On this page
Quick answer: what is an AI agent?
An AI agent is software that uses a large language model to pursue a goal on its own. It reads your data, decides the next step, uses tools such as code editors, calendars, payment systems or databases, and takes actions in other systems with limited supervision. A chatbot answers a question. An agent does the work.
Key takeaways
- The agents people actually pay for share six traits: they use the customer's own data, live inside the workflow, take actions, vouch for the result, show up at the trigger, and ship with a screen built for one job.
- Agents now exist for most functions: coding, support, voice, sales, legal, healthcare, finance, recruiting, security, meetings and research.
- Pricing is moving from seats to outcomes. Some agents charge per resolved conversation instead of per user.
- The best way to evaluate any agent is to ask one question: what exact work does it finish without a human copying and pasting?
Start with the job you need done
- Write and ship codeCursor, Claude Code, Codex, Devin
- Resolve support ticketsFin, Sierra, Decagon, Parloa
- Answer phone callsRetell, Vapi, ElevenLabs, Bland
- Book sales meetings11x, Artisan, Clay, Apollo
- Do legal, health or finance workHarvey, Abridge, Rogo, Hebbia
- Capture meetings and researchGranola, Perplexity, Manus
- Automate workflowsZapier Agents, n8n, Lindy
- Serve Indian languagesGnani.ai, Sarvam AI, Yellow.ai
- Pick one for your teamA six-step checklist
What is the difference between an AI agent, a chatbot and a copilot?
A chatbot responds, a copilot suggests, and an agent acts. The difference is who finishes the task. With a chatbot or copilot, a human still carries the result into another system. With an agent, the software completes the work itself, inside your systems, and reports back.
| Type | What it does | Who finishes the task | Example |
|---|---|---|---|
| Chatbot | Answers questions in a chat window | The human | A website FAQ bot |
| Copilot | Suggests text or code while you work | The human | Autocomplete in an editor |
| Workflow automation | Runs fixed if-this-then-that steps | The rules you wrote | A Zapier-style zap |
| AI agent | Plans, uses tools, takes actions, checks results | The software, with human oversight | Fin resolving a refund end to end |
Two ideas make an agent different from earlier automation. First, it decides the next step itself instead of following a fixed script. Second, it can change things in other systems: issue a refund, book a slot, update a record, open a pull request.
What makes an AI agent worth paying for? Six traits
The agents that win customers share six traits. Each one removes a step where a human used to copy, paste, check or remember. If an agent has none of them, it is usually a chatbot with a new name.
Uses the user's own data
Works from the customer's files, not a pasted excerpt.
- Example
- Harvey works on a law firm's own documents
- Why it matters
- Answers are specific to the firm, not generic
Lives inside the work
Sits in the tool where the work happens.
- Example
- Cursor in the code editor, Granola in the meeting
- Why it matters
- No copy and paste between apps
Takes actions
Changes things in other systems.
- Example
- Fin processes refunds and payments
- Why it matters
- The task is finished, not just advised
Vouches for the result
A human signs off, or you pay only for the outcome.
- Example
- Fin charges per resolution; Sierra sells outcomes
- Why it matters
- Risk moves from the buyer to the vendor
Turns up at the trigger
Acts when an event happens, not only when asked.
- Example
- Poke messages you first on apps you already use
- Why it matters
- Work starts without a prompt
A screen built for one job
An interface designed for one workflow, not a generic chat box.
- Example
- a16z argues a chat box makes first-class product hard
- Why it matters
- Better experience than a general chat tool
ChatGPT is the useful contrast. It is a general tool that waits to be asked, so it is strong at breadth and weaker at the six traits above. The a16z point, from its State of Consumer AI 2025 report, is that delivering a first-class experience inside a general chat interface is hard.
“…hard to deliver a first-class product experience inside the ChatGPT interface.”
Which AI agents were named in the source text, and what does each one actually do?
