The best AI agents for business in 2026, by the job they do

Meronym builds custom AI agents, sells none of the products in this list and has no referral or partner deal with the companies that do. We read every vendor's own pricing pages and documentation on 27 September 2026. We didn't run trials, so what each agent can do is what its maker documents, quoted and linked, and all prices are in US dollars.

What AI agents for business do, and how few businesses let one act

An AI agent chooses its next step and acts in your systems through tools: it reads the ticket, looks up the order, issues the refund or hands the case to a person. A chatbot answers questions, and a workflow follows a path someone drew in advance. The difference that matters when you buy one is the write: what the agent may change on its own, and what waits for a person.

The jobs businesses give agents are unglamorous. In a Reddit thread on small businesses using agents, which ranks for these searches, owners and the people who build for them name four:

• First-line support. One owner is "currently using an AI agent to automate customer service (T1)" (January 2025): the work of an AI customer service agent.
• Booking appointments by phone. An automation agency owner says "Most of the small businesses want to automate the appointment bookings process", the job of AI voice agents.
• Qualifying leads and following up on invoices. "Think lead pre-qualification, invoice nudges, or FAQ replies," another reply says (April 2025): work for an AI sales agent and AI agents for finance.
• Documents and books. A builder lists "document data extraction, bookkeeping, email, sms, content, and meeting recording review pipelines".

The first figure is from the US Census Bureau's Business Trends and Outlook Survey, released on 24 September 2026. It counts AI of any kind, not only agents, and the question changed in November 2025, so it can't be set beside the 9 to 10% figures quoted before then; the smallest firms match the national rate, and firms with 250 or more employees reach 44.0%. The second is from the Main Street AI Monitor, a probability sample of 1,070 small-business workers by the US Chamber of Commerce Foundation and Ipsos, published on 17 June 2026: half use AI, mostly to draft and summarise, and "Just 6% say they use it to automate workflows with minimal human involvement."

The third is from Gartner's 2026 survey of CIOs and technology executives, in which more than 60% expected to deploy agents within two years. McKinsey's 2026 survey shows the gap by size: 40% of respondents from organisations with more than $1 billion in revenue report scaling agents, against 22% at smaller ones.

So where AI is in use at all, it mostly drafts and summarises, and few businesses let it act. The first agent worth buying finishes one job, with a person approving whatever it changes until it has earned more.

How we chose

We started from every product recommended on the pages ranking for AI agents for business and two neighbouring searches, "ai agents for small business" and "best ai agents": 82 across 14 pages, 60 of them named only once. We added six none of them named: five agents built into software businesses already run (Microsoft 365 Copilot, Gemini Enterprise, ServiceNow's AI agents, Atlassian's Rovo and GitHub's Copilot cloud agent) and one enterprise support agent, Decagon. That made 88. On 27 September 2026 we checked 30 of them on their makers' own pricing pages and documentation: the 20 named on two or more lists, leaving out Reclaim.ai, a calendar tool, and nexos.ai, recommended only by its own publisher, plus the six we added and four the lists name once (ChatGPT's business plan, Sierra, Gumloop and Amazon's AgentCore). Twenty-five passed all five rules:

• Generally available now. A waitlist, a preview or a beta as the main offer rules a product out.
• Sold to businesses at a published price, or a stated way of pricing. "Contact sales" is enough only when the vendor says what it charges for.
• Finishes a job in a business system. It takes an action or hands over finished work; answering in a chat window isn't enough.
• Says what it can change, and how a person steps in. Its documentation names the actions it takes and the approval or handover around them.
• Named on at least two of the lists, or part of a platform businesses already pay for.

Make's agents, named on three lists, failed the first rule: the app is in open beta, and Make says "product functionality and pricing may change". Manus, named on two, failed the fourth: none of the pages we read describes a step where a person approves what it does. Gumloop is recommended only on its own list. Sierra and Decagon sell only on enterprise contracts and are named on one list between them, so they're mentioned in the customer service group rather than given entries. Salesforce and HubSpot each sell one agent for service and another for sales, so each has two entries; with the three developer platforms in their own table, that makes the 25.

