8,811 mentioned Claude or Anthropic — a rate of 1.8 per thousand. In 2025 the rate rose to 8.5 per thousand. In the first eight months of 2026 it reached 67.3 per thousand: 1,189 messages.

The step is not gradual. Two inflections are visible in Figure 1, and both are legible as product events rather than as fashion. The first, in May 2025, is the month Claude Code became widely available and the word “Anthropic” enters the corpus at all. The second and much larger step begins in January 2026, when agentic desktop use arrived and the community's vocabulary shifts within weeks — from asking what Claude is to describing what they have built with it.

Raw counts flatter Anthropic here, because the community itself grew. Figure 2 removes that effect by expressing every quarter as mentions per 1,000 messages posted in the same quarter. The shape survives normalization intact.

* 2026 covers January 1 – August 27 only. The 2026 rate is therefore a rate over a partial year, not an annualized projection.

The clearest evidence is not that Claude rose. It is that in the room dedicated to this exact subject, Claude and ChatGPT traded places.

The community's #ai channel is the single best-controlled subset in the corpus: 8,397 messages whose entire purpose is discussing artificial intelligence. Within it, in 2024, ChatGPT and OpenAI appeared in 19.1% of messages and Anthropic in 4.4% — a ratio of roughly one Claude message for every four about OpenAI. By 2026 the figures are 3.1% and 19.2%. The ratio is now better than six to one the other way.

This is a 27-fold swing in relative share inside a fixed forum, among substantially the same people. Note also what happened to Google: NotebookLM and Gemini peaked in 2025, reaching 6.0% of the channel, and then declined to 2.7% — a reminder that this community will pick a tool up and put it down, and that Claude's trajectory is not simply novelty-seeking.

A loud minority can manufacture a curve. Three independent tests say this is not that.

Test one: how many people. In the first half of 2024, exactly one member of 59 active had mentioned Claude. By the first half of 2026, 60 of 184 active members had — 32.6%. Across the full window, 71 distinct members discussed Claude, out of 313 who ever posted anything.

Test two: how many days. In 2024, 13 of 256 active days contained an Anthropic mention (5%). In 2025, 95 of 365 (26%). In 2026, 208 of 239 (87%). The subject is no longer episodic. It is ambient.

Test three: how many rooms. Anthropic mentions appeared in 5 channels in 2024, 11 in 2025, and 16 in 2026 — escaping #ai and #legaltech into #marketing, #hiring, #money, #referrals, #pi-channel, #seo, #vendors, #systems, and #wins. A tool discussed in the hiring channel is no longer a tool. It is infrastructure.

Counting mentions tells you what people talk about. Measuring the messages tells you how they feel about it.

Six behavioral measures separate Claude messages from the corpus baseline. Every one of them points the same direction, and the largest gap is the most telling.

Claude messages are longer: median 180 characters against 124 for the corpus, and 4.52% run past 1,000 characters against 1.10% overall — a four-fold elevation in the long-form share. They generate more discussion: threads started by a Claude message draw 6.72 replies on average against 5.00 for all threads. They are far more interrogative: 24.7% contain a question mark against 13.6% of all messages, meaning lawyers are teaching each other rather than announcing conclusions.

And they are 5.5 times more likely to say “I built.” Build language — “I built,” “we built,” “I made,” “I created,” “spun up” — appears in 3.09% of Claude messages and 0.56% of all messages. For comparison, OpenAI-family messages sit at 2.4%.

This is the finding that a mention count cannot reach. In this community, other software is discussed the way software is normally discussed: as a purchase, a comparison, a complaint. Claude is discussed the way a workshop is discussed. The dominant speech act is not evaluation. It is show-and-tell.

Base: 48,696 all messages; 1,393 Claude messages. Thread measures use 5,212 and 310 thread parents respectively.

The single most surprising result in this study: the largest category of legal Claude use is not drafting. It is building.

We classified all 1,393 Claude-mentioning messages against fourteen non-exclusive use categories. A message can fall into several — a lawyer describing an intake dashboard hits three at once — so shares sum past 100%. The ranking is what matters.