Six named agents illustrate the traits: Harvey, Cursor, Granola, Fin, Sierra and Poke. Here is who built each, why, and the exact work it does. Figures are as reported in the sources linked, as of October 2026.
| Agent | What it does | Trait it illustrates |
|---|---|---|
| Harvey | Works on a law firm's own documents | Uses the user's own data |
| Cursor | AI coding agent inside the code editor | Lives inside the work |
| Granola | Meeting notes agent that sits in the meeting | Lives inside the work |
| Fin (Intercom) | Support agent that processes refunds and payments, priced per outcome | Takes actions, vouches for the result |
| Sierra | Customer service agents sold on "a job well done" | Vouches for the result |
| Poke | Messages you first on apps you already use | Turns up at the trigger |
| ChatGPT | A general chat tool that waits to be asked | Counter-example |
a16z is cited in the source text as a research source, not as an agent.
Harvey
The legal agent that works on a firm's own documents
Visit Harvey- Built by
- Harvey, founded in 2022 in San Francisco by Winston Weinberg (a former litigator) and Gabriel Pereyra (a former DeepMind researcher), per this company profile.
- Why it was built
- Legal work is document-heavy, high-value and slow. Harvey's thesis is that agents can run legal workflows from start to finish, not just answer questions.
- What work it does
- Executes tasks such as mergers and acquisitions work, due diligence, contract drafting and document review, according to Pulse 2.0's coverage of the March 2026 raise.
- Typical use cases
- A deal team reviewing hundreds of contracts for change-of-control clauses during due diligence; an associate drafting a first-pass agreement from a firm's own precedents; an in-house team running compliance checks across policies; litigators researching case law and summarizing long document sets.
- Scale
- Harvey said in its March 2026 announcement that it raised $200 million at an $11 billion valuation, with more than 100,000 lawyers across 1,300 organizations and over 25,000 custom agents running on the platform. A later report said it was in talks at $15.5 billion; that was not confirmed at the time of writing.
- Uses the user's own data
- Vouches for the result because a lawyer signs off
“…becoming the system through which legal work gets done.”
Cursor
The coding agent inside the editor
Visit Cursor- Built by
- Anysphere, a company founded by four MIT dropouts, per an a16z-based report.
- Why it was built
- To make AI the organizing principle of the code editor instead of an add-on. Cursor is a fork of Microsoft's open-source VS Code.
- What work it does
- Takes plain-English instructions and writes, edits, debugs and refactors code across many files. Its agent mode runs multi-step commands, and background agents work on tasks and open pull requests on their own.
- Typical use cases
- Migrating a codebase from one framework to another across hundreds of files; fixing a bug by describing it in words; generating tests; running a background agent overnight on a ticket; reviewing pull requests with Bugbot, its GitHub code-review tool launched in July 2025.
- Scale
- Reported annualized revenue of about $4 billion by mid-2026, per Amplifying's tracker. SpaceX agreed on June 16, 2026 to acquire Anysphere in a reported $60 billion all-stock deal; sources differ on whether the deal had closed by October.
- Lives inside the work
- Takes actions in the codebase
Granola
The meeting agent that sits in the meeting
Visit Granola- Built by
- Granola, a London company founded in 2023 by Chris Pedregal and Sam Stephenson.
- Why it was built
- People in back-to-back meetings need accurate notes without a visible bot joining the call or typing everything themselves.
- What work it does
- Records meeting audio on the user's computer, transcribes it, and merges the transcript with the user's own typed notes into structured notes. Its Spaces feature lets teams share notes with access controls, and its API and integrations feed meeting context to tools such as Claude and ChatGPT, per The Next Web.
- Typical use cases
- Capturing customer discovery calls; turning a one-on-one into action items; preparing a brief by asking questions across every past meeting with one account; sharing a product review summary with a team.
- Scale
- Raised $125 million at a $1.5 billion valuation in March 2026, led by Index Ventures with Kleiner Perkins participating.
- Lives inside the work
Fin
The customer support agent that takes actions and charges per outcome
Visit Fin- Built by
- The company formerly known as Intercom, founded in Dublin in 2011. It launched Fin in 2023 and renamed the whole company Fin in May 2026.
