The best AI agents for work, compared on what they are allowed to touch

Prices and features checked: 2026-11-09. Prices in the currency the vendor shows, checked on the date shown; we do not convert.

Verdict by user type

An "AI agent for work" is a model that is allowed to take actions, not just answer: open your mail, fill a form, run a workflow, write to a CRM. That permission is the whole point and the whole risk, so this comparison is organised around it.

Our criteria, stated first: what the agent is allowed to touch and how you limit it; the billing unit and the entry price on the vendor page as read on 2026-11-09; whether the vendor states where data is processed and whether it trains on your content; which models you can choose; and how much of the setup a non-developer can do alone. Two of the seven products below (Zapier and n8n) have affiliate programmes; ChatGPT and Claude have none, and Microsoft, Lindy and Relevance AI are not in our catalogue. The order does not depend on commission.

One statement up front, because every other page in this search result implies the opposite: we have not run our own benchmark of these agents. Our plan called for a twelve-task run log, and we will publish it when we have it. Until then, this page compares documented permissions, prices and licences, and cites controlled studies of AI-assisted work rather than our own impressions.

You are…	Start with	Why
A solo knowledge worker who already pays for a chat assistant	Claude Pro or ChatGPT Plus	Browser and connector agents are included in the subscription you already have; the blast radius is your own accounts
A Microsoft 365 organisation	Microsoft 365 Copilot, then Copilot Studio for custom agents	Prompts stay inside the M365 service boundary; agents inherit existing permissions
An operations person who wants agents inside existing automations	Zapier Agents	9,000+ apps, a free tier of 400 activities a month, and a hard cap per run
A team in the EU that needs to say where data lives	n8n Cloud	Hosted data stored in Frankfurt; human approval for tool calls is a built-in feature
A manager who wants an "AI employee" for inbox, calendar and calls	Lindy	Credit-based, approvals built in, SOC 2 and GDPR statements on the pricing page
An enterprise buying a fleet of agents with audit logs	Relevance AI (or n8n/Zapier Enterprise)	Sold only through sales, with SSO, RBAC and audit logs in the enterprise bundle

What the adoption numbers do and do not say

Agents are being bought faster than they are being proven. In McKinsey's 2025 state-of-AI survey, 23 percent of respondents said their organisation is scaling an agentic system somewhere and a further 39 percent said they are experimenting B^[10]. Deloitte's Q4 enterprise survey found 26 percent exploring autonomous agents "to a large extent" and 42 percent "to some extent", with regulatory compliance the top barrier for 38 percent, up from 28 percent a wave earlier A^[11]. Microsoft's 2025 Work Trend Index reports that 82 percent of the 9,037 leaders surveyed are confident they will use "digital labour" to expand capacity within 12 to 18 months A^[18]. The Stanford AI Index 2026 puts organisational AI adoption at 88 percent in 2025 and adds, in the same sentence, that "AI agent use remains early" A^[13].

Against that, Gartner's press release predicts that more than 40 percent of agentic-AI projects will be cancelled by the end of 2027 for cost, unclear value or weak risk controls; it is an analyst estimate from a headline and we grade it C C^[12]. In our Future Demand Radar the "AI agents and delegated work" category scores 59.7 out of 100 in the baseline scenario, with a p10–p90 band of 55.2 to 64.1, the highest of twelve categories and classed "build now". Our baseline scenario expects demand for the category to keep growing through 2028 with a trough risk if the cancellations Gartner predicts materialise. That is a model output with stated uncertainty, not a fact.

Search demand matches the picture: "best ai agents" is searched about 2,400 times a month in the US with a three-year compound growth rate of about 240 percent, and "ai agents platforms" another 2,400 (DataForSEO, retrieved 2026-11-09). People are shopping. The evidence on whether the shopping pays off is thinner, and the next section is about that.

The evidence on delegating work to AI

Three results matter here. First, the "jagged frontier": in a field experiment with 758 BCG consultants, GPT-4 users completed 12.2 percent more tasks, 25.1 percent faster and at more than 40 percent higher rated quality on tasks inside the model's capability, but were 19 percentage points less likely to be correct on a task designed to sit outside it B^[14]. An agent extends that frontier into actions, and the edge is no easier to see.

Second, a meta-analysis of 106 experiments found that human–AI combinations on average performed worse than the better of the human or the AI alone (g ≈ −0.23), with gains concentrated in content-creation tasks and losses in decision tasks A^[15]. Delegating a decision to an agent and then rubber-stamping it is the configuration that loses.