Software construction and custom tooling appears in 32.5% of Claude messages. Automation, agents, and scheduled work appears in 18.8%. Together they eclipse document drafting (10.8%), which is the category every legal AI product is designed around, and legal research and citation checking (8.4%), which is the category every legal AI product markets.

This is an inversion of the industry's assumed hierarchy. The legal AI market was built on the premise that lawyers want a better brief. The evidence here is that a meaningful slice of lawyers, given a general-purpose tool, would rather build the thing that writes the brief — and then keep the thing.

What the categories look like in practice

Counts establish the shape. The following are verbatim from the corpus, anonymized per the note in §11. They are selected to illustrate the categories in Figure 6, not to represent typical enthusiasm.

In 2026, the two most valuable legal AI companies in the world are worth a combined $16.6 billion. In this community of roughly 500 practicing lawyers, they are worth 82 messages.

Harvey raised at an $11 billion valuation in March 2026 and was reported in August to be raising again at $15.5 billion. Legora tripled its valuation to $5.55 billion with a $550 million Series D in the same month. These are the flagships of the thesis that law needs its own AI.

Across 48,696 messages and 29 months, Harvey appears in 77 messages from 15 distinct members. Legora appears in 14 messages from 7 members. Claude and Anthropic appear in 1,393 messages from 71 members. The mindshare ratio against Harvey and Legora combined is 17 to 1 by message and roughly 5 to 1 by distinct member.

Two honest qualifications. First, this is not the segment Harvey and Legora sell to; they sell to large firms and corporate legal departments, and a community of solos and small firms is not their market by design. Second, purpose-built legal vendors do appear here in force — just not the ones with the headline valuations. Supio (164 messages, 36 members), EvenUp (78, 22), and CoCounsel (60, 27) all show substantially broader member reach than Harvey, because they serve personal injury and litigation work that this community actually does.

That second point sharpens rather than weakens the finding. When these lawyers pay for legal-specific AI, they pay for narrow tools that do one workflow. The general-purpose model is what they reach for everywhere else — and “everywhere else” is expanding.

Valuations as reported publicly and cited in References. Anthropic's is its May 2026 Series H figure and is included for scale, not for comparability — Anthropic is not a legal software company.

A study that finds only what its author hoped to find is not a study. Nine percent of Claude messages in this corpus contain the vocabulary of failure, risk, or constraint, and that material is the most useful part of the record.

Hallucination, caught by the lawyer

“Claude hallucination issue that I caught because I'm an ethical lawyer, but beware… I downloaded all of the cases I wanted to use in my oral argument notebook… Well, one of the cases that I know the facts very very well were all wrong… I can only imagine however that more and more attorneys that just blatantly trust Claude and other AI systems are going to seriously get themselves in trouble and perhaps even innocently so.”

Community member · #ai · 8 March 2026 · 18 thread replies

Privilege and confidentiality are unsettled, and the community knows it

The most-replied privilege question in the corpus was three lines long: “Dumb question: how are you using Claude Co-Work and protecting privileged information?” It drew 34 replies. Another member's warning was blunter:

“I would check with your IT/Network provider as Claude Co-Work can access every single document on your system, create false prompts and create permanent damage. The appeal is real, but the security threats are next level.”

Community member · #ai · 23 March 2026

Members also circulated 2026 federal decisions reaching opposite conclusions on whether conversations with AI tools are privileged — which is to say this community is tracking the case law faster than most bar associations are writing guidance.

The purpose-built tools sometimes simply win

The author's own most-replied message on the subject is a concession, and it drew 35 replies:

“Just used Westlaw's Co-Counsel 2.0 and it was impressive. Legal specific platforms like this are efficient. It is easier than trying to coax claude into doing what I want it to do.”

Ryan McKeen · #ai · 19 April 2026 · 35 thread replies

Usage limits are a real operational constraint

“When a usage limit is reached in Claude Cowork so that I'm blocked out for a couple of hours, can I disconnect from the internet long enough to take the lap top home…? I'm sitting in my office scared to leave before my usage is unblocked at 8:50.”