Where a vendor's page left something out, the entry says so. The last entry, a custom agent, is ours: it sits outside the 25 and the rules, and it carries the same fields as the rest.

Ten questions to ask before you pick one

Take these into any demo. A product that can't answer one in writing has told you something.

• Does it finish the task, or prepare a draft? An agent that drafts replies can save a person time; one that resolves the request completes the task end to end.
• Where does it have to live? The agent built into the system the job happens in needs no integration project; one outside it needs a connection to every system it reads or writes.
• What can it change? Ask for the list of actions: refunds, record updates, emails sent. Anything missing from the list it can't do, and anything on the list it eventually will.
• Which actions wait for a person, and is that the default? Some products ask before every change until you switch that off; others approve nothing until you switch it on.
• Who does it hand over to, and with what? A handover should reach a named queue with the conversation and what the agent already tried.
• What exactly is the unit you pay for, and who decides it happened? A resolution, an outcome, a conversation and a credit are counted differently by every vendor; Relevance AI's documentation says a tool run that fails "will still count as one Action". Our page on the AI customer service agent explains why the definition of a resolution matters more than its price.
• What must you already pay for? HubSpot's customer agent needs HubSpot Professional or Enterprise, Agentforce for Service is sold from Salesforce's Core edition up, $195 a user a month, and ServiceNow's autonomous IT agents come only in ITSM Prime; the edition can cost more than the agent. Each entry's "Needs you to run" line says what it needs.
• What happens when the allowance runs out? Some agents pause until next month, some keep working and bill you, and Botpress's paid plans add capacity you can't switch off.
• Where does your data go? Whether the vendor or its model providers train on it, where it's stored, and whether you can host it yourself.
• How do you test it on your own cases before it acts? A trial that can publish, a sandbox, or a period where it drafts and people send.

Ask what you must already pay for, and what the unit is, first. Those two decide whether an agent is possible at all, and what it costs once it works.

Customer service agents

These answer customers and take small actions, and most charge per resolved conversation: from about $0.20 (the standalone Lyro, depending on the model) and $0.50 (HubSpot) to $2.00 (Zendesk's pay-as-you-go rate and Salesforce's Help Agent) among the ones we checked. Start with the one built into your helpdesk. Two firms sell only on enterprise contracts with no public price: Sierra, which wrote in December 2024 that it "gets paid only when we complete a task for you", and Decagon, which offered per-conversation or per-resolution pricing in a post the same month. For the platforms a large organisation standardises on, see our guide to enterprise AI agents.

Zendesk AI agents: best for teams already on Zendesk, with resolutions included in the plan

What it does. Zendesk's AI agents resolve requests over messaging, email and voice, and on its pricing page they "Connect to external systems to read, write, parse data, and trigger actions". Zendesk also sells Forethought's agents since completing its purchase of Forethought in March 2026.

Needs you to run. A Zendesk Support plan, from $19 per agent per month on annual billing, for AI agents on email; AI agents on messaging need a Suite plan, from $55.

Price. Plans include automated resolutions (5 per agent a month on Support Team, up to 10,000 a year); after that they cost $1.50 each on a commitment or $2.00 pay-as-you-go. Only a verified resolution, one a second model confirms, is billed. The trial runs 14 days with no card.

For what it can't reach, see our Zendesk AI agents page.

Fin: best for paying only for resolved conversations, on the helpdesk you already have

What it does. Fin, from the company formerly called Intercom, answers customers over chat, email, SMS, WhatsApp and social, and its maker says it "updates accounts, processes payments and refunds". Salesforce completed its purchase of Fin on 10 September 2026.

Needs you to run. Nothing in particular: Intercom's helpdesk, or yours, from Salesforce and HubSpot to Zendesk and Gorgias.

Price. $0.99 per outcome, at most one per conversation. An outcome is a resolution or a completed hand-off, and a customer asking for a person costs nothing. On another helpdesk the minimum is 50 outcomes a month. The trial runs 14 days with no card.

For what it can't reach, see our Intercom AI agents page.

