AI Assistants — the Top 15 and What to Learn from Them

As of 11 October 2026. A review of the best existing AI assistants — open and closed — chosen by market traction and strength of idea. Purpose: to learn the technical and market lessons before I build my own assistant — for myself and for Polish professionals who cannot send client data to the USA (a local machine on the desk keeps all the data).

⚠️ This review goes stale within weeks. Numbers (users, revenue, valuations) carry a date and a source; figures published by the companies themselves I mark (company data). GitHub stars from the GitHub API on 11 October 2026. I tested nothing here on my own machine — this is a review of documentation, announcements and analyses. The legal parts are educational material, not legal advice.

---

Table of contents

• Market map: four kinds of assistants
• The fifteen
• The 10 most important lessons
• What this means for a local assistant for Polish professionals
• How I chose
• Files in this directory
• Related parts of the repository

---

Market map: four kinds of assistants

Market details, numbers, acquisitions and failures: 02-MARKET.md.

---

The fifteen

Score 0–100 using the weights from the section How I chose. Full profiles (9 areas each): 01-PROFILES.md.

Who did not make the fifteen but is worth knowing: 01-PROFILES.md.

---

The 10 most important lessons

1. Trust beats features. The best professional assistants (Abridge, OpenEvidence, Harvey) tie every answer to a source. For a lawyer and a doctor, "show where you know this from" matters more than eloquence.
2. Workflow, not chat. Granola, Cursor, Abridge win because they sit where the person already works — in the editor, in the consulting room, in the meeting — and not in a separate chat window.
3. Context is engineering. Manus and OpenClaw show that an agent's quality depends on what goes into the context and in what order (a stable prefix, a plan file, memory in files) — especially important with a local model.
4. The channel is the product. "Always-on" assistants work through messengers (Telegram, WhatsApp, iMessage) and in the background; for an ordinary user that is more natural than an app.
5. Autonomy needs brakes. Approvals scaled to external consequences, a sandbox per task, secrets outside the prompt. The 722 OpenClaw security advisories in 2026 are the cost of fast growth without these.
6. A custom open-weight model is a moat. Harvey (Tenet) and Cursor (Composer) fine-tune open models on their own data — cheaper and better in their domain than a general frontier model.
7. Evaluation in practice, not in a benchmark. Sierra measures pass^k — whether the agent does the same thing well k times in a row. For an assistant that someone trusts with their clients, repeatability is what counts.
8. The business model follows the value. From seats to usage to pay-per-outcome; but sudden pricing changes (Cursor) and the cost of frontier models (Canva) can hurt.
9. Europe is underserved. Some of the giants' new agentic features skip the EEA, and professional secrecy (art. 6 of the Prawo o adwokaturze (Act on the Bar), art. 3 of the ustawa o radcach prawnych (Act on Legal Advisers)) and § 23e of the Kodeks Etyki Adwokackiej (Code of Bar Ethics) require "due protection" of confidential information, while a cloud in the USA additionally carries CLOUD Act risk — this is a gap for a local product (04-POLAND-AND-SOVEREIGNTY.md).
10. Local can make sense. Open harnesses (OpenClaw, Hermes, browser-use) with a local model work without the internet for work on documents and files; the weak spots are fresh knowledge from the web, Polish voice and vision.

---

What this means for a local assistant for Polish professionals

• Start with one profession and one workflow — e.g. a law firm: review of contracts and pleadings with citations to the source; a consulting room: a visit note. Models: Harvey, Abridge, Noxtua.
• Everything on the desk: model, document search, transcription, memory, logs. Treat the cloud as an exception for data unrelated to the client. Architecture and the offline matrix: 03-TECHNICAL-PATTERNS.md.
• Citations and verification as a core feature — the Kodeks Etyki Adwokackiej (Code of Bar Ethics) § 23e (adopted by the NRA (Naczelna Rada Adwokacka, National Bar Council) on 12–13 June 2026; date of entry into force not established) requires personal verification of AI outputs. Legal details and competitors' prices: 04-POLAND-AND-SOVEREIGNTY.md.
• Price anchors in Poland: Beck-Noxtua PLN 590 net/month per user, Noa Notes PLN 199/month + VAT — a local assistant with one-off hardware has to compare with this.
• My rig (4 slots of DeepSeek-V4-Flash, text only) is good for a prototype; for a client, a smaller, quiet machine for the desk is what counts (hardware and prices).

---

How I chose

I reviewed about 60 candidates in five research areas (everyday, executors, always-on, professionals in law and finance, professionals in medicine, code and other professions) plus the Polish and European market. Each received scores of 1–5:

Score = 20 × (0.35·T + 0.25·I + 0.25·L + 0.15·A). Constraints: each of the four types at least twice, current products only, both open and closed, at least one European example betting on privacy (Noxtua). Ties are broken by traction, then by editorial order (ChatGPT, Gemini and Cursor have 87 pts each and T=5). Next in line: UI-TARS 76, Legora 74, Heidi 74, M365 Copilot 74, Genspark 72, Poke 72.

---

Files in this directory

Related parts of the repository

• Agent harnesses for local models — a technical review of harnesses, including Hermes Agent.
• Course: agentic applications — architectures, memory, proactivity, channels, security.
• Course: local models — hybrid of a local model + Claude.