- Why it was built
- To resolve customer conversations end to end rather than deflect them to a human queue.
- What work it does
- Answers and resolves queries across live chat, email, WhatsApp, SMS, phone and Slack, powered by a proprietary model called Apex, per MarTech. It can process payments and refunds.
- Typical use cases
- Checking an order and issuing a refund; changing a subscription; answering account questions using the help center; qualifying inbound leads (billed at a higher rate); working on top of Zendesk or Salesforce instead of only Intercom's inbox.
- Pricing
- $0.99 per resolved conversation, charged once per conversation, per published pricing summaries. Human seats are billed separately on the full helpdesk.
- Scale
- Salesforce signed a definitive agreement on June 15, 2026 to acquire Fin for about $3.6 billion, with closing expected in Salesforce's fiscal Q4 2027, per Business Wire. Reports put Fin's agent revenue near $100 million of roughly $400 million in total ARR.
- Takes actions
- Vouches for the result through outcome pricing
“…enable every company to become an agentic enterprise.”
Sierra
Customer experience agents sold as “a job well done”
Visit Sierra- Built by
- Sierra, founded in 2023 by Bret Taylor (former Salesforce co-CEO) and Clay Bavor (former Google Labs lead).
- Why it was built
- To let large consumer brands deploy their own branded agents that handle service, sales and retention, and pay only when the agent finishes the job.
- What work it does
- Runs agents across voice, chat, email and messaging that process insurance claims, return orders and manage subscriptions. Its tools include Ghostwriter, which builds agents from documents or plain-language instructions, and Insights, which analyzes conversations and agent actions, per this organization profile.
- Typical use cases
- A retailer handling returns and exchanges by voice; an insurer collecting claim details; a subscription brand managing cancellations and plan changes; a bank handling routine account queries.
- Scale
- Reported over 40% of the Fortune 50 as customers and about $200 million ARR in May 2026; raised $950 million that month at a valuation above $15 billion.
- Vouches for the result
- Takes actions in backend systems
Poke
The assistant that texts you first
Visit Poke- Built by
- The Interaction Company of California in Palo Alto, founded by Marvin von Hagen and Felix Schlegel. It was reported acquired by Cognition, the company behind Devin, in July 2026.
- Why it was built
- The founders' thesis is that people do not want another app; they want to text an AI the way they text friends.
- What work it does
- Lives in iMessage, SMS, WhatsApp and Telegram, connects to services such as Gmail, Google Calendar, Notion and Asana, and messages you proactively with one-tap actions.
- Typical use cases
- Drafting and sending email replies; rescheduling a meeting; flagging an unpaid invoice and letting you pay it from the thread; booking travel; setting reminders and recurring automations.
- Scale
- On June 4, 2026 it became the first third-party AI agent approved on Apple's Messages for Business, per TechWyse; it reports about 100 million messages relayed.
- Turns up at the trigger
- Lives inside the work
Which AI agents write and ship code?
Coding is the most mature agent category. The tools split into three jobs: agents you work beside in an editor, agents you hand a ticket to, and agents that turn a prompt into a running app. Comparison guides such as daily.dev's 2026 review reach a similar conclusion: there is no single best agent, only a best fit per job.
