Every software vendor now sells an AI agent, and most of the advice about AI agents for business is a list of products. That skips the questions that decide whether an agent makes or loses money: which job it should do, what it will really cost to run, and who is accountable when it gets something wrong.
This guide answers those questions. You get a plain explanation of what business agents are, where the 2026 evidence says they pay off, a five-question test for any use case, a tested ROI calculator you can run with your own numbers, a checklist for spotting agent washing, and a 90-day plan for a first rollout. If you want product names and prices, our roundup of the best AI agents covers the tools; this piece is about strategy.
What are AI agents for business?
An AI agent is a program that uses a large language model to work toward a goal: it decides what to do next, calls tools such as your CRM, ticketing system or ERP, checks the result and continues until the task is done or it needs a person. A chatbot answers; an agent acts. Our guide to agentic AI vs generative AI explains that difference in depth.
Anthropic's engineers draw a useful line between two designs. In their words, "Workflows are systems where LLMs and tools are orchestrated through predefined code paths," (opens in a new tab) while agents "dynamically direct their own processes and tool usage." In practice, many business deployments sit between the two: a fixed process with one or two AI steps that make decisions inside clear limits. We cover those patterns in AI agentic workflows.
Agentic systems often trade latency and cost for better task performance
That trade is the heart of every business case. An agent is slower and more expensive per task than a fixed script, so it has to earn its keep on work where flexibility matters.
Where enterprise AI agents stand in 2026
Adoption is real but uneven. McKinsey's State of AI 2026 report (opens in a new tab), based on a survey of 1,719 participants run from May 4 to June 8, 2026, found that 40 percent of respondents at organizations with more than $1 billion in revenue report scaling AI agents, up from 27 percent a year earlier. Among smaller organizations the figure held flat at 22 percent.
Value is lagging behind activity. In the same survey, 37 percent of respondents attribute at least some EBIT impact to AI use, about the same as the year before. The group McKinsey calls AI high performers, about 6 percent of respondents, behave differently: nearly three-quarters of them report fundamentally redesigning workflows because of AI, against one-quarter of everyone else, and they are more than three times as likely to be scaling agents in most business functions.
Analysts also expect a shakeout. In June 2025, Gartner predicted that over 40 percent of agentic AI projects will be canceled by the end of 2027 (opens in a new tab), citing "escalating costs, unclear business value or inadequate risk controls," and advised pursuing agentic AI only where it delivers clear value or ROI. Finance teams are feeling the gap: in a Gartner survey of 204 finance leaders (opens in a new tab) in March 2026, 63 percent said AI implementation in 2025 was slower than they expected.
Where AI agents pay off: five business functions
Agents work best where tasks are frequent, follow rules you can write down, rely on information the agent can reach, and produce results that are easy to check. McKinsey's respondents most often report scaling agents in IT, knowledge management and software engineering. Here is how those traits play out in the functions most businesses look at first. For named company deployments in each area, see our agentic AI examples.
Customer support
Support is the classic first use case. Anthropic notes that support interactions (opens in a new tab) follow a conversation flow while needing access to external information and actions, which is exactly what agents are built for. Gartner predicts that by 2029 agentic AI will autonomously resolve 80 percent of common customer service issues (opens in a new tab) without human intervention, leading to a 30 percent reduction in operational costs. Treat that as a forecast, not a benchmark.
The strongest measured evidence so far comes from AI that helps people rather than replaces them. A study of 5,179 customer support agents (opens in a new tab), published in the Quarterly Journal of Economics in 2025, found that access to a generative AI assistant raised issues resolved per hour by 14 percent on average and by 34 percent for novice and low-skilled workers, with minimal impact on the most experienced staff. Good first tasks: order status, password resets, refunds within a set limit, and drafting replies for a person to approve.