HubSpot's customer agent: best for HubSpot Service Hub teams, at $0.50 a resolution

What it does. It "resolves questions, takes action in your CRM, and hands off to your team with full context", HubSpot says: resetting a password, booking a meeting, and passing to a person when its confidence is low. It's one of what HubSpot calls its "core agents"; since 23 July 2026 they're managed from Agent Hub, which is in public beta with Agent Builder. The ranking lists still call them Breeze.

Needs you to run. HubSpot Professional or Enterprise (Service Hub Professional from $90 a seat a month on annual billing) and a chat, WhatsApp or Facebook channel.

Price. $0.50 per resolution, meaning a conversation not handed to a person for 72 hours, paid in HubSpot Credits at $0.010 each. Buying at least one Service Professional or Enterprise seat gives you the agent free for 14 days.

For what it can't reach, see our HubSpot integration services page.

Agentforce Help Agent: best for Salesforce service teams that want an agent grounded in Salesforce Knowledge

What it does. Help Agent, generally available since July 2026, is "designed for one job: helping enterprises to resolve their customer issues from start to finish, on day one". It manages cases, schedules appointments and updates orders, and hands the full context to your team on escalation.

Needs you to run. Service Cloud's Advanced edition, $395 a user a month billed annually, whose card on Salesforce's pricing page lists the Help Agent, with self-service add-on licences for per-resolution billing.

Price. $2 per resolution with no person involved; escalations are free, and the actions inside a resolved session aren't billed on top. Salesforce's other meters are Flex Credits, at $0.10 an action, or $2 per conversation.

For what it can't reach, see our Salesforce integration services page.

Tidio Lyro: best for small businesses that want an agent in their website chat, priced by the conversation

What it does. Lyro answers customers from your own support content in chat, email and social channels, and through Lyro Actions it "can update addresses, log new leads, or trigger refunds, all via secure APIs across your systems". Each action can be set to ask the customer to confirm first, and Lyro hands over to a person on request.

Needs you to run. A free Tidio account. Running Lyro inside Zendesk, Salesforce or another helpdesk needs Tidio's Plus or Premium plan, or the standalone version at getlyro.ai.

Price. From $39 a month billed monthly, or $32.50 billed annually, for 50 Lyro conversations, up to $700 a month for 1,000. A Lyro conversation is any thread with at least one reply from Lyro. The first 50 are free, once, and there's a 7-day trial. The standalone Lyro at getlyro.ai charges only for conversations it resolves, "from ~$0.20" each depending on the model.

Botpress: best for acting on refunds and accounts, on a monthly conversation allowance

What it does. Botpress "Takes real action on refunds, account updates, and multi-step workflows" in Zendesk, Intercom, HubSpot or Salesforce, with escalation rules deciding when it hands over.

Needs you to run. Nothing: it works on top of your helpdesk, or as one.

Price. Free for 25 conversations a month, with no top-ups or overages; Plus is $150 a month billed annually for 250, with packs of 100 at $65, and Team $750 for 1,500. A conversation is any exchange with at least two messages from the customer, resolved or not. On Plus and Team, each conversation includes $0.10 of AI usage; past that, Botpress buys $10 of AI spend at cost each time your workspace reaches 95% of its limit.

Sales agents

These work leads inside the CRM you already pay for. Our page on the AI sales agent says most teams should buy one rather than build it, and when that stops being true.

Agentforce Engagement: best for Salesforce teams that want every new lead emailed, answered and handed to a rep

What it does. Agentforce Engagement, called Agentforce SDR at launch, works new leads by email: Salesforce says it sends each prospect an introductory email, replies, answers questions and passes qualified leads to reps. A Require Manual Approval setting holds every email for a rep.

Needs you to run. Salesforce Enterprise, Performance or Unlimited edition with Sales and Einstein for Sales.

Price. Flex Credits at $0.10 an action, or $2 per conversation, one per lead from the first email. A newer "Leads Worked" meter counts each person once per 30 days; we found no published price for it.

HubSpot's prospecting agent: best for HubSpot teams, from Starter up, that want outreach researched and drafted for review

What it does. HubSpot says its prospecting agent researches accounts, finds the right contacts and drafts outreach, drawing on data providers such as ZoomInfo and Apollo. You can review each draft before it goes, or switch it to send on its own.

Needs you to run. HubSpot Starter, Professional or Enterprise.

