Most articles on AI agents for business are lists of tools. The harder question is which work to hand over. In a benchmark that set agents loose in a simulated small company, the best one finished 30% of its tasks on its own. That's useful if you pick the right 30%, and expensive if you don't. This guide is about picking it.
What are AI agents for business?
AI agents for business are AI systems that carry out a piece of work by deciding their own next steps and using company tools, such as email, a CRM or a ticket queue, rather than only answering a person who then acts. An assistant drafts the reply. An agent reads the request, looks up the order, writes the reply, sends it and closes the ticket.
This guide is for the people choosing where a company of 20 to 1,000 people should use agents first: COOs, operations and IT leads, and finance and HR leads who've been pitched one. It covers what an agent is, what the evidence says agents can finish, how to choose the first task, what the main platforms charge and how to start. It isn't a tool roundup, and we don't sell or resell any AI tools.
Assistant, workflow or agent: which do you need?
Anthropic, the AI lab behind Claude, draws the line clearly in its guide for builders. "Workflows are systems where LLMs and tools are orchestrated through predefined code paths." "Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage." Its advice to its own customers is to start with "the simplest solution possible, and only increasing complexity when needed", because "agentic systems often trade latency and cost for better task performance."
| Assistant | Workflow | Agent | |
|---|---|---|---|
| Who decides the next step | A person | The steps you wrote | The AI |
| Good for | Drafting, summarising, answering | The same steps every time | Tasks where the steps vary |
| What it touches | What the person pastes in | The systems you wired up | Whatever its access allows |
| Cost per task | Low and predictable | Low and predictable | Higher and variable |
| What can go wrong | A bad answer someone reads | A step fails | It acts on a bad decision |
A lot of what's sold as an agent is a workflow with an AI step in it, and that's often what you want. If you can write the steps down, write them down. An agent earns its higher cost where the steps change from one case to the next: a service request that might need a refund, a replacement or a question back to the customer.
What can AI agents actually finish on their own?
Less than the demos suggest, and more each year.
The most useful test for a business buyer is TheAgentCompany, a research benchmark. It builds a simulated small software company where agents work the way staff do: browsing the web, writing code, running programs and messaging coworkers. In its latest version, "the most competitive agent can complete 30% of tasks autonomously." The authors found that "a good portion of simpler tasks could be solved autonomously", while "more difficult long-horizon tasks are still beyond the reach of current systems."
The trend is moving fast. METR, a research group that measures what AI systems can do, found in March 2025 that the length of tasks agents "can complete autonomously with 50% reliability has been doubling approximately every 7 months" over six years. Read that as a direction, not a forecast for your company. Tasks an agent fails today may be within reach next year, and tasks it manages today are still only finished about half the time at the edge of its range.
Large companies are moving anyway. In KPMG's survey of 314 US business leaders at companies with $1 billion or more in revenue, fielded in July and August 2026, 62% said they were building, deploying or developing agents, up from 53% the quarter before. Microsoft's 2025 Work Trend Index, a survey of 31,000 workers in 31 markets, found 82% of leaders "confident that they'll use digital labor to expand workforce capacity" within 12 to 18 months. Neither survey shows that the agents pay off. They show the vendors will keep putting agents in front of you.
Where do AI agents pay off in a business?
Agents pay off on short, frequent tasks whose result someone can check quickly and undo easily, using only the access that one task needs. Anthropic names two areas where agents already work well, and both fit that pattern. In customer support, "success can be clearly measured through user-defined resolutions." In coding, "code solutions are verifiable through automated tests."
Those are four tests, and every candidate task should pass all of them before an agent gets it.
- It happens often. An agent costs time to set up and money to run. A task done twice a month rarely repays either. A task done 200 times a week can.
- It's quick to check. Someone has to be able to tell, in seconds, whether the agent got it right: the ticket was resolved, the invoice matched, the answer cited the right policy. If checking takes as long as doing, you've added work.
- It's easy to undo. A wrong draft costs nothing. A wrong payment, a deleted record or a rejected job applicant costs a great deal. Keep agents to actions you can reverse until they've earned more.
- It needs narrow access. The fewer systems and records it can reach, the less a mistake can spread. Read access to one queue is a different risk from write access to the finance system.
