AI agents for business: the jobs they do well, by department

In this guide

• What are AI agents for business?
• AI agent use cases by department: the master table
• AI agent examples, department by department: customer service, sales, marketing, finance and admin, operations, HR and recruiting, research and strategy, IT and engineering
• Single-purpose agents or a platform where agents work together?
• How do you start using AI agents in a business?
• How much do AI agents for business cost?
• Common mistakes
• How Olano helps
• FAQ

What are AI agents for business?

An AI agent is a language model given tools, memory and a goal: it reads, decides on a step, acts through a connected system, checks the result and carries on. Unlike rule-based AI business automation, which follows a path drawn in advance, an agent chooses its next step, so it copes with messy work and needs a boundary. See what AI agents are and what running one takes for the mechanics, and AI automation vs AI agents for when a fixed workflow is better.

Each job below has a starting autonomy level. It is our recommendation, not a feature of any product:

• Alone. It reads, sorts, summarises or drafts internally. Nothing leaves the business.
• Approve. It prepares a reply, a post, a record change or an invitation, and a person approves each one. "Approve, then alone" means routine cases can run unattended once the agent has a clean record.
• Person. Money moves, data is deleted or the decision is about a person: the agent prepares, a person decides and acts.

AI agent examples, department by department

Four jobs per team, what runs alone, what waits, and the first agent we would start with. Systems for each job are in the table above.

Customer service

• Answer repeat questions on prices, hours, delivery and returns, from your own documents.
• Triage and route every message by topic and urgency, with a summary for the person it goes to.
• Order and booking status, looked up in your system instead of "we'll check".
• Complaints and refund requests: the facts gathered, the reply and a resolution drafted.

Alone: triage, lookups and drafts. Waits for a person: replies until the agent has a track record, and every refund or exception to policy. First agent: a front desk on the channel most enquiries use, every reply approved for two weeks. See automating enquiries with human approval and WhatsApp AI customer service.

Sales

• Lead capture and qualification: every enquiry logged in the CRM and scored against your criteria.
• Pre-call research briefs on the company, the people and what changed since the last conversation.
• Follow-ups on quiet deals, flagged past your threshold and drafted.
• Proposal first drafts from your templates, price list and call notes.

Alone: research, scoring and flags. Waits for a person: anything sent to a prospect, and every price, discount or contract term. First agent: CRM upkeep: enquiries captured, stale deals flagged, follow-ups drafted for the salesperson. See the AI CRM and sales assistant and missed lead recovery.

Marketing

• Social posts and newsletters drafted in your voice from your own material.
• Campaign reporting: what ran, what it cost and what moved, every week.
• Reviews and mentions watched, with the ones needing an answer flagged and a reply drafted.
• SEO research and content briefs a writer can use.

Alone: research, reporting, monitoring and drafts. Waits for a person: every post, newsletter and public reply, because publishing cannot be taken back. First agent: a content drafter that queues everything for approval. See the AI social media agent and the approval-first social workflow.

Finance and admin

• Invoice and bill intake: supplier, amount and due date extracted from emailed invoices and matched to purchase orders.
• Overdue invoice reminders drafted on a schedule, firmer at each step.
• Month-end reconciliation prep: bank exports compared with the ledger, mismatches listed.
• Budget against actual, with every variance over your threshold flagged.

Alone: extraction, matching and reports. Waits for a person: ledger entries, customer reminders, and every payment, which the agent should have no way to make. First agent: invoice chasing with read access to the accounting system. In Olano, financial actions wait for approval even in Trusted, the most permissive oversight mode.

Operations

• Shared inbox triage: labels, summaries, priorities and draft replies.
• Moving data between systems: an enquiry becomes a CRM record, a calendar entry and an invoice line without retyping.
• Supplier and order chasing for late deliveries and missing documents.
• Weekly operations report from the systems holding the numbers.

Alone: triage, summaries and reports. Waits for a person: messages to suppliers and customers, and record changes until the mapping has proved itself. First agent: inbox triage with drafts. See AI for operations teams, the inbox triage and follow-ups playbook and the AI inbox assistant.

HR and recruiting

• Application screening summaries against your written criteria, so a person shortlists faster.
• Interview scheduling across the panel's and the candidate's calendars.
• Policy questions on leave, claims and benefits, answered from your handbook with the section cited.
• Onboarding checklists for each new joiner, tracked until done.

Alone: summaries, checklists and, once tested, internal answers. Waits for a person: every message to a candidate, and every decision about who progresses, is hired or leaves. The EU AI Act lists systems that analyse and filter job applications as high-risk (Annex III, point 4); see our EU AI Act guide. First agent: an internal handbook assistant, low-risk and easy to check.

Research and strategy

• Competitor monitoring: pricing pages, launches and hiring, in a weekly brief with sources.
• Industry news digest, with a line on why each item matters to you.
• Supplier and product comparisons against your requirements, with prices and the date checked.
• Tender and RFP monitoring, with a first-pass fit summary for each notice.

