Frequently Asked Questions

An AI agent is software that uses artificial intelligence to autonomously complete tasks, make decisions, and interact with users or systems without requiring step-by-step human instruction. Unlike basic AI tools that respond to single queries, AI agents can plan multi-step workflows, use external tools (web search, code execution, API calls), and take actions inside connected software systems on your behalf. Examples include coding assistants that can write, test, and commit code; customer service agents that can resolve tickets without human involvement; and research agents that can autonomously gather and synthesise information from multiple sources.

Chatbots typically respond to single-turn questions using pre-scripted flows or basic language models — they answer what you ask and stop. AI agents go significantly further. They can maintain context across long, complex conversations; autonomously plan and execute multi-step tasks without per-step prompting; use external tools like search engines, code interpreters, and databases; and take real actions inside connected systems such as updating a CRM record, creating a Jira ticket, or deploying code. The key distinction is autonomy — an agent acts; a chatbot responds.

AI agents are broadly categorised by their primary function. The main categories tracked on AI Agent Square include: Coding agents (GitHub Copilot, Cursor, Devin), Customer service agents (Intercom Fin, Zendesk AI), Writing agents (Jasper, Copy.ai), Sales agents (Gong, Salesloft), Research agents (Perplexity, Elicit), Data & analytics agents, Marketing agents, Voice & audio agents (ElevenLabs), and Video agents (Synthesia). There are also general-purpose AI assistants (ChatGPT Enterprise, Claude for Enterprise) that span multiple use cases. See our full category list for all 20 categories.

Most commercial AI agents are built on top of frontier LLMs from OpenAI (GPT-4o, GPT-4o mini), Anthropic (Claude 3.5 Sonnet, Claude 3 Opus), Google (Gemini 1.5 Pro, Gemini Flash), or Meta (Llama 3). Many agents allow you to select your preferred model — Cursor, for example, lets users switch between GPT-4o, Claude 3.5 Sonnet, and other models. Some vendors, like Google (Gemini) and Anthropic (Claude), operate their own proprietary models. For specialised tasks, fine-tuned models trained on domain-specific data often outperform general frontier models.

Start by defining your use case precisely — "we want to automate tier-1 customer support tickets" is far more useful than "we want AI." Then evaluate agents on five dimensions: (1) Task fit — does it actually solve your specific workflow? (2) Integration — does it connect to your existing tools (Salesforce, Zendesk, GitHub, Slack)? (3) Pricing at scale — model the cost at your actual expected usage volume, not the headline figure. (4) Security and compliance — does it meet your data residency, SOC 2, GDPR requirements? (5) Support quality — is there a meaningful SLA and customer success engagement? Use our comparison tool to evaluate options side by side.

Yes, always. A structured 30–90 day proof-of-concept (POC) with defined success metrics is essential before committing to an annual contract. Your POC should include: a representative sample of real production tasks; baseline measurements of current process performance (time, cost, error rate) to compare against; a representative user group (not just enthusiastic early adopters); and pre-agreed success criteria. Any serious vendor will support a structured POC. If a vendor is unwilling to offer a trial period, that should raise questions about their confidence in the product's performance in your environment.

Key questions for AI agent procurement include: What is your data processing agreement (DPA) and where is data stored? What are your uptime SLAs and what happens when you miss them? How is my data used — is it used to train your models? What does your security certification cover (SOC 2 Type II, ISO 27001, HIPAA BAA)? Who are your subprocessors? What are the most common reasons customers churn? Can you provide references from customers in our industry and at our scale? What are the contract renewal terms and price escalators? Download our full procurement checklist for 50 questions to ask every vendor.

ROI measurement for AI agents depends on the use case. For coding agents: measure developer velocity (PRs merged per week, time from issue to deployment), code review time, and bug rate. For customer service AI: track resolution rate (% tickets resolved without human involvement), average handle time, CSAT scores, and cost per resolution. For writing agents: volume of content produced per writer per week, content review cycles, time to publication. Establish baseline metrics before deployment, then measure the same metrics at 30, 60, and 90 days post-deployment.

Pricing varies enormously by category and scale. Individual and small team plans: coding agents cost $10–$25/user/month; writing agents $20–$125/month; productivity agents $10–$20/user/month; research agents $20/month. Enterprise pricing is negotiated and typically involves annual commitments with volume discounts. See our comprehensive Pricing Guide for a full breakdown by category, including comparison tables and total cost of ownership analysis.

