4 AI drops worth watching: May 15
Anthropic: Claude for Small Business
On May 13, Anthropic launched Claude for Small Business, a package that connects Claude to QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, and Microsoft 365. Fifteen ready-to-run agentic workflows ship at launch across finance, operations, sales, marketing, HR, and customer service. A nationwide tour starts in Chicago on May 14, alongside a free “AI Fluency for Small Business” course developed with PayPal.
Anthropic is meeting small business owners inside the tools they already use, rather than asking them to learn a new chat interface. The 15 workflows are prebuilt playbooks for routine work like running payroll, closing the monthly books, or chasing invoices.
The take. The bet is distribution, not capability. Most small businesses will not adopt AI on its own merits; they will adopt it inside QuickBooks because their accountant told them to. The connector list reads like a map of every category small businesses live in, which is the point. The free education course suggests Anthropic thinks the bottleneck is fluency, not features. Watch whether the prebuilt workflows do real work or just demo well: if half of these have meaningful adoption six months from now, this becomes a defining playbook for embedded vertical AI.
OpenAI: Codex everywhere
On May 14, OpenAI made Codex accessible via the ChatGPT mobile app. Builders can monitor, steer, and approve coding tasks in real time across devices and remote environments. The release is part of a wider week-long Codex expansion that also moved the agent onto the web and into sandboxed Windows.
Coding agents used to require a desktop terminal. Now one can run on a remote machine while the developer is on the bus, with the phone acting as a control surface for approvals and course corrections.
The take. The interesting shift is from “agent in a window” to “agent in your pocket”. Coding agents have a real bottleneck around long-running tasks: builders want to fire them off and check progress later, not babysit a chat. Mobile makes that natural. The competitive question is whether async + mobile becomes the default shape of AI dev tools. Anthropic’s Claude Code is moving the same direction; the field will converge inside a quarter. The understated part of this announcement is the remote-environment piece, which is what separates “neat” from “production-relevant”.
IBM: Granite Embedding R2
On May 14, IBM Granite released two new multilingual embedding models under Apache 2.0. The 311M-parameter model scores 65.2 on MTEB Multilingual Retrieval, second among open models under 500M parameters. The 97M-parameter sibling scores 60.3, the highest among sub-100M open multilingual models. Both support 200+ languages with enhanced retrieval on 52, code retrieval across 9 programming languages, and a 32K-token context window.
Embeddings are how AI apps find relevant text chunks in big collections. IBM just open-sourced two of the strongest small ones, with permissive licensing, working across 200+ languages and 9 programming languages, on hardware that does not need to be enormous.
The take. The numbers matter for a specific kind of builder: anyone running retrieval-augmented generation on edge hardware, multilingual customer support, or any system where 1B-parameter embeddings are too expensive on latency or cost. Sub-100M open multilingual that scores 60+ on MTEB is a real unlock for those use cases. The 32K context window aligns with where RAG is heading, which is less retrieval and more direct stuffing of long documents. IBM is one of the few large vendors shipping enterprise-grade open releases under permissive licenses with this level of language coverage. Quietly important if retrieval is in the stack.
Ardent: Postgres sandboxes for coding agents
On May 13, Ardent (YC P26) launched. The tool spins up production-like Postgres clones in under six seconds at terabyte scale, using logical replication and DDL triggers on top of branching engines like Neon. Clones run with copy-on-write and a proxy layer for per-clone credentials and access control. PII redaction runs on branches via registered SQL. BYOC supports full data residency.
Coding agents that touch databases are dangerous because they can break production. Ardent gives each agent its own real-data copy to test against, spun up in seconds and with sensitive fields scrubbed.
The take. Agent + database safety is currently solved by either denying access (which limits agent usefulness) or by accepting risk (which limits deployments). A fast clone-per-agent approach is the obvious third path; the engineering question is whether it is cheap enough that teams actually use it. Sub-six-second clone time at TB scale is the load-bearing claim. If it holds at production scale, this becomes a default for any team running coding or analytics agents against real schemas. The strategic dependency on Neon as the branching primitive is worth watching, since the moat lives in the proxy, redaction, and orchestration layer above it.
More drops at dropwatch.ai. Want them in your inbox? Subscribe below.