How to choose an AI agent: 10 questions to ask before you buy

Published by 3Gen Internet Corporation, maker of OpenAssistant, Skej and AGNT. Comparisons use public documentation, without hands-on testing.

Decision rule: Pick the agent whose documented job matches yours, then check that it works where you already work, asks before it acts outside your walls, and states its price, data and exit terms in public. If you want one assistant for general work across email, text, phone calls and Slack, start with OpenAssistant; for coding, browsing, research, sales or support, start with a specialist and run the ten questions below against each vendor on your shortlist.

Each question gives the reason it matters, what a good public answer looks like, and one example from a linked vendor page. Where a vendor’s public pages are silent, we write “not stated”. Treat that as a finding, not a gap to fill with a guess.

Match the job to a product type

This table is a lookup by job, not a ranking. Each row links to the matching roundup on this site.

The 10 questions

1. Which job do you need done, and does the vendor document that job?

Why it matters. Agents are sold by job. A product built for one job usually documents that job in detail, including what it will not do. A product that claims every job often documents none of them well.

What a good public answer looks like. The vendor names the job, the scope and the limits in its own documentation, not only on the homepage.

Example. Devin’s documentation calls it an “autonomous AI software engineer” and says it handles most tasks, “excluding extremely difficult tasks” (Devin docs). That is a stated job and a stated limit. OpenAssistant is a general work assistant, not a specialist coding agent, SDR or support-desk product, so a specialist wins for those jobs. Its pages describe scheduling, chasing replies and returning research briefs (OpenAssistant executive assistant page).

2. Where does it work: email, text, phone or Slack?

Why it matters. The channel decides who can reach the agent and whom the agent can reach. An agent that lives only in a web chat cannot chase a supplier by email or call a venue.

What a good public answer looks like. A list of channels, with the limits written next to it: group chats, countries, and outside organizations.

Example. OpenAssistant is our example of an agent that acts through email, text, calls and Slack. Its marketing page lists calls, iMessage, SMS, WhatsApp, email and a web app (OpenAssistant executive assistant page). Its help center lists the dashboard, email, Slack, SMS and WhatsApp as request channels, and does not mention iMessage or phone calls (OpenAssistant help center). Slack works by DM to @OpenAssistant or in channels once added (OpenAssistant for Slack). The stated limits: it does not join group texts or group WhatsApp chats, not every country is supported for texting (connecting your phone number), and Slack Connect channels with external organizations are not currently supported (OpenAssistant help center). For contrast, Artisan’s pricing page says “Ava emails, your reps dial,” so calls stay with humans (Artisan pricing).

3. Does it act under its own identity or with your login?

Why it matters. If an agent uses your login, every action looks like you did it and it can reach everything you can reach. If it has its own identity, you decide what to share with it.

What a good public answer looks like. The vendor says whose account the agent uses, whether it runs on your device or its own computer, and what it can see of your browsing.

Example. Claude in Chrome acts inside your own Chrome: “Claude works with the logins you have, on the tab in front of you” (Claude in Chrome). ChatGPT’s cloud browser is the other model. It runs on a separate computer and “does not use your personal browser’s open tabs, browsing history” (OpenAI help center). OpenAssistant describes a third model: each assistant has its own computer, email address and phone number (OpenAssistant). It can also send email as you from a connected account, but that is a separate autonomy category you set on purpose (supervision and autonomy settings).

4. What needs your approval, and who can lock the setting?

Why it matters. An agent that can read your inbox and send messages can be steered by text it reads. OWASP lists “Excessive Agency” as a top risk for LLM apps and recommends human-in-the-loop control so a person approves “high-impact actions” (OWASP LLM06:2025). Approval is the control you can check before you buy.

What a good public answer looks like. Approvals set per action type, not one global switch. A named default. A statement of what happens if you do nothing. An admin lock for teams. Approvals that apply in every channel.