Sales operations
In sales, agents earn their place on the administrative work around selling rather than the selling itself: researching an account before a call, enriching and deduplicating CRM records, logging call notes, routing inbound leads and drafting follow-up emails. These tasks are repetitive and the output is easy to check. Keep a person on anything that commits the company to a price, a discount or contract terms.
AI agents in finance
Finance has plenty of rules-bound, high-volume work: matching invoices to purchase orders, chasing missing receipts, flagging expense claims that break policy, and preparing reconciliation workpapers for review. The audit trail matters as much as the automation, so pick tools that log every action the agent takes. Gartner's finance analysts warn CFOs not to measure success by how many tools are live; as Director Analyst Marco Steecker put it, "They must not mistake activity for impact." (opens in a new tab) In that survey, 66 percent of finance organizations using AI named efficiency and productivity as a top benefit.
AI agents for marketing
Marketing teams use agents to turn one brief into channel variants, monitor campaign performance and summarize what changed, assemble competitor research, and keep product listings or the content calendar up to date. The risk is brand and legal: an agent that publishes directly can put a false claim in front of customers. A sensible default is that agents draft and people publish, at least until you have weeks of reviewed output.
IT and internal knowledge
These are the areas where McKinsey's respondents most often report scaling agents. Typical jobs are triaging help desk tickets, resetting access, answering policy questions from internal documents and, in engineering, writing and testing code. Coding suits agents because, as Anthropic points out, code can be checked with automated tests. Internal knowledge agents usually rely on retrieval-augmented generation to answer from your own documents.
| Function | Good first tasks | Keep a human on | Measure |
|---|---|---|---|
| Customer support | Order status, resets, refunds under a set limit, drafted replies | Complaints, exceptions, anything above the refund limit | Resolution rate, reopen rate, CSAT, cost per resolution |
| Sales operations | Account research, CRM updates, lead routing, follow-up drafts | Pricing, discounts, contract terms | Hours saved per rep, data quality, response time |
| Finance | Invoice matching, receipt chasing, expense policy checks | Payments, journal entries, write-offs | Cycle time, exceptions caught, error rate |
| Marketing | Content variants, campaign summaries, research briefs | Anything published, claims, regulated copy | Output per person, revision rate, time to launch |
| IT and knowledge | Ticket triage, access requests, policy Q&A, code with tests | Production changes, permissions, security exceptions | Time to resolve, deflection rate, test pass rate |
A five-question test for any agent use case
Before you buy anything, run each candidate task through these questions. A task that fails two or more is better handled by a fixed workflow, a simple script or a person.
- Is it frequent? Hundreds or thousands of times a month justifies setup and monitoring. A task done twice a week rarely does.
- Can you write the rules down? If your best employee cannot explain how they decide, the agent will not decide well either.
- Can the agent reach what it needs? It needs access to the systems and data involved, through connectors or a standard such as the Model Context Protocol.
- Can you check the result cheaply? Tests, totals that must match, or a quick human glance. Unverifiable output is where errors pile up.
- Is a mistake recoverable? Start where an error costs a reply or a correction, not a payment, a legal commitment or a customer.
How to estimate AI agent ROI before you buy
Vendor ROI claims usually count the time saved and leave out the costs. A fair estimate subtracts four things from the labor you save: the cost of every attempt, including failed ones; the time people spend reviewing the agent's work; platform or license fees; and the one-off setup cost, which decides your payback period.
The script below does that arithmetic. We ran it with Python 3.13; it uses only the standard library. The inputs are a hypothetical example, a support team handling 6,000 tickets a month at 8 minutes each with a loaded labor cost of $40 an hour, so replace every number with your own. For real per-task prices, check your vendor's pricing page; our best AI agents guide lists several.