Third, the one randomised trial of agentic coding tools on real work found that experienced open-source developers were 19 percent slower with early-2025 AI tools, while believing afterwards they had been 20 percent faster B^[16]. Sixteen developers, 246 tasks, mature codebases; narrow, but the gap between felt and measured speed is the finding to carry into every agent trial you run. Measure, do not ask.

The human-factors literature adds the mechanism: people over-trust automated aids, and complacency and automation bias fall with accountability and training rather than with better interfaces alone A^[17]. Our playbook for managing AI agents turns that into a delegation routine; the comparison below tells you which product gives you the controls to run it.

The seven agents, by what they can touch

Claude (Projects, Claude Code, Chrome). The pricing page lists Free at USD 0 without Claude Code; Pro at USD 17 a month on annual billing (USD 200 up front) or USD 20 monthly, with Claude Code included; Max at USD 100 for "5x Pro usage"; Ultra at USD 200; Team at USD 20 per seat annual or USD 25 monthly, Premium seat USD 100 or USD 125; Enterprise at USD 20 per seat annual with a 20-seat minimum plus usage at API rates ^[1]. Connectors are on all plans; "Claude in Chrome" and the Microsoft 365 and Outlook integrations start at Pro. Model training is "opt-out" on individual plans and "none by default" on Team and Enterprise ^[1]. What it can touch: whatever connectors and browser tabs you authorise, plus your file system if you run Claude Code locally. EU data residency: not stated on the pricing page; US-only inference is offered as an API option at 1.1x price.

ChatGPT (agent mode, Codex). The pricing page rendered its plan names (Free, Go, Plus, Pro, Business, Enterprise) and its feature table, but the price amounts did not render in the text we could read, and the alternative URL refused our request, so this page carries no ChatGPT prices: see pricing ^[2]. The feature table marks Codex "Limited" on Free and Go, "Yes" on Plus and "Expanded" on Pro, and paid plans are "priced per user per month" ^[2]. Agent mode is not named on the page as such.

Microsoft 365 Copilot and Copilot Studio. Microsoft 365 Copilot is USD 30 per user a month paid yearly ^[3]; the Copilot Business add-on for small businesses is listed at USD 21 yearly with a promotional USD 18 between 1 July and 31 December 2026, or USD 25.20 monthly ^[4]. Custom agents are built in Copilot Studio, sold in packs of 25,000 Copilot Credits at USD 200 per pack a month, with pay-as-you-go for consumed credits ^[3]. The data statement is the clearest of the seven: "prompts and responses stay within the Microsoft 365 service boundary", are not used to train the models, and the page names GDPR, ISO/IEC 27018 and the Microsoft DPA ^[4]. What it can touch: everything the user already has permission to see in Microsoft 365, which is the strength and the thing to audit first.

Zapier Agents. Agents are billed in "activities, not tasks": a trigger, a knowledge-base answer, an action, a web browse or scrape, a web search or a Chrome-extension message each count as one. Free includes 400 activities a month with a cap of 10 per run; the paid Agents tier starts at about USD 33.33 a month on annual billing with 1,500 activities and 40 per run, and the agent asks before continuing when it hits the cap ^[5][6]. Core Zapier plans run from Free (100 tasks) through Professional at USD 19.99 a month annual for 750 tasks to Team at USD 69 ^[5]. Reach: 9,000+ apps. Knowledge: "FAQs, docs and public links". EU residency and SOC 2: not stated on either page we read.

n8n AI agents. The AI Agent node sits inside a workflow and is billed per execution: Starter at EUR 20 a month annual includes 2,500 executions and 1,600 Assistant credits; Pro at EUR 50 includes 10,000 executions and up to 9,600 credits; Business at EUR 667 (40,000 executions) is self-hosted only ^[7]. "Human approval for tool calls" is listed on Starter, Pro and Enterprise, and hosted data "is stored within the EU, on servers located in Frankfurt, Germany" ^[7]. An agent that calls a model fifteen times in one run is still one execution; the model provider bills the calls. Self-hosting is under the Sustainable Use License, which our open-source agents comparison explains.

Lindy. Credit-based "AI employees": Free gives USD 50 of credits that last seven days; Team is USD 29.99 a month per active seat for 3,000 pooled credits, with higher tiers up to 1.75 million credits and top-ups at USD 10 per 1,000 ^[8]. The page sizes work as "Everyday Asks 2–250 credits", "Deep work 250–1,000" and "Big Builds 1,000–2,500", which makes the monthly bill depend on how often you ask for deep work. 1,500+ integrations plus MCP, two to five connected inboxes depending on plan, meeting scheduling, phone calls, and "anything with outside impact waits for approval". The FAQ states "Lindy is SOC 2 and GDPR compliant"; EU data residency is not stated; HIPAA with a BAA is Enterprise only ^[8].