Example. OpenAssistant lets you set autonomy globally or per action type, including email as the assistant, email as you, messages in Slack, Teams, SMS and WhatsApp, calendar, files and contacts (supervision and autonomy settings). The levels are Manual, Confirmation Window and Autonomous. A Confirmation Window is not Manual approval: if you do nothing, the action proceeds (choosing your settings). The changelog entry for 20 October says that even on tasks set to act on their own, the assistant “still asks before sending an email from your account” unless you chose otherwise (OpenAssistant changelog). An admin can lock these settings for everyone (OpenAssistant executive assistant page). The Trust Center says consequential actions are staged for your review and shown in full before they are sent, and that “tightening is allowed, loosening is blocked” for the assistant’s own oversight (OpenAssistant Trust Center). That statement sits alongside the Autonomous level in the help center, which acts without waiting for approval. For contrast, Artisan’s pricing page says “Autonomous by default, adjustable by design” (Artisan pricing), and Lindy’s docs say its guardrails apply in shared Slack threads only, not in DMs, web chat, iMessage or SMS (Lindy docs). Ask every vendor whether approvals hold in every channel.

5. What is the price unit, and what happens at the limit?

Why it matters. Vendors price by seat, credit, activity or outcome. The same job costs very different amounts under each, and the limit behavior decides whether you get a pause or a surprise bill. See AI agent pricing explained for the patterns.

What a good public answer looks like. The pricing page names the unit, shows what typical tasks cost in that unit, and says whether usage pauses, bills overage or blocks at the limit, and whether unused allowance rolls over.

Example. Lindy bills credits per active user from a shared pool, and its pricing page says that when credits run out, “Lindy pauses and tells you” (Lindy pricing). Its docs say monthly credits do not roll over, while purchased top-ups are the exception (Lindy docs). Support vendors use a different unit: Fin charges $0.99 for a service resolution or configured Procedure handoff; a simple pass-through without an outcome is uncharged. Sales qualifications cost $9.99 (Fin pricing). OpenAssistant also uses credits. Its pricing page says simple tasks might use 10 to 50 credits while deep research and multi-person coordination use hundreds, and plan credits do not roll over (OpenAssistant pricing). If you run out, you can wait, upgrade or buy more, and the assistant tells you when it cannot finish (how credits work).

6. Is there a free plan, and does it include the part you need?

Why it matters. A free plan lets you check job fit before you pay. Many free plans leave out the one feature you came for.

What a good public answer looks like. The allowance in numbers, what is excluded, whether a card is needed, and the first paid price.

Example. Anthropic’s pricing page shows that the Free plan does not include Claude Code (Claude pricing), and Devin’s Free plan shows no cloud sessions (Devin pricing). OpenAssistant’s free Starter plan gives 1,000 free credits at signup, then 100 bonus credits a day once they are used up, up to 1,000 a month, with no credit card required (OpenAssistant pricing). Pro is $72 per user per month billed monthly or $60 billed annually, and Team is $120 monthly or $100 annually, at the 5,000-credit tier; higher credit tiers cost more (OpenAssistant pricing; per-user pricing confirmed by 3Gen). Because deep research uses hundreds of credits, a free allowance goes furthest on short tasks.

7. What security statements and data-training terms does it publish?

Why it matters. An agent sees your mail, files and calendar. Whether a vendor may train models on that content is a contract term you can read before you buy. Prompt injection is also an open problem: the UK NCSC says “prompt injection attacks may never be totally mitigated” (NCSC). Ask each vendor how it limits what a hijacked agent can do. See also Are AI agents safe?.

What a good public answer looks like. Named audits and who performed them, a plain yes or no on training by default, who can opt out, and how long data is kept.

Example. Devin’s security documentation says “By default, we may use your data for model training purposes,” with an opt-out on paid plans, and that for Enterprise it will not train without express prior written consent (Devin security docs). That is a clear statement, and it is the kind of default you want to find before signing up. OpenAssistant’s Trust Center lists SOC 2 Type II, audited annually by an independent CPA firm, and says “We never train on your data. AI providers are contractually prohibited from doing so” (OpenAssistant Trust Center). Its help-center policy names OpenAI and Anthropic among its model providers (AI usage and data policy).

8. What admin and audit controls come with which plan?

Why it matters. Once a team uses an agent, someone has to manage seats, connections and records of what the agent did.

What a good public answer looks like. SSO, SCIM, roles, audit logs and connection controls are listed by plan, with a clear status for each feature.