Community member · #ai · 6 August 2026 · 26 thread replies

Context management, token cost, lost project state after a plan change, connectors that need reconnecting, and a Word integration that would not authenticate all appear repeatedly. This is not the conversation of a community in a honeymoon. It is the conversation of a community that has committed and is now dealing with consequences — which is a considerably stronger signal.

Every important technology in the history of the legal profession entered from below and was dismissed on the way in.

The photocopier was a convenience for secretaries before it restructured discovery. The fax machine was undignified before it became the only acceptable way to serve a document. Email was a toy, then it was the file. Westlaw on a terminal was a curiosity next to a wall of reporters. In each case the profession's serious institutions evaluated the serious product, and the actual change arrived through the cheapest general-purpose version of the thing, adopted by individuals who did not ask permission.

The data in this paper describes that pattern in progress, with timestamps.

What makes this cycle different from the fax machine is the direction of the connection. Prior tools sat beside the practice. This one reaches into it. One member, asked directly for this paper, put the mechanism precisely:

“It's connecting directly to the sources of truth: client files, email, calendar, notes, westlaw, attorney infrastructure (document forms, templates), and web browser. And its ‘artifacts’ (custom dashboards) are incredibly powerful. It was the first to do that. Now Google and ChatGPT are releasing competitors, but the first mover always has an advantage.”

Community member, statement to the author, August 2026

That is also the sentence that explains Figure 6. If a tool can reach the sources of truth, the highest-value thing to do with it is not to ask it for a document. It is to wire it to the practice — and wiring is building. The 32.5% build share is not lawyers playing at software engineering. It is lawyers discovering that the integration is the product, and that nobody is going to build their integration for them.

Three implications follow from the data, offered as hypotheses rather than conclusions.

1. Adoption in this segment is not a procurement event, so it will not appear in procurement data. The curve in Figure 1 was produced by individuals with a $20 to $200 monthly subscription. No RFP generated it. Any market map built from enterprise contracts will therefore systematically under-count the segment that employs most lawyers.

2. The moat is the loop, not the wrapper. The community reached this conclusion out loud and in public: in the author's own summary of a frontier-lab talk, posted to #ai in June 2026: “The harness doesn't matter in a year. The models get so good you don't need it… Your software stack doesn't matter much. If it can connect to Claude you are good.” The durable asset is the firm's own captured judgment — its files, its templates, its reasoning, written down where a model can read it. That asset does not transfer with a vendor switch, which is exactly why it is a moat.

3. Access to justice is the underpriced consequence. The foreclosure vignette in §7 is one message with thirteen reactions, and it describes a case that would not have been taken a year earlier. If the marginal cost of competent help falls far enough, the binding constraint on legal aid stops being lawyer hours. Nothing in this corpus proves that at scale. Everything in it suggests the experiment is already running, unfunded and unmeasured, in solo practices.

Sixteen months before that message, in April 2025, the same author wrote in a quieter channel: “The advances in the underlying tech (Claude, GPT, Gemini) are going to eat a lot of SAAS in the near future.” At that point the entire corpus contained four Anthropic mentions. The prediction is in the record with a date on it, which is the only reason it counts for anything.

This paper makes a strong claim from a single community, and the following constraints bound it.

One community, self-selected. Best Era members joined a paid network organized around firm growth and, latterly, AI. They are not representative of the American bar. They are plausibly representative of its early-adopting small-firm segment, which is the claim made here and no more.

Founder effect. The author is a co-founder of the community and its single most prolific voice on this subject — 441 of the 1,393 Claude messages, or 32%. Removing the author entirely leaves 952 messages from 70 other members, and every trend survives: the normalized rate among all other members still runs 2.05, 6.49, and 61.71 mentions per 1,000 messages across 2024, 2025, and 2026 — a 30-fold rise. Readers should treat the author as a participant in the phenomenon being measured, not an observer of it.