Relevance AI. On 2026-11-09 the pricing page showed a single Enterprise card: custom pricing, "Talk to sales", "Custom Vendor Credits", unlimited agents and tools, 2,000+ integrations, agent evaluations, SSO, RBAC and audit logs ^[9]. No self-serve plan, no free tier, no public price, and the privacy FAQ did not render. Treat it as an enterprise purchase.

Security and permissions: what to check before you connect anything

• Scope. Give the agent its own account where you can (a service mailbox, a restricted CRM role) rather than your own credentials. Microsoft 365 Copilot and Zapier Agents inherit the connecting user's permissions, which means an over-permissioned user produces an over-permissioned agent.
• Approval gates. n8n and Lindy document approval before tool calls or "outside impact" actions ^[7][8]; Zapier caps activities per run and asks before continuing ^[6]. Turn these on for anything that sends, pays or deletes.
• Data residency. Only n8n (Frankfurt) states a region on its pricing page ^[7]. Microsoft states a service boundary and names GDPR and ISO/IEC 27018 ^[4]. Lindy states SOC 2 and GDPR compliance without a region ^[8]. Claude, ChatGPT, Zapier and Relevance AI did not state residency on the pages we read.
• Training on your data. Claude: opt-out on individual plans, none by default on Team and Enterprise ^[1]. Microsoft: not used to train ^[4]. The others did not state it on the pages we read.
• Logs. You cannot manage what you cannot replay. Zapier, n8n and Lindy show step logs; Relevance AI lists audit logs on Enterprise. Keep them for the agent's first month at minimum.

Who should not buy an agent yet

• Anyone whose task is a judgement call with no ground truth. The meta-analysis above found losses precisely there A^[15].
• Teams without a person who will own the agent. A Gartner-style cancellation is a management outcome, not a model failure C^[12]; our analysis of why agent projects fail goes through the common causes.
• Regulated data without a residency statement from the vendor. Wait for the contract, not the marketing page.
• People who expect to feel faster. The developer trial found the feeling and the stopwatch disagreed B^[16].

If what you want is a better chat assistant rather than an actor in your systems, our explainer on what an AI agent is draws the line. If you want to build an agent rather than subscribe to one, the no-code agent builders comparison covers Zapier, Make, n8n, Lindy, Dify and Voiceflow with cost at three run volumes.

What to do next

Pick the one task you would delegate first, write down how you will measure it (time, error rate, rework), then choose the narrowest agent that can do it. For a solo worker that is usually the assistant you already pay for. For a team inside Microsoft 365 it is Copilot. For automations, Zapier Agents below about 1,500 activities a month and n8n when you need per-run billing, approval gates and a stated EU region. The Tool Finder turns your task list, team size and residency requirement into a shortlist with the same dated prices. We will add task-run results to this page when we have run them; until then, measure your own.

Frequently asked questions

What is the best AI agent for work in 2026?

It depends on what you let it touch. For a solo worker, the agent features inside Claude Pro (USD 17–20 a month) or ChatGPT are the cheapest start. For Microsoft 365 organisations, Microsoft 365 Copilot at USD 30 per user a month keeps prompts inside the service boundary. For automations, Zapier Agents (400 free activities a month) or n8n (EUR 20 a month, Frankfurt hosting). We have not run our own benchmark; prices checked 2026-11-09.

Are there free AI agents?

Zapier Agents includes 400 activities a month on its Free plan, Lindy gives USD 50 of credits that expire after seven days, Claude's Free plan includes connectors but not Claude Code, and n8n's self-hosted Community Edition costs only the server. Microsoft Copilot and Relevance AI have no free agent tier on the pages we read on 2026-11-09.

Do AI agents actually make people more productive?

The controlled evidence is mixed. Consultants using GPT-4 were faster and better inside the model's capability but less accurate outside it (Dell'Acqua et al. 2023); a meta-analysis of 106 experiments found human–AI pairs did worse than the better of the two alone on decision tasks (Vaccaro et al. 2024); and experienced developers were 19 percent slower with early-2025 agentic tools while believing they were faster (METR 2025). Measure your own task rather than trusting the feeling.