Price. $1 for each lead it recommends outreach for, paid as 100 HubSpot Credits. Before 14 April 2026 it charged monthly for every contact enrolled.

Clay: best for sales teams that research and score accounts before anyone writes to them

What it does. Claygents "are agents in Clay that perform tasks on your behalf. They take inputs, follow your instructions, and write a structured output": account research, contact finding and lead scoring, written into Clay's tables. Clay's newer Account Agents, in open beta since August 2026, pick a next step from the actions you allow and update your CRM, with the write-back approved by a person.

Needs you to run. A Clay account. Account Agents need the Launch plan or above, connected to your CRM, data warehouse or Gong.

Price. Free for 500 actions and 100 data credits a month; Launch costs $185 a month billed monthly ($167 annually) and Growth $495 ($446). An agent uses one action per prompt plus data credits, from $0.05 each. The trial runs 14 days.

Marketing agents

Our page on AI agents for marketing says to buy the tool for drafting, variants, repurposing and scheduling, and to build only when claims must come from an approved list or approvals differ by audience.

Jasper: best for marketing teams producing on-brand content at volume

What it does. Jasper's agents, more than 100 of them since its relaunch in June 2025, handle research, drafting, personalisation, translation and search optimisation in your brand voice, and send finished content to tools such as Adobe Workfront and Box. Jasper says people stay in control by approving outputs; we found no agent action that changes records in other systems.

Needs you to run. A Jasper seat; the advanced agents and custom agents need the Business plan.

Price. Pro costs $69 a seat a month billed monthly or $59 billed yearly; Business is custom, with a 12-month commitment. Advanced agents run on credits, 40 to 100 a run, and Jasper doesn't publish a credit's price. The Pro trial runs 7 days.

Office work and research agents

These do a team's everyday work: documents, email, meetings, research and spreadsheets. Start with the one inside the suite you already pay for. All of them charge per seat, with heavier use metered on top.

Microsoft 365 Copilot: best for teams whose documents, email and meetings live in Microsoft 365

What it does. The seat includes Microsoft's Researcher and Analyst agents, generally available since June 2025, and Facilitator for meetings. Copilot Cowork takes on multi-step work: it "plans the steps, picks the right skills and apps, pauses for your approval at key checkpoints". Agents you make with Agent Builder use knowledge only; actions need Copilot Studio.

Needs you to run. A qualifying Microsoft 365 business or enterprise plan.

Price. $30.00 per user per month, paid yearly. The Business add-on, for up to 300 users, is $21.00 paid yearly ($18.00 in the first year on orders by 31 December 2026) or $25.20 billed monthly. The one-month trial needs a card.

For what it can't reach, see our Microsoft 365 AI agents page.

Gemini Enterprise: best for Google Workspace teams, or one agent app across Workspace and Microsoft 365

What it does. Google's agent app, renamed from Agentspace in October 2025, holds search, Google's agents such as Deep Research, and no-code agents from Workflow Builder, generally available since 3 September 2026. Its workflows act through connectors to Gmail, Slack, Jira, ServiceNow, SharePoint and Outlook, with human approval steps. Workspace Studio, included in Workspace Business and Enterprise plans, builds agents inside Gmail, Drive and Sheets.

Needs you to run. Nothing for the Business edition: "No IT setup needed."

Price. Business starts at $21 per seat per month, for up to 300 seats; Standard and Plus start at $30 through sales. Google doesn't say whether that's a monthly or an annual term. The trial runs 30 days. On Standard and Plus, daily quotas are pooled across the edition's users rather than capped per person.

For what it can't reach, see our Google Workspace AI agents page.

ChatGPT Business: best for a small team that wants shared agents without an IT project

What it does. Workspace agents, "an evolution of GPTs", are agents a team builds once and shares in ChatGPT or Slack; they run code, use connected apps, and can be made to ask permission before sensitive steps such as sending an email. ChatGPT Work, launched in July 2026, works across apps and files to "turn a goal into finished work". Standard seats also include Codex, OpenAI's coding agent. The plan was called ChatGPT Team until August 2025.

Needs you to run. At least two paid seats.