What good first agents look like
In a company of a few hundred people, these usually pass:
- IT and operations: sort incoming requests, answer the routine ones from your own documentation, and pass the rest to the right person with a summary.
- Finance: match supplier invoices against purchase orders and flag mismatches for a person to resolve, without paying anything.
- HR: answer staff questions about leave, benefits and expenses from the handbook, with a link to the section it used.
- Sales operations: prepare a brief before each customer meeting from the CRM and recent emails, and suggest CRM updates for the account owner to accept.
And these usually don't, at least at first: approving refunds or payments, running the month-end close, and anything that ranks, scores or rejects people. Decisions about people also bring in the law. Colorado's new AI law, for example, requires notice and an explanation when software materially influences a hiring decision (see our guide to the Colorado AI law).
When an agent goes wrong, it acts
The four tests exist because an agent's mistakes aren't just wrong answers. In July 2025, according to the OECD's AI incident monitor, Replit's AI coding assistant "ignored explicit instructions not to alter code, deleted a live production database, and fabricated data to cover up the error." It had been told not to change anything. Nothing in its access stopped it. That's the difference between an agent you've told to be careful and an agent that can only do one safe thing.
Anthropic makes the same point from the builder's side: with an agent "you must have some level of trust in its decision-making", and it recommends "extensive testing in sandboxed environments, along with the appropriate guardrails."
Where will your first AI agents come from?
For most companies of 20 to 1,000 people, from software you already pay for. You rarely need to build one from scratch, and the governance you already have for those vendors carries over.
- Microsoft 365: Copilot Studio builds agents for Microsoft 365. Microsoft says the Copilot Studio included with Microsoft 365 Copilot lets "the licensed users build and use internal agents within Microsoft 365" at no additional cost.
- Salesforce: Agentforce adds agents to Salesforce, priced per conversation, per action or per user (see the costs below).
- Google Workspace: Google's new coworker agents, announced on 8 October 2026, each get "its own Workspace account, including an email address, calendar, Drive, and presence in your company directory." The Gemini agent is "now in preview for Financial Services and Legal, and coming soon to Government, Healthcare, and Retail."
Agent features can also arrive inside tools you didn't buy as agents, such as a help desk or a CRM. The first job is to know which ones you already have. Our Shadow AI Audit builds that list, including AI features vendors have turned on for you.
How much do AI agents cost?
Agents are usually billed by use, not by seat, so the cost depends on how often the agent runs and how much work each run takes. That's the real change from assistants, which most companies buy at a fixed price per person. These are the published rates as of October 2026.
Microsoft Copilot Studio
Copilot Studio meters agents in Copilot Credits. Microsoft sells them at "$200.00/pack/month for 25,000 Copilot Credits", or through a pay-as-you-go meter that "requires no up-front license commitment". Each thing an agent does uses a set number of credits:
| What the agent does | Copilot Credits |
|---|---|
| A classic (pre-written) answer | 1 |
| A generative answer | 2 |
| An agent action | 5 |
| Grounding in your tenant's Microsoft Graph data | 10 |
| Agent flow actions, per 100 | 13 |
Microsoft's own example is a customer support agent that gives four pre-written answers and two generative answers per customer, for 900 customers a day: 7,200 credits a day. At the pack price, that's about $57.60 a day.
Two details matter for budgeting. Employee-facing agents are "No charge" when the user has a Microsoft 365 Copilot licence and the agent works under that user's identity, so an internal agent for licensed staff may cost nothing extra. And if prepaid capacity runs out, enforcement starts at 125% of it, and "Custom agents are disabled." Set a monthly limit for each agent in the Power Platform admin center so one busy agent can't switch off the others.
Salesforce Agentforce
Salesforce publishes three ways to pay: "$ 2" per conversation; Flex Credits at "$ 500" per 100,000 credits, where "Agentforce actions are 20 Flex Credits" (about $0.10 an action); or per-user add-ons from "$ 125" a user a month. The per-action meter suits agents that do a few things often. The per-conversation price suits customer-facing chat.
Google hasn't published a price for the Gemini agent or coworker agents. It does let you set "a hard limit on a project's AI spend", after which "that project's agent pauses."
Whatever the platform, write the running cost into the case before you start, and cap it.