There are six main pricing models in the AI agent market: Per-seat (fixed monthly fee per user — predictable, common in coding tools); Usage-based (per API call, token, or task — scales with use); Tiered subscription (Free/Pro/Business/Enterprise tiers with feature gates — most common SaaS model); Outcome-based (per resolved conversation or completed task — aligns incentives, common in customer service AI); Credits/prepaid (purchase credits upfront — flexible but can waste); and Enterprise/custom (negotiated annual contracts with volume discounts and SLAs). Many vendors use hybrid models combining two or more of these.

Many AI agents offer free tiers or trials. GitHub Copilot Free, Cursor Free, Copy.ai Free, ElevenLabs Free, Perplexity Free, and Notion AI (limited) all provide genuinely useful free plans. Others like Jasper and Intercom Fin do not offer free plans but provide 7–14 day trials. Enterprise-focused agents (Salesforce Einstein, Gong) typically require a demo and sales engagement before any trial access. Free tier limitations usually include usage caps, restricted model access, and limited integrations. Our agent reviews note whether a free tier is genuinely useful or essentially a preview-only experience.

Our reviews combine editorial hands-on testing with verified enterprise buyer ratings. Each agent is scored across six dimensions: Features (20%), Pricing & Value (20%), Ease of Use (20%), Support Quality (15%), Integration Ecosystem (15%), and Reliability (10%). Editorial scores are derived from structured testing using standardised tasks relevant to each category. User scores are collected from verified enterprise IT buyers, procurement professionals, and department heads — we do not include unverified consumer reviews. See our full Methodology page for details.

No. Our editorial reviews and scores are completely independent. Vendors cannot pay to improve their scores, change editorial content, or receive preferential placement in organic listings. We do offer clearly labelled sponsored listing placements (marked "Sponsored") as a revenue stream, and we earn affiliate commissions when you sign up for a tool via our links. Neither of these commercial relationships influences our editorial scoring or review content. If you observe what you believe is a conflict of interest in any review, contact us.

Use our Submit Agent form to submit your tool. We review all submissions and aim to publish qualified agents within 30–60 days of submission. Qualification criteria include: publicly available product (generally available or open beta), published pricing (or clear enterprise sales process), at least basic security documentation, and a genuine AI agent use case (not purely a chatbot or basic automation tool). Submission does not guarantee a specific score — all published reviews reflect our independent editorial assessment.

Pricing data is refreshed monthly. Full editorial reviews are updated whenever an agent releases a major version update, significant new features, or pricing changes. Each review page shows the last review date at the top. For rapidly evolving agents, reviews may be updated quarterly or more frequently. If you notice outdated information on any page, please contact our editorial team.

Enterprise-grade AI agents typically offer: SOC 2 Type II certification, GDPR and CCPA compliance, data encryption at rest (AES-256) and in transit (TLS 1.3), SSO/SAML authentication, role-based access controls, audit logging, configurable data retention, and the option to sign a Data Processing Agreement (DPA). However, security posture varies significantly between vendors. Always request a completed security questionnaire, review their latest SOC 2 report, request a list of subprocessors, and confirm your data residency requirements are met before deployment. For healthcare and financial services, additionally verify HIPAA BAA availability and relevant regional compliance.

Enterprise AI governance should cover four areas: Access control — who can use which agents, with what permissions; Data governance — what data the agent can access and process, and how outputs are classified; Output quality — human review workflows for high-stakes outputs (legal documents, financial analysis, customer-facing content); Audit and compliance — logging of agent actions and outputs for regulatory and internal review. Larger organisations should appoint an AI programme lead and develop an acceptable use policy before broad deployment. Our enterprise buyer's guides cover governance frameworks in depth.

The most important integrations depend on your use case, but the most commonly required in enterprise evaluations are: Identity — SSO via Okta, Azure AD, or Google Workspace; Collaboration — Slack and Microsoft Teams for notifications and approvals; CRM — Salesforce and HubSpot for sales/service agents; Ticketing — Zendesk, ServiceNow, Jira for workflow agents; Code repositories — GitHub and GitLab for coding agents; Data — Snowflake, BigQuery, or your enterprise data warehouse for analytics agents. Always test integration depth, not just existence — a "native integration" might mean a simple webhook rather than a full bidirectional API.