Price. Standard seats cost $20 per user per month on annual billing or $25 monthly, and Premium seats $100 or $125. Use beyond the allowance is paid in credits (agent mode costs 30 a message), and we found no dollar price for a credit. No trial is stated.

Claude Team: best for handing off multi-step document work, with approval before every write by default

What it does. Claude Cowork takes a goal "and it works across your files and tools", through connectors to Google Drive, Gmail, GitHub, Microsoft 365 and Slack, and hands the work back for review. Anthropic says "By default, Claude asks before taking an action", and on Team plans write tools need approval for each task unless an owner turns on a setting that is off by default. Claude Code comes with every Team seat.

Needs you to run. A paid plan: Team takes 2 to 150 seats.

Price. Standard seats cost $20 per seat per month on annual billing or $25 monthly, and Premium seats $100 or $125, with five times the usage. Enterprise is $20 a seat plus all usage at API rates. Anthropic lists trials only for sales-assisted Enterprise, not for Team.

Lindy: best for a small team that wants one assistant in Slack and email that asks before it changes anything

What it does. Lindy works through Gmail, Slack, Notion, HubSpot and more than 1,000 other integrations, and its pricing page says "Anything with outside impact waits for approval such as sending an email, updating a ticket, posting to another channel, publishing a doc."

Needs you to run. A seat for everyone who uses it, including anyone who mentions it in Slack.

Price. Plus costs $29.99 per user per month with 3,000 credits, Pro $99.99 with 15,000 and Max $199.99 with 35,000; a job uses 2 to 2,500 credits. The 7-day trial comes only through a Slack invitation.

Notion: best for teams whose documents and tasks already live in Notion

What it does. Notion has two kinds of agent: Notion Agent, a personal assistant since September 2025, and Custom Agents, which since February 2026 run team work on a schedule or trigger, such as answering repeat questions or routing tasks. "You decide what it can read and write across Notion, Slack, Mail, Calendar, and MCP integrations", and "Every agent run is logged, and all changes are reversible."

Needs you to run. A Notion workspace; Custom Agents need the Business or Enterprise plan.

Price. Business costs $20 a member a month billed yearly or $24 monthly. Custom Agents then run on Notion credits at $10 per 1,000 a month, shared across the workspace, and are free to try.

IT and HR helpdesk agents

These take requests from staff rather than customers: password resets, access, equipment, policy questions. Our pages on the AI service desk and AI agents for HR cover when those requests need an agent of their own.

ServiceNow AI Agents: best for IT and HR teams already running ServiceNow

What it does. ServiceNow's agents "act autonomously to get work done" across IT, customer service and HR, and reach other systems over MCP; its developer guide has agentic workflows ask the user to confirm by default, and recommends the same for tools that change records. ServiceNow completed its purchase of Moveworks in December 2025.

Needs you to run. The ServiceNow AI Platform; the autonomous "AI Agents for ITSM" come only in the top package, ITSM Prime.

Price. Contact sales: the agents come in ServiceNow's Foundation, Advanced and Prime packages, quoted per customer. We found no trial.

For what it can't reach, see our ServiceNow integration services page.

Atlassian Rovo: best for teams whose requests and projects live in Jira and Confluence

What it does. Rovo agents create and edit Jira work items and Confluence pages and use connected tools such as Slack. Atlassian says an agent asks "for confirmation before executing consequential tools that may mutate data across systems", and it can do only what its user is allowed to.

Needs you to run. A paid Atlassian Cloud plan of Jira, Confluence or one of its Collections, with nothing extra to buy.

Price. Credits pooled across the organisation: 25, 70 or 150 per user a month on Jira and Confluence Standard, Premium and Enterprise. A basic agent interaction uses 10, and extra use costs $0.01 a credit, from 31 August 2026.

For what it can't reach, see our Jira AI agent page.

Bookkeeping agents

Intuit's QuickBooks agents: best for small businesses that keep their books in QuickBooks Online

What it does. Intuit AI, as Intuit's pages now call the set, keeps the books categorised and reconciled (Accounting AI), proposes invoice reminders (Payments AI) and covers sales tax, leads and finance. Intuit Intelligence, in beta, drafts invoices for review and sends reminders with your approval, and payroll runs only when you approve it.