An agent is a build, so give it a case and a proof
Most companies treat a new agent like a new feature: someone switches it on and tells the team. It's closer to a new hire, or a new system, and it should go through the same gate as anything else you spend money on.
In our framework, that gate is Build. Every build gets a one-page case: the value it frees at the bottleneck, the full running cost, an owner and a backup, a stop-loss and kill criteria. Then it's proved in two stages. First in a sandbox that can't reach live systems. Then in a limited live trial with real users, in the systems they already use. An agent that can't pass the first stage never gets live access. One that can't pass the second gets switched off without anyone having to argue about it, because the kill criteria were written before it started.
The rules for agents add the rest. Each agent gets its own login with limited reach, never a borrowed human account, and anyone can stop it. Our free lesson on rules for agents shows how to set them, and two-stage proof covers the trial. For the six decisions to make before agents spread across the company, see our guide to AI agent governance.
The large companies in KPMG's survey are still catching up on this. 49% had defined high-risk uses where agents may not make decisions on their own, and only 30% were building controls into agents alongside monitoring and evaluation, down from 43% two quarters earlier.
How to start with AI agents in your business
01 · List the agents you already have
Find the agent features already switched on in your Microsoft, Google, Salesforce and other subscriptions, and who is using them. This gives you the starting list, and it often shows an agent already doing work nobody approved.
02 · Find the bottleneck before you pick the tool
Ask where work waits longest, not where an agent would be impressive. Our lesson on naming the bottleneck shows how to find one and put a value on it. The value is what the agent has to beat.
03 · Pick one task that passes all four tests
Often, quick to check, easy to undo, narrow access. If no task at the bottleneck passes, the bottleneck needs a workflow or a redesign first, not an agent.
04 · Write the one-page case, with a cost cap
Name the owner and backup, the access it needs and nothing more, the monthly spending limit, and the result that would make you switch it off. Use the vendor's published rates and your real volumes to estimate the cost.
05 · Prove it in two stages, then review it
Run it in a sandbox, then in a limited trial with the people who do the work today, with a person checking every result at first. Set a review date. At the review, keep it, fix it or retire it on the evidence, and only then consider the next task or wider access.
Deploying an agent is one case of AI deployment in general: the work, the owner and the proof matter more than the tool. If you'd like help choosing the first task and running the trial, our AI implementation team does exactly that.
What's still open
- Pricing. Agent pricing is still settling. Microsoft and Salesforce both bill agents through usage meters, and Google hasn't published prices for its agents. Expect the rates in this article to change.
- Capability. The benchmarks above are moving fast. A task that fails the "quick to check" test because the agent gets it wrong too often may pass within a year.
- Rules. State laws on automated decisions, such as Colorado's from January 2027, apply to agents that influence decisions about people. More are coming.
We'll update this article when the major platforms change their agent pricing or when a new large study changes what agents can be trusted to finish. The date at the top of the page shows when we last checked.
Sources
- TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks, arXiv, September 2025 (v3): the simulated company, the 30% completion rate and the finding on simpler and long-horizon tasks.
- Building effective agents, Anthropic, December 2024: the definitions of workflows and agents, the advice to start simple, and the customer support and coding examples.
- Measuring AI Ability to Complete Long Tasks, METR, March 2025: the seven-month doubling of the length of tasks agents complete with 50% reliability.
- Billing rates and management, Microsoft Copilot Studio documentation, October 2026: the Copilot Credit rates, the customer support example, the Microsoft 365 Copilot inclusion and overage enforcement.
- Copilot Studio pricing, Microsoft, October 2026: the price of a Copilot Credit pack and the pay-as-you-go meter.
- Agentforce pricing, Salesforce, October 2026: per-conversation, Flex Credit and per-user prices.
- Welcome to Gemini at Work 2026, Google Cloud, October 2026: coworker agents' accounts, the preview, and spend caps.
- KPMG AI Quarterly Pulse Survey, Q3 2026, KPMG, September 2026: the share of large US companies building or deploying agents, and their controls.
- 2025: The year the Frontier Firm is born, Microsoft Work Trend Index, April 2025: the survey of 31,000 workers and leaders' plans for agents.
- Replit AI coding tool deletes live production database and fakes data, OECD AI Incidents Monitor, July 2025 (secondary): the incident summary.