| Agent | Built by | What it does | Typical use cases |
|---|---|---|---|
| Claude Code | Anthropic | Terminal and desktop coding agent that reads a repo, edits files, runs commands and tests, and can run parallel sub-agents | Large multi-file refactors, debugging, migrations, building features end to end |
| OpenAI Codex | OpenAI | Coding agent across CLI, IDE extension, desktop app and cloud delegation | Background pull-request work, parallel tasks, terminal workflows |
| GitHub Copilot (agent mode and coding agent) | GitHub, owned by Microsoft | In-editor completions and chat, plus an issue-to-pull-request flow | Teams already on GitHub; assigning an issue and reviewing the resulting PR |
| Devin | Cognition | Cloud "AI software engineer" that takes a scoped ticket and works in its own environment, reachable on the web and Slack | Backlog tickets, bug fixes, small features, library upgrades |
| Devin Desktop (formerly Windsurf) | Cognition | Desktop IDE built around an agent; its local agent reads the whole repository | Refactor-heavy work in an agent-first editor |
| Replit Agent | Replit | Builds, runs and deploys an app from a prompt inside Replit's browser environment | Prototypes, internal tools, small products without local setup |
| Lovable | Lovable (Sweden) | Generates full web apps from plain-language prompts | Founders and marketers building landing pages, MVPs and dashboards |
| Bolt.new | StackBlitz | Browser-based app generation and editing | Quick prototypes and demos |
| v0 | Vercel | Generates UI components and pages, typically React and Next.js | Designers and front-end teams turning a prompt or screenshot into interface code |
| Jules | Asynchronous agent that works on GitHub repositories and proposes changes | Background bug fixes and dependency updates |
Cursor, covered above, sits in the editor-first group. Open-source options such as Cline, Aider and OpenCode let developers bring their own model keys. Pricing and model choices in this category change quarterly, so check each vendor's current plan before you buy.
Which AI agents handle customer support, voice calls and sales?
Customer-facing agents are the second most mature category after coding, because the work is repetitive, measurable and expensive to staff. They split into support agents that resolve tickets, voice agents that answer and place phone calls, and sales agents that research, write and book meetings.
Customer support agents
| Agent | Built by | What it does | Typical use cases |
|---|---|---|---|
| Fin | Fin (formerly Intercom) | Resolves conversations across chat, email, WhatsApp, SMS, phone and Slack | Refunds, order status, account questions, lead qualification |
| Sierra | Sierra (Bret Taylor, Clay Bavor) | Branded agents for large consumer companies, priced per outcome | Returns, claims, subscription changes at Fortune 500 scale |
| Decagon | Decagon | Platform for building and tuning support agents across chat, email and voice | Enterprise ticket resolution with custom workflows |
| Salesforce Agentforce | Salesforce | Agents built inside the Salesforce platform and its data | CRM-connected service, sales and commerce tasks |
| Zendesk AI agents | Zendesk | Automated resolutions inside the Zendesk helpdesk, billed at roughly $1.50 each on a committed plan, per one pricing comparison | Teams already running support on Zendesk |
| Ada | Ada | AI customer service automation for e-commerce, fintech, SaaS and gaming | Automated resolution of repeat questions |
| Parloa | Parloa (Berlin) | AI Agent Management Platform to design, test and run phone and chat agents; raised $350 million at a $3 billion valuation in January 2026, per TechCrunch | Insurance, travel and enterprise contact centers; named customers include Allianz, Booking.com and SAP |
| Crescendo | Crescendo | AI-native contact center that blends agents and human staff | Outsourced-style support with AI doing first-line work |
Gartner predicts that by 2029, agentic AI will “autonomously resolve 80% of common customer service issues”.
Voice agents
Voice agents answer inbound calls, place outbound calls, qualify leads, book appointments and hand complex cases to a person. Most teams choose between a developer platform, a no-code builder, or a voice-quality specialist. Prices below are per minute as of July 2026, from a vendor-by-vendor comparison and exclude telephony and model costs where noted.
| Agent platform | What it does | Best for | Price signal |
|---|---|---|---|
| Vapi | Developer platform to orchestrate voice agents with your own model | Engineering teams building custom voice agents | About $0.05 per minute plus pass-through costs |
| Retell AI | Build, deploy and monitor phone agents with no-code or code | Production call operations | $0.07 to $0.31 per minute |
| Bland AI | Closed in-house stack for large-scale outbound calling | High-volume outbound campaigns | $0.11 to $0.14 per minute |
| Synthflow | No-code voice agent builder | Operators without engineers: booking, qualification, surveys | From about $0.09 per minute |
| ElevenLabs Agents | Conversational agents on top of ElevenLabs' speech technology, deployable on phone, web and messaging | Customer-facing calls where voice quality and languages matter | $0.08 per minute beyond plan minutes plus model costs |
| PolyAI | Enterprise voice agents for customer service, raised $86 million at a $750 million valuation in December 2025, per TechCrunch | Large contact centers | Custom |
Typical voice agent use cases include after-hours answering for clinics, appointment confirmation, order status calls, collections reminders, lead qualification and outbound surveys.
