On Friday, January 30, Anthropic released industry-specific plugins for Claude Cowork, its agent that reads files, organises folders and drafts documents for a user. The plugins cover sales, finance, data, marketing and legal work. By Tuesday, February 3, investors had decided the launch was a threat to the companies that sell software and data to those same professions, and CNN reported that an exchange-traded fund tracking the software industry fell 5.69% that day, its worst day since April.
The sharpest moves were in legal and financial information. Thomson Reuters fell 15.83% on Tuesday, its biggest single-day drop on record; LegalZoom fell 19.68%; RELX, which owns LexisNexis, fell 14%; FactSet fell 10.51%. Several of those names recovered a little on Wednesday.
/ 01What happened
Cowork itself is not new. What changed is that Anthropic packaged it for specific lines of work. A general assistant that can draft a memo is a productivity tool; a legal plugin that can review a contract, or a finance plugin that can pull and summarise market data, starts to look like a substitute for a product someone already pays for by the seat.
That is the reading the market took. CNN quoted Thomas Shipp, head of equity research at LPL Financial, summarising investor logic as “Why do I need to pay for software, the thinking goes.” Toni Kaplan, an equity analyst at Morgan Stanley, said the launch “adds to investors’ fear that AI-native companies will be able to break into the legal tech space.” Analysts at Aurelion Research called the sell-off “sentiment driven.”
Later in the week Anthropic released a new model, Claude Opus 4.6, and pitched it for financial research tasks such as screening and due-diligence data gathering. Fortune reported that the sell-off spread further across financial data, with FactSet down again and S&P Global, Moody’s and Nasdaq also falling. Gartner analysts, in a research note quoted by Fortune, struck a cooler tone: “Predictions of the death of SaaS and enterprise applications are premature,” while describing the plugins as “potential disrupters for task-level knowledge work.”
/ 02Why it matters
Markets price expectations, not deployments, and one week of trading says little about whether agents will replace legal research platforms or financial terminals. But the episode is a useful signal of where buyers and vendors think the line now sits. Agents have moved from answering questions to performing tasks inside a professional workflow: reading a folder of documents, extracting what matters, producing a work product.
That shift changes the risk profile. When a chatbot gets a fact wrong, a person usually catches it before it matters. When an agent drafts the contract summary, fills the spreadsheet and files the result, the error travels further before anyone looks. The companies that lost value this week sell, among other things, curated and verified content: case law that has been checked, financial data with a known lineage. That curation is exactly what a general-purpose agent working from a user’s files does not have by default.
/ 03The traceability question
For regulated work, the useful question is not whether an agent can produce a plausible draft, but whether every statement in that draft can be traced to a source a reviewer can check. A legal summary needs to point at the clause it summarises. A due-diligence note needs to show which filing each number came from. Without that, the reviewer has to redo the work to trust it, and much of the promised saving disappears.
This is where structure helps. If the documents an agent works from are first turned into a knowledge graph — parties, obligations, dates, amounts, each linked to the passage it came from — the agent’s output can carry those links forward. The reasoning becomes something that can be inspected and replayed rather than a fresh guess each time. It also gives an organisation a way to keep its own definitions (what counts as a “material change”, which entity is the counterparty) instead of inheriting whatever the model assumes.
An agent that does the work is only useful if a reviewer can check the work faster than redoing it.
/ 04What to watch
- Evidence in the output. Watch whether agent vendors start shipping citations to specific passages and data rows as a default, not an option.
- Where the data lives. Professional firms will ask whether client files leave their environment when an agent processes them. Expect demand for deployments that run inside the customer’s own perimeter.
- Incumbents’ response. The data companies hit this week own curated, verified content. Their counter is likely to be agents grounded in that content, which is a reasonable bet if they can show the provenance.
The sell-off priced in substitution, but what legal and financial data vendors sell is checked content with a known lineage. An agent drafting due-diligence notes from a folder of files inherits none of that unless someone builds it.
For a diligence team, the useful graph links each figure in a memo to the filing, table and row it was read from, each counterparty to its legal entity, and each obligation to the clause that creates it. The Ingestion Hub turns those PDFs, spreadsheets and SQL extracts into an ontology that analysts refine themselves, so the firm’s own definition of a material change governs the output rather than the model’s guess. Because reasoning over the graph is deterministic, the same question about exposure returns the same answer and the same evidence chain next quarter, which is what a compliance reviewer will ask to replay. The reviewer checks links instead of redoing the research.
How we’d model it
Memo figure—derivedFrom→Filing rowObligation—createdBy→Contract clauseCounterparty—identifiedAs→Legal entity
- Anthropic's new AI tool sends shudders through software stocks — CNN (via ABC17 News) abc17news.com ↗
- Anthropic's Claude triggered a trillion-dollar selloff. A new upgrade could make things worse — Fortune fortune.com ↗
See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.