Needs you to run. QuickBooks Online; which agents you get depends on the plan, with reconciliation from Plus.

Price. Nothing extra: QuickBooks Online costs $38, $85, $140 or $340 a month by plan, with half off for three months or a 30-day trial.

For what it can't reach, see our QuickBooks integration services page.

Build your own agent, without code

For a job no product covers but whose steps you can describe, and someone on the team who builds automations. The agent is yours to test and to keep working.

Microsoft Copilot Studio: best for building agents that act across Microsoft 365 and your other systems

What it does. Copilot Studio "is a platform for building and managing agents" that act through flows and APIs, connect to more than 1,400 systems and MCP servers, and publish where staff and customers are. Its agent flows can require multistage approvals, still in preview.

Needs you to run. An Azure subscription.

Price. Copilot Credits: $200 for 25,000 a month, or $0.01 each pay-as-you-go. In Microsoft's examples, a generative answer uses 2 and an agent action 5. Staff with a Microsoft 365 Copilot licence use your agents at no charge, and the trial can't publish.

AI by Zapier: best for teams whose automations already run in Zapier

What it does. Zapier is folding its Agents product into AI by Zapier, an agent step in the Zap editor, generally available since July 2026. It uses actions from more than 9,000 apps, and each tool has a "Require approval before running" switch, off by default.

Needs you to run. A paid Zapier plan to publish an agent step.

Price. Zapier tasks: one per run plus one per tool call, times the model's rate (1x, 3x, or 5x for the default tier). Plans start at $19.99 a month on yearly billing, and the standalone Agents plans remain, from free (400 activities a month) to $400 a year (1,500).

n8n: best when the agent has to run on your own servers

What it does. n8n's AI Agent node: "Connect a chat model and one or more tools, and the agent decides which tools to call to complete a task." A tool can be any n8n integration, an HTTP request, code or MCP, and chosen tools can wait for a person's approval in Slack, Telegram or n8n Chat.

Needs you to run. n8n Cloud or your own servers, and your own model account.

Price. The self-hosted Community Edition is free. Cloud costs $20 a month for 2,500 executions or $50 for 10,000, on annual billing. An execution is "a single run of your entire workflow", however many steps. The Cloud trial needs no card.

For what it can't reach, see our n8n automation agency page.

Relevance AI: best for building agents that follow your existing workflows, with model costs passed on without markup

What it does. Relevance AI sells agents "for sales, customer success, marketing and HR" that, its documentation says, "can follow established workflows and use the same applications as your human workforce", on autopilot or with a person approving.

Needs you to run. Nothing stated; it's sold through sales.

Price. Contact sales: the pricing page lists only Enterprise. Usage is counted in Actions (one tool run) and Vendor Credits (the model's cost, without markup), and top-ups cost $80 per 1,000 Actions.

Build your own agent, with code

For a team with engineers who would rather own the agent's code. These are frameworks and platforms, not finished agents, so the prices below are the platform's alone; the model is billed on top. Each can make a tool wait for a person: CrewAI's paid platform through an email review step, Google's Agent Development Kit through a tool confirmation feature it marks experimental, and AgentCore through a policy that can demand a recorded human approval. Google renamed Vertex AI as Gemini Enterprise Agent Platform on 22 April 2026; the lists that rank for this search still use the old name.

Software development agents

Claude Code comes with every Claude Team seat, and Codex with ChatGPT Business Standard seats; both are in the office entries above.

GitHub Copilot cloud agent: best for software teams that want issues turned into pull requests

What it does. GitHub's cloud agent, called Copilot coding agent until April 2026, "can work independently in the background to complete tasks, just like a human developer": it plans, changes code on a branch and opens a pull request, and you can approve its plan first.

Needs you to run. A paid Copilot plan and repositories on GitHub.com.

Price. Copilot Business costs $19 per seat per month with 1,900 AI credits, and Enterprise $39 with 3,900; a credit is $0.01, plus GitHub Actions minutes. We found no trial of Copilot Business.

For what it can't reach, see our GitHub AI agents page.