Sales agents
Sales agents fall into four jobs: autonomous outbound reps, research and enrichment, engagement and coaching, and CRM-native assistants.
| Agent | Built by | What it does | Typical use cases |
|---|---|---|---|
| 11x (Alice for email and outreach, Julian for voice) | 11x | Autonomous digital sales rep that researches, writes, follows up and books meetings | Outbound prospecting at volume |
| Artisan (Ava) | Artisan | Autonomous AI SDR covering lead sourcing, outreach and reply handling | Early-stage teams without an SDR team |
| Clay | Clay | Research and enrichment layer with more than 150 data sources and AI research agents | Building precise target lists and personalized talking points |
| Apollo | Apollo.io | Prospecting database with an AI assistant and sequencing | Finding contacts and running outreach in one tool |
| Regie.ai | Regie.ai | Content engine and auto-pilot agents with human-in-the-loop controls | Writing sequences, scripts and personalized messages at scale |
| Gong | Gong | Call recording and deal intelligence | Coaching, forecasting, spotting deal risk |
| Agentforce and HubSpot Breeze | Salesforce, HubSpot | CRM-native agents | Lead follow-up and record updates inside the CRM |
Which AI agents work in legal, healthcare, finance, HR, security and data?
Industry agents win by combining a general model with the documents, workflows and compliance rules of one profession. They tend to sell to regulated, high-stakes teams, so trust, citations and human sign-off matter more than raw speed. Valuations and revenue figures in this section come from third-party trackers and should be treated as approximate.
| Domain | Agent | Built by | What it does | Typical use cases |
|---|---|---|---|---|
| Legal | Legora | Founded in Stockholm in 2023 as Leya by Max Junestrand, August Erseus and Sigge Labor | Collaborative AI workspace that works inside Word and Outlook; reported $5.6 billion valuation in April 2026 and about $100 million ARR, per Sacra and YesPress | Document review, due diligence, drafting, regulatory work |
| Legal | Spellbook | Spellbook | Microsoft Word add-in for contract drafting and review | Redlining and drafting contracts at small and mid-sized firms |
| Legal | CoCounsel | Thomson Reuters | Legal AI tied to Westlaw and Practical Law content | Legal research, drafting, document analysis |
| Legal | EvenUp | EvenUp | Builds claim packages for plaintiff personal injury firms; says it processes about 10,000 cases a week | Demand letters and case preparation |
| Healthcare | Abridge | Abridge | Ambient clinical documentation that turns visits into notes inside the health record | Doctors' visit notes, reducing after-hours charting |
| Healthcare | Ambience Healthcare | Ambience | Operating-system approach covering notes, coding and billing | Documentation plus coding across specialties |
| Healthcare | Nabla | Nabla (France and US) | Ambient scribing and dictation with multilingual support | Clinical notes in multiple languages |
| Healthcare | Hippocratic AI | Hippocratic AI | Patient-facing agents focused on non-diagnostic tasks | Appointment prep, medication reminders, post-discharge follow-up calls |
| Finance | Rogo | Founded in 2021 by former investment bankers | Agents for banking workflows; reported $160 million Series D at about $2 billion in April 2026, per AllMind's market review | Comparable companies, CIM drafting, diligence memos, Excel model roll-forwards |
| Finance | Hebbia | Hebbia | Document analysis platform that links outputs to source passages | Diligence and credit review needing traceable citations |
| Finance | AlphaSense | AlphaSense | Market-intelligence search over a large library of filings, transcripts and research | Competitive and earnings research |
| HR | Paradox (Olivia) | Paradox, founded 2016; acquired by Workday in 2025 | Conversational recruiting agent that screens, schedules and answers candidates | High-volume hourly hiring: retail, restaurants, logistics |
| HR | Eightfold AI | Eightfold | Talent platform across recruiting, internal mobility, succession and reskilling | Matching candidates and employees to roles |
| Security | Dropzone AI | Dropzone AI | AI analyst that investigates security alerts | Triage of security operations center alerts |
| Security | Big Sleep | Google (Project Zero and DeepMind) | Agent that hunts for software vulnerabilities | Finding previously unknown bugs in widely used code |
| Data | Julius AI | Julius AI | Chat-style data analyst that cleans data and builds charts | Analysis of spreadsheets and databases without code |
| Data | Hex Magic | Hex | AI features inside Hex's collaborative data notebook | Writing SQL and Python, explaining and fixing queries |
For the security and data rows, the descriptions come from vendor positioning rather than independent testing, so verify capabilities in a trial before buying.