Mentions are not usage. Everything here measures talk. A lawyer who uses Claude silently every day contributes nothing; a lawyer who mentions it once contributes a data point. The direction of this bias is unknown.

Partial final period. 2026 data covers January 1 through August 27. The 2026 rate is a rate over eight months, not an annualized figure, and the 2026 H2 diffusion bar in Figure 4 covers two months against six for every other bar.

Term sets are imperfect. Regular expressions catch names, not concepts. A message describing an agentic workflow without naming a vendor is invisible to this method. Conversely, a member sharing a news link about Anthropic counts identically to one describing their own build. In 2026, 144 of 1,189 Anthropic messages (12.1%) are link shares carrying fewer than 60 characters of the member's own text.

Public channels only. Direct messages, private channels, and huddle audio are outside the export. The huddle transcript file was empty. Substantial technical discussion in this community happens in direct messages and is not captured.

Anonymity. At the author's direction, every quotation in this paper other than the author's own is anonymous. Speakers are identified only by channel and date. No firm size, practice area, jurisdiction, role, or seniority is given, and no member is described in a way that would distinguish them from the other 312 people who posted in this corpus. Two members consented to be named and were anonymized anyway, on the view that a paper about a private community should not make anyone identifiable as the price of illustrating a statistic.

Editorial marks inside quotations. No wording has been changed. A plain ellipsis (…) marks omitted text. A bracketed ellipsis ([…]) marks an identifying detail removed. Square brackets around words mark an identifying detail replaced with a neutral equivalent — used twice, both times to remove a practice area or geography that would have named the speaker. Every figure in this paper is computed from the unaltered corpus; these marks affect only the quoted illustrations.

Reproducibility. The corpus is a private Slack export and cannot be published. The patterns, category definitions, and every count in this paper are reproducible against any Slack export using the specifications in Appendix B, and the author will confirm individual figures on request.

All patterns are case-insensitive PCRE, applied to the concatenation of a message's text field with the titles, preview text, and URLs of its attachments. \b denotes a word boundary.

Product family term sets

Deliberate exclusions. \bopus\b and \bsonnet\b were excluded from the Anthropic set: both appear in the corpus in non-Anthropic senses (a poetry discussion, a business name), and including them would have added mentions without confidence. \bmcp\b, \bskill\b, and artifact were also excluded, since each is now used generically. Every exclusion lowers the Anthropic count.

Build language. \bi built\b | \bi've built\b | \bwe built\b | \bi made\b | \bi created\b | \bbuilt (a|an|my|this)\b | \bspun up\b | here'?s what i built | i wrote a

Problem and risk language (the 9% figure in §9). hallucinat | wrong | broke | broken | fail | frustrat | disappoint | usage limit | privileg | confidential | security | risk | ethic

Use category patterns (Figure 6)

Suggested citation

External sources

1. Bloomberg, “Legal AI Startup Harvey Raises Funds at $11 Billion Valuation,” 25 March 2026. bloomberg.com
2. CNBC, “Legal AI startup Harvey valued at $11 billion in funding round,” 25 March 2026. cnbc.com
3. SiliconANGLE, “Legal AI startup Harvey reportedly raising $500M at $15.5B valuation,” 7 August 2026. siliconangle.com
4. Crunchbase News, “Swedish Legal Tech Startup Legora Triples Valuation To $5.55B With $550M Series D Led By Accel,” March 2026. news.crunchbase.com
5. TechCrunch, “Legal AI startup Legora hits $5.6B valuation and its battle with Harvey just got hotter,” 30 April 2026. techcrunch.com
6. Legora, “Legora raises $550 million Series D to fuel US growth,” March 2026. legora.com
7. Value Add Pulse, “Anthropic Valuation, Funding & News — August 2026,” reporting a $965B Series H valuation (May 2026) and a $65B annualized revenue run rate (late July 2026). valueaddvc.com

Primary source

Best Era, LLC Slack workspace export, 15 April 2024 – 27 August 2026. 54 public channels, 48,696 messages, 504 member accounts. Held by the author; not publicly available.