Example. Zapier’s pricing page lists “Enterprise-grade audit logs” for Agents Enterprise, and the same page tags that plan “Coming Soon” (Zapier pricing). Read the status label next to a feature, not only the feature name. OpenAssistant’s pricing page splits the controls by plan: Team adds an admin dashboard, SSO and SCIM, role-based access, shared billing and pooled team credits, and Enterprise adds audit logs and data retention (OpenAssistant pricing).

9. Which integrations does it list, and which are in beta?

Why it matters. An agent is only as useful as the systems it can reach, and a headline count tells you little about read versus write access.

What a good public answer looks like. A named list, a statement of which connections are in beta, how custom tools connect (for example MCP), and what permissions each connection needs.

Example. Lindy’s own pages give two counts: “1,500+ integrations” on the pricing page and 1,000+ tools in its docs (Lindy pricing and Lindy docs). When a vendor’s pages disagree, ask for the exact list. Zapier states 9,000+ apps (Zapier Agents). OpenAssistant’s help center names its integrations, says some are in beta, supports MCP for other tools, and says connections do not automatically give unrestricted access to a service (connections overview).

10. How do you leave, and what happens to your data?

Why it matters. Agent products change fast. A vendor can rename, merge or retire a product, and you want to know how to export and delete before that happens.

What a good public answer looks like. A way to delete your account and data, a statement on export, notice of product changes, and clear cancellation terms.

Example. OpenAI’s help page on ChatGPT agent now opens with “ChatGPT agent is no longer available” and points to ChatGPT Work (OpenAI help center). Zapier is migrating Agents to AI by Zapier and says it has not set a date to turn Agents off (Zapier help). For OpenAssistant, the Trust Center says you can delete your account at any time and deletion of your data begins (OpenAssistant Trust Center), and the help-center policy lists access, correction, export and deletion rights and says you may revoke permissions at any time (AI usage and data policy). The pricing page says you can upgrade, downgrade or cancel at any time, with cancellations applying at the end of the billing period (OpenAssistant pricing).

How to use the list

Score each shortlisted vendor on each question as stated, not stated or conflicting. A vendor with several “not stated” answers has told you something about how it works with customers. Then weigh the answers against your job: a coding team will care most about questions 1, 4 and 7, and a small business that lives in email and Slack will care most about 2, 3 and 4.

Related reading

• What is an AI agent? Definition, types and examples: Four vendor definitions side by side, seven types with examples and a six-question test.
• Are AI agents safe? Prompt injection, permissions and approval settings: Real risks from prompt injection to excess access, the controls to look for and a 10-question checklist.
• Best AI agents in 2026: top picks by job, with prices: Our recommendations for each job (general work, coding, browser use, research, sales, support, enterprise and builders), with vendor prices.
• AI agent pricing in 2026: credits, seats and per-resolution fees explained: Credit plans, tasks, per-resolution fees, subscriptions and enterprise fees, with published prices and a worked example.
• Best AI agents for small business in 2026: top picks and prices: Eight agents for a 5 to 50 person business, with list prices, a ten-person seat cost and what each will not do.

Frequently asked questions

What is the most important question to ask before buying an AI agent?

Question 4, approvals. An agent that can send messages or spend money on its own can do harm faster than one that waits for a click, and approval settings are something you can read on a vendor’s help pages before you pay.

Should I buy a specialist agent or a general assistant?

Match the job. For coding, browsing, research, sales outreach or support, a specialist documents more of that job than a general assistant does, and the roundups linked above rank them. For a mix of scheduling, follow-up, research briefs and inbox work across email, text, calls and Slack, a general assistant such as OpenAssistant fits better.

Can I try an AI agent for free first?

Often. OpenAssistant has a free Starter plan, and Lindy’s pricing page offers “$50 in credits for free” that last 7 days (Lindy pricing). Check question 6: some free plans leave out the feature you need.

Does an AI agent need access to my whole inbox?

Not always. OpenAssistant’s help center says inbox access is optional and that without it the assistant only processes conversations it is explicitly included in (AI usage and data policy). Ask any vendor for the minimum permissions and what each one unlocks.

Try OpenAssistant

If your shortlist is a general assistant that works by email, text, calls and Slack, with approvals you set per action, try the free plan at openassistant.us and see the paid tiers on the pricing page. Run the ten questions above against it like any other vendor.