Alone: nearly all of it, which makes research the safest place to start. Waits for a person: any decision to act on the findings. Treat what the agent reads on the web as untrusted input (see prompt injection and AI agents). First agent: a weekly competitor brief. See competitor monitoring with AI agents and the AI market research agent.

IT and engineering

• Pull request review notes flagging risky changes and missing tests.
• Bug report triage: labels, duplicates spotted, missing details requested.
• Internal IT questions answered from your runbooks.
• Alert and incident summaries sent to the person on call.

Alone: summaries and notifications. Waits for a person: review comments at first, and always merges, deployments, deletions and access changes. First agent: bug report triage, where every label it applies is visible and easy to correct.

In their own words (checked 20 October 2026): Salesforce says Agentforce launches agents "for sales, service, HR, and more"; HubSpot's Agent Hub agents are "built on your CRM data"; Zendesk's AI agents run in Zendesk or an existing service platform; Intercom's Fin covers service, sales, ecommerce and success. ChatGPT agent works on its own virtual computer and asks before actions of consequence, with 40 agent requests a month on Plus; Gemini for business works under each person's existing permissions; Claude takes tasks on a schedule.

The case for a platform is the handoff. Google's own page names the problem: "Today almost every tool has an agent. Each one knows a sliver of your business and none of them talk to each other." When a WhatsApp enquiry must become a CRM record, a booking and an invoice, agents that delegate to each other under one set of approvals do the work of several products that each see part of it. The cost is that you design the team; see how to choose an AI agent platform, Olano vs ChatGPT for business and Olano vs Lindy.

How do you start using AI agents in a business?

One job, one channel, approvals on, and a number written down first. It works the same for AI agents for small business teams and for one department of a larger company:

1. Pick one job from the table, the one costing most today. Research is the safest start; customer enquiries usually pay back fastest.
2. Write down today's number: volume, hours, response time, enquiries missed (see how to measure AI ROI).
3. Connect one channel and only the systems the job needs, read-only wherever reading is enough.
4. Turn approvals on and review every output for two weeks, correcting the procedure as you go.
5. Widen only what has earned it. At about ninety days, compare with the before number, then add a second job or stop.

Still choosing? Read where to start with AI in your business, or take the free AI readiness assessment, which suggests a first agent.

How much do AI agents for business cost?

Four lines: the platform, model usage, one-off setup per job, and the time someone spends reviewing the work. Model usage follows the shape of the job: an agent that reads twenty web pages per question costs far more than one that sorts email, so estimate per job, not per agent. What an AI agent actually costs to run builds the estimate.

Pricing shapes differ too. Salesforce prices Agentforce through Flex Credits, Resolutions or per-user licensing, and Microsoft sells Copilot Studio in packs of 25,000 Copilot Credits at US$200 a month (both checked 20 October 2026). Per seat, per usage or per deployment compares the shapes. Olano plans start at US$25 a month with included credits and no per-seat pricing; see pricing.

Common mistakes

• Starting with the hardest job. Pricing, hiring and key accounts depend on judgement. Start where work repeats and errors are cheap to spot.
• One agent for every department. Its instructions start to contradict each other; multi-agent vs single agent gives the signals for splitting.
• Twenty agents in week one. Each needs maintaining, and most businesses need fewer than they think.
• Approvals nobody reads. An unchecked queue stalls the work it holds. Name the person who checks it.
• More access than the job needs. An agent that drafts reminders does not need to move money.

How Olano helps

Olano runs a team of agents across departments, on infrastructure we operate.

• A starting point for every team. 180+ specialist agents ship across 19 starter teams, from customer service, sales and marketing to finance, operations, engineering, research, HR and legal. Install the two or three that match your jobs.
• Your channels and systems. 15+ messaging channels, including WhatsApp, Telegram, Slack, email, SMS, Signal, WeChat and Lark, and 450+ connector tools.
• A real team. Agents sit in an org chart, hand work to their manager, reports and approved peers, and share an issue tracker with your people; each keeps its own memory in files you can read and correct.
• Approvals per agent. Trust levels 0 to 4 and an oversight mode: Supervised holds every write to an external system, high-risk operations need a person at every level, and every decision is recorded. See security and approvals and trust.
• Improvement you control. When you switch Cortex on, it runs Fully Autonomous by default; rewrites of an agent's own instructions, outbound sends and irreversible actions always wait for a human.
• Predictable cost. Plans from US$25 a month, unlimited agents and people, your own model keys on every plan, and per-agent budgets that can stop at a cap.

The limits: Olano does not answer phone calls (voice arrives only as voice notes on WhatsApp, Telegram and Signal), and a catalogue agent is a starting point, so you still choose the job and review the work. Building your agents is not in the plan price: you do it in the dashboard, or Olano Studio does it with you. More on the platform page.