Which AI agents handle meetings, research, browsing, workflows and India-specific work?
This group covers the horizontal agents that help any knowledge worker, plus the Indian vendors built for Indian languages and phone-first customers. It also has the fastest churn: several well-known products changed hands or shut down in 2026.
Meetings and note-taking agents
| Agent | What it does | Typical use cases |
|---|---|---|
| Granola | Local, bot-free notes that blend your typing with the transcript (covered above) | One-on-ones, customer calls, product reviews |
| Otter.ai | Transcription and meeting summaries with a bot or recorder workflow | Interviews, lectures, team meetings |
| Fireflies.ai | Meeting bot that records, transcribes and makes calls searchable; reported to have more than 16 million users and a $1 billion valuation, per The Next Web | Sales call libraries, team knowledge search |
| Fathom | Call summaries for Zoom, Google Meet and Teams | Customer and sales calls |
| Read AI | Meeting summaries and analytics | Recaps and engagement tracking |
The same Next Web report notes that notes-plus-summary has become a commodity, and that Notion, Microsoft and Google are building the feature into their suites. The differentiator is what the agent does with the meeting afterward.
General-purpose, research and browser agents
| Agent | Built by | What it does | Status to know |
|---|---|---|---|
| Manus | Monica (formerly Butterfly Effect), a Chinese-founded startup | Runs tasks autonomously in a cloud computer; its Wide Research feature runs many sub-agents in parallel | One review notes the site now says Manus is part of Meta, per eesel |
| Genspark | Genspark | Turns a prompt into a finished deliverable; has a phone-calling module | Reported $385 million Series B in April 2026; credit-based billing |
| Perplexity (Comet and Deep Research) | Perplexity | Comet is an agentic browser that navigates sites and fills forms; Deep Research produces cited reports | Comet launched on Android on August 19, 2026, per Tech Insider |
| ChatGPT (Work, Codex and Deep Research) | OpenAI | General assistant with research and agent features | Atlas, OpenAI's standalone browser, was discontinued on August 9, 2026 and its browser-agent features moved into ChatGPT and Codex, per TechRadar. Wikipedia reports the earlier ChatGPT agent mode was removed in early August 2026 in favor of ChatGPT Work |
| Gemini Deep Research and Gemini Agent | Research reports and browser tasks, including Chrome Auto Browse | Folded into Chrome rather than shipped as a separate browser | |
| Claude (Claude in Chrome, Cowork, Claude Code) | Anthropic | Assistant with a browsing agent in Chrome and an agentic desktop app for knowledge work | Available in Anthropic's apps |
Typical uses: market research reports with citations, competitor tracking, form filling and booking on websites, building slide decks or spreadsheets from a brief, and long multi-source research that would otherwise take an analyst a day.