When no product reaches the job

A custom agent: best for a job that crosses systems no product connects, or runs on rules only your business has

What it does. Meronym, the publisher of this guide, builds it. We scope one workflow in two weeks, write down what the agent may do alone and what it has to escalate, and give you a fixed quote and your engineer's name before any build; write actions wait behind an approval gate until the agent passes an evaluation on your own reviewed cases. This guide does not cite a published client result for our entry.

Needs you to run. Nothing new: it's built on the systems you have.

Price. After the two-week scope, $4k–$12k and credited against the build, a pilot on one workflow is $8k–$35k over four to six weeks, and a production agent $16k–$82k over eight to fourteen. The running cost is quoted apart.

Every entry in one table

Screenshot this one. It has a row for every entry above; the three developer platforms are in their own table. "Can it change records" means it can write to a system, not only answer; the entries say what waits for a person.

What AI agents for business cost at your volume

Every vendor counts something different, so the same month of work bills differently from one product to the next.

Take a support team with 1,000 customer conversations a month, 500 of which the agent resolves with no person involved. The 50% is a round number for the sum, not any vendor's claim; put in your own.

Per-resolution bills rise and fall with the share the agent resolves; per-conversation bills don't. Seats work differently again. For ten people on annual billing, Microsoft 365 Copilot's Business add-on costs $210 a month ($180 in the first year on orders by 31 December 2026) on top of the Microsoft 365 plan, and the enterprise seat $300; Gemini Enterprise's Business edition costs $210 (Google doesn't state the term), and ChatGPT Business and Claude Team $200 each. Lindy's Plus plan is $299.90 a month for ten users. Heavier use is metered on top of each seat, and that's where the bill moves.

At some volume the unit price costs more than owning the agent. The next section turns that into a rule.

When a bought agent stops being enough

Buy first. For most jobs the right first agent is the one in the system where the job already happens, and our pages on the AI customer service agent, the AI sales agent and AI agents for marketing say the same. Building starts to pay under three conditions:

1. The job needs a system no product reaches. List every system the agent must read or write. If the product you're considering has no connector for one of them, that part of the job remains with your team or needs a custom integration.
2. An incorrect answer or action could create a regulated event. If one reply could reach a regulator or a court, you need your own rules, logs and sign-off; our page on AI agents for compliance covers what that takes.
3. Your volume makes the unit price cost more than owning the agent. Multiply your monthly resolutions by the unit price and by twelve. At 5,000 resolutions a month, Fin's $0.99 comes to $59,400 a year and a $2.00 resolution to $120,000. Compare that with a build, $16k–$82k for a production agent of ours, with the running cost quoted apart.

Two documented failures show why the write matters more than the model. In Moffatt v. Air Canada, decided on 14 February 2024, British Columbia's Civil Resolution Tribunal held the airline liable for its website chatbot's wrong answer about bereavement fares. The tribunal read Air Canada's defence as suggesting that the chatbot was "a separate legal entity that is responsible for its own actions", called that "a remarkable submission", and ordered the airline to pay C$812.02 in all. Whatever agent you buy, you own what it tells your customers.

In July 2025, Replit's coding agent deleted a production database during a code freeze on SaaStr founder Jason Lemkin's project. Replit's chief executive called it "Unacceptable and should never be possible" and announced automatic separation of development and production databases. An agent with write access will one day use it wrongly, which is why the approval questions matter more than the price. Gartner predicted in May 2026 that by 2027, 40% of enterprises will demote or decommission autonomous agents because of governance gaps found only after incidents in production.

Two kinds of job don't need an agent at all. A job whose steps never change, however many there are, is a fixed workflow: work for business process automation services. And help with one person's inbox and documents is what the office seats above already cover.

If your job meets one of the three conditions, that's the work behind our AI agent development services: the custom agent in the list, scoped in two weeks to a fixed quote.

Your next step

Pick the job, then the agent that lives where that job already happens, and try it on last month's real cases with a person approving every write. Fifteen of the 24 product entries have a free plan or a trial, among them Fin and Zendesk for 14 days, QuickBooks Online for 30, Botpress's free plan and n8n's Community Edition on your own servers.

If the job meets one of the three conditions above, send it to us instead: use the Scope an agent button at the end of this page.