Which AI agent is best for GDPR and EU data residency?

n8n is the only product of the seven whose pricing page states a region: hosted data is stored in the EU, in Frankfurt. Microsoft states that prompts and responses stay within the Microsoft 365 service boundary and names GDPR and ISO/IEC 27018. Lindy states SOC 2 and GDPR compliance without naming a region. Claude, ChatGPT, Zapier and Relevance AI did not state residency on the pages we read on 2026-11-09.

What is the difference between an AI agent and an AI assistant?

An assistant answers; an agent acts. The agent can open your mail, run a workflow or write to a database, which is why permissions, approval gates and logs matter more than the model. Our explainer on AI agents draws the line in detail.

How much does Microsoft Copilot Studio cost for custom agents?

Copilot Studio is sold in packs of 25,000 Copilot Credits at USD 200 per pack a month, with pay-as-you-go billing for credits consumed beyond that. Microsoft 365 Copilot itself is USD 30 per user a month paid yearly. Both prices were read on 2026-11-09.

Sources and evidence

1. Claude pricing (Free, Pro, Max, Ultra, Team, Enterprise; Claude Code inclusion; training policy) — Anthropic (2026), vendor page A claude.com/pricing (retrieved 2026-11-09)
2. ChatGPT pricing (plan names and feature table; price amounts did not render) — OpenAI (2026), vendor page B chatgpt.com/pricing/ (retrieved 2026-11-09)
3. Microsoft Copilot Studio pricing (Copilot Credits packs; Microsoft 365 Copilot per-user price) — Microsoft (2026), vendor page A www.microsoft.com/en-us/copilot/pricing/copilot-studio (retrieved 2026-11-09)
4. Microsoft 365 Copilot pricing (Copilot Business add-on, bundles, data statements) — Microsoft (2026), vendor page A www.microsoft.com/en-us/microsoft-365-copilot/pricing (retrieved 2026-11-09)
5. Zapier pricing (task tiers; Agents activities and Agents Pro price) — Zapier (2026), vendor page A zapier.com/pricing (retrieved 2026-11-09)
6. Zapier Agents (activity definition, per-run caps, knowledge sources) — Zapier (2026), vendor page A zapier.com/agents (retrieved 2026-11-09)
7. n8n pricing (executions, AI Agent node, human approval for tool calls, Frankfurt hosting) — n8n GmbH (2026), vendor page A n8n.io/pricing/ (retrieved 2026-11-09)
8. Lindy pricing (credits, Team plan, approvals, SOC 2 and GDPR statement) — Lindy (2026), vendor page A www.lindy.ai/pricing (retrieved 2026-11-09)
9. Relevance AI pricing (Enterprise card only) — Relevance AI (2026), vendor page B relevanceai.com/pricing (retrieved 2026-11-09)
10. The state of AI in 2025: Agents, innovation, and transformation (survey figures as reported) — McKinsey & Company, via ppc.land (2025), report B ppc.land/most-companies-still-pilot-ai-programs-despite-widespread-ado (retrieved 2026-11-09)
11. The State of Generative AI in the Enterprise: Generating a new future (Q4 report) — Deloitte (2025), report A www.deloitte.com/us/en/about/press-room/state-of-generative-ai.html (retrieved 2026-11-09)
12. Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 — Gartner Newsroom (2025), report C www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts (retrieved 2026-11-09)
13. Artificial Intelligence Index Report 2026, Chapter 4: Economy — Stanford HAI (2026), report A hai.stanford.edu/assets/files/ai_index_report_2026_chapter_4_economy.p (retrieved 2026-11-09)
14. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell'Acqua F, McFowland E, Mollick ER, et al. (2023), RCT B doi.org/10.2139/ssrn.4573321 (retrieved 2026-11-09)
15. When combinations of humans and AI are useful: A systematic review and meta-analysis — Vaccaro M, Almaatouq A, Malone T (2024), meta-analysis A doi.org/10.1038/s41562-024-02024-1 (retrieved 2026-11-09)
16. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — Becker J, Rush N, Barnes E, Rein D (METR) (2025), RCT B doi.org/10.48550/arXiv.2507.09089 (retrieved 2026-11-09)
17. Complacency and Bias in Human Use of Automation: An Attentional Integration — Parasuraman R, Manzey DH (2010), systematic review A doi.org/10.1177/0018720810376055 (retrieved 2026-11-09)
18. 2025 Work Trend Index Annual Report: The year the Frontier Firm is born — Microsoft WorkLab (2025), report B www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-fro (retrieved 2026-11-09)