Workflow automation and productivity agents
| Agent | Built by | What it does | Typical use cases |
|---|---|---|---|
| Lindy | Lindy | No-code builder for business agents | Inbox triage, meeting scheduling, lead follow-up |
| Zapier Agents | Zapier | Agents that work across Zapier's large app library | Cross-app tasks such as updating a CRM after a form fill |
| n8n | n8n (Berlin) | Workflow automation with AI agent nodes, self-hostable | Technical teams wiring models into their own stack |
| Gumloop | Gumloop | No-code AI workflow builder | Marketing and operations automations |
| Notion Agent | Notion | Agent that works on pages and databases in your workspace | Updating docs, summarizing projects, drafting from notes |
| Microsoft Copilot | Microsoft | Assistant and agents inside Microsoft 365 | Drafting in Word, summarizing Teams meetings, building Excel analysis |
India-focused agents
Indian enterprises need agents that handle code-switching between languages, regional accents, WhatsApp and phone-first customers. These vendors build for that.
| Agent | Built by | What it does | Typical use cases |
|---|---|---|---|
| Yellow.ai | Yellow.ai, founded 2016 | Conversational agents across voice, chat, WhatsApp and email; claims 135+ languages in total and about 20 Indian languages on voice, per Acefone | Banking, retail, telecom and healthcare customer service |
| Haptik | Part of the Reliance Jio group | Enterprise chat agents extended into voice | E-commerce, banking and travel support |
| Gnani.ai | Gnani.ai | Voice agents trained on Indian speech; builds its own speech and reasoning models and sells voice biometrics. Its own materials cite 200+ enterprises and 30 million daily calls, per Gnani's comparison page | Collections, sales calls, insurance and banking contact centers |
| Sarvam AI | Sarvam AI | Full-stack sovereign AI: Samvaad voice agents for Indian languages and Arya for enterprise workflows, per Sarvam | Loan servicing, insurance renewals and claims calls in 10+ Indian languages |
| Skit.ai | Skit.ai | AI phone agents | Banking collections and insurance calls |
| Bolna | Bolna | Voice agent orchestration for Indian-language calls | Developers building Indian-language calling agents |
Vendor figures in this table are self-reported or from comparison sites, so request a pilot with your own call recordings before committing.
How do you choose the right AI agent for your team?
Start from the job, not the tool. Name the task and its finish line, check the agent against the six traits, pick the autonomy level you can supervise, model the cost at your volume, and test on your own data before you commit.
- Name the job and its finish line. "Refund issued and customer notified" or "pull request opened with passing tests" is a finish line. "Help with support" is not.
- Score the agent on the six traits. Does it read your data, sit in your workflow, take actions, stand behind the result, act at the trigger, and offer an interface built for the job? Two or fewer usually means a chatbot.
- Pick the autonomy level. Suggest only, act after approval, or act alone. Start one level more cautious than you want.
- Model the pricing. Seats, per-minute, credits and per-outcome pricing behave differently at scale. At Fin's $0.99 per resolution, 1,000 resolved conversations a month costs $990 before seat fees; at 10,000 it is $9,900. Compare that with your current cost per resolved ticket.
- Pilot with your own data. Run 50 to 100 real cases and measure resolution rate, error rate and the quality of hand-offs to humans. Vendor benchmarks and your data rarely match.
- Plan for change. In 2026 alone, Intercom became Fin and agreed to be acquired by Salesforce, Windsurf became Devin Desktop, OpenAI shut down Atlas, and Poke was reported acquired by Cognition. Keep your data exportable and avoid building on a single vendor's proprietary format.
Pick by job
| Job to be done | Start with |
|---|---|
| Write code while you stay in control | Cursor, GitHub Copilot |
| Delegate whole tickets or refactors | Claude Code, OpenAI Codex, Devin |
| Resolve customer support tickets | Fin, Sierra, Decagon |
| Answer or place phone calls | Retell AI, Vapi, ElevenLabs Agents; Parloa or PolyAI at enterprise scale |
| Outbound prospecting and meeting booking | 11x or Artisan, fed by Clay or Apollo |
| Legal drafting, review and diligence | Harvey, Legora, Spellbook |
| Capture and search meetings | Granola, Fireflies.ai, Fathom |
| Cited research and web tasks | Perplexity, Manus, Gemini, Claude, ChatGPT |
| Connect your apps with agents | Zapier Agents, n8n, Lindy |
| Voice agents in Indian languages | Gnani.ai, Sarvam AI, Yellow.ai |
Frequently asked questions about AI agents
What are AI agents, with examples?
AI agents are software systems that pursue a goal by deciding the steps, using tools and acting inside other systems. Examples include Cursor for coding, Fin for customer support, Harvey for legal work, Granola for meetings, Poke for personal tasks by text, and Retell AI for phone calls.
What is the difference between an AI agent and a chatbot?
A chatbot answers questions in a chat window and leaves the task to you. An AI agent finishes the task itself by reading your data and changing things in other systems, such as issuing a refund or opening a pull request. Agents can also start work on a trigger instead of waiting for a prompt.
Which AI agent is best for coding?
There is no single best one. Claude Code and OpenAI Codex are often chosen for hard multi-file work and delegated tasks, Cursor for editor-first daily coding, GitHub Copilot for GitHub-centered teams, and Devin for well-scoped tickets. Choose by how much control you want to keep.
How much do AI agents cost?
Pricing models differ. Cursor's Pro plan is listed at $20 a month, Fin charges $0.99 per resolved conversation, voice agent platforms run from about $0.05 to $0.31 per minute, and enterprise agents such as Sierra and Harvey are quoted privately. Model your own volume before comparing.
What is outcome-based pricing for AI agents?
Outcome-based pricing means you pay only when the agent completes a defined result, such as a resolved support conversation, instead of paying per user or per call. Fin and Sierra both use it. The contract's definition of "resolved" decides what you actually pay, so read it closely.
What are voice agents used for?
Voice agents answer and place phone calls using speech recognition, a language model and speech synthesis. Common jobs are after-hours answering, appointment booking, order status, payment reminders, lead qualification and surveys, with a hand-off to a human for complex cases.
Will AI agents replace human workers?
Agents are taking over repetitive, well-defined tasks, but results are mixed. A 2026 buyer's guide cites data that AI sales-rep tools churn at 50 to 70 percent a year, and in legal work a lawyer still signs off. Most teams shift people toward review, exceptions and relationships.
Which Indian companies build AI agents?
Yellow.ai, Haptik (part of the Reliance Jio group), Gnani.ai, Sarvam AI, Skit.ai and Bolna build agents for Indian languages and phone-first customers. Most focus on voice and customer service for banks, insurers, retailers and telecom companies.
Conclusion: what should you remember about AI agents?
The useful question is not "which AI agent is best" but "which agent finishes this exact job inside my systems, and who stands behind the result." Harvey, Cursor, Granola, Fin, Sierra and Poke each answer that for one job. The market around them changes monthly, so test with your own data and keep your options open.
Sources and notes
Figures are as reported in the pages below, checked on October 5, 2026. Valuations, revenue and user numbers are company-reported or from third-party trackers and may differ between sources. Use cases are drawn from vendor descriptions and press coverage; where a scenario is illustrative, it is labeled as typical.
- Harvey: funding announcement, March 2026; company profile
- Cursor: Wikipedia company page; Amplifying tracker
- Granola: The Next Web, March 2026; Pulse 2.0
- Fin: Salesforce announcement via Business Wire; MarTech; pricing summary
- Sierra: organization profile
- Poke: TechWyse on Apple Messages for Business; Layer3 Labs review
- Coding agents: daily.dev comparison; Dupple comparison
- Support and voice agents: TechCrunch on Parloa; voice platform comparison; Zendesk and Fin pricing comparison
- Sales agents: Autobound buyer's guide; Mutiny sales agent guide
- Legal, finance and HR: Sacra on Legora; AllMind equity research review; Workday and Paradox announcement
- General-purpose and browser agents: TechRadar on Atlas; Tech Insider on Comet; eesel on Genspark and Manus
- India: Sarvam; Gnani comparison; Acefone voice platform guide
- AEO and GEO guidance: Trio Media, how to write blog content for AI search; Superblog AEO guide
- Source text on the six traits: The passage that frames this guide, which cites a16z's State of Consumer AI 2025