# Rough monthly ROI for one AI agent use case.
# Every input is an assumption: replace each one with your own numbers.
def agent_roi(tasks_per_month, minutes_per_task, hourly_cost,
completion_rate, cost_per_attempt, review_share,
review_minutes, platform_fee, setup_cost):
completed = tasks_per_month * completion_rate # tasks the agent finishes
saving = completed * minutes_per_task / 60 * hourly_cost
run_cost = tasks_per_month * cost_per_attempt # you pay for failed attempts too
review_cost = completed * review_share * review_minutes / 60 * hourly_cost
total_cost = run_cost + review_cost + platform_fee
net = saving - total_cost
payback_months = setup_cost / net if net > 0 else None
return saving, total_cost, net, payback_months
for rate in (0.35, 0.15):
saving, cost, net, payback = agent_roi(
tasks_per_month=6000, minutes_per_task=8, hourly_cost=40,
completion_rate=rate, cost_per_attempt=0.50, review_share=0.10,
review_minutes=3, platform_fee=1500, setup_cost=30000)
months = f"{payback:.1f} months" if payback else "never"
print(f"{rate:.0%} completed: saves ${saving:,.0f}, costs ${cost:,.0f}, "
f"net ${net:,.0f}/month, payback {months}")35% completed: saves $11,200, costs $4,920, net $6,280/month, payback 4.8 months
15% completed: saves $4,800, costs $4,680, net $120/month, payback 250.0 monthsThe second line is the lesson. With identical prices, an agent that finishes 15 percent of tickets instead of 35 percent goes from paying back in under five months to barely covering its own running costs. Completion rate is the number that decides your business case, and you only learn it by piloting on your own data.
| Item | 35% completed | 15% completed |
|---|---|---|
| Tickets the agent finishes | 2,100 | 900 |
| Labor saved | $11,200 | $4,800 |
| Agent attempts (6,000 at $0.50) | $3,000 | $3,000 |
| Human review of 10% of finished tickets | $420 | $180 |
| Platform fee | $1,500 | $1,500 |
| Net benefit | $6,280 | $120 |
| Payback on $30,000 setup | 4.8 months | 250 months |
Two more things belong in the business case even though they are harder to price: the cost of the errors that slip through review, and whether the saved hours turn into real savings or better service. Hours saved that nobody redeploys do not show up in the P&L.
How to spot agent washing
Gartner uses the term agent washing for vendors that rebrand existing products, "such as AI assistants, robotic process automation (RPA) and chatbots," (opens in a new tab) as agentic AI, and in 2025 it estimated that only about 130 of the thousands of agentic AI vendors are real. In May 2026 it warned again that agent washing relabels conventional automation as agentic (opens in a new tab), raising the risk of misaligned investments and lock-in. Ask any vendor these questions:
- What does the agent decide by itself? If every step is fixed in advance, it is a workflow. That may be fine, but price it like one.
- What happens when it is unsure? Look for confidence thresholds, escalation to a named person and a record of the handoff.
- Which systems can it write to, and with what permissions? You want scoped, per-agent credentials, not a shared admin account.
- Can we see the full action log? Every tool call, input and output should be reviewable after the fact.
- What is the completion rate on tasks like ours? Ask for a pilot on your own data rather than a demo on theirs.
- How is it priced when it fails? Per conversation, per action and per credit pricing all charge for unsuccessful attempts differently.
Governance: set the rules before you roll out
Agent counts can grow fast once teams can build their own. Gartner predicts that by 2028 an average global Fortune 500 enterprise will have over 150,000 agents in use (opens in a new tab), up from fewer than 15 in 2025, while only 13 percent of organizations think they have the right AI agent governance in place. Its six recommended steps are a good skeleton for a policy of any size:
- Establish agent governance and policies: who may build and share agents, and which connectors are allowed.
- Build a central inventory of every agent, including ones teams built without telling IT.
- Define each agent's identity, permissions and life cycle, and retire agents nobody uses.
- Govern the information agents can reach, and keep it current so they do not overshare or act on stale data.
- Monitor behavior and fix agents that go beyond their scope.
- Train people and build a culture of responsible use.
McKinsey's high performers add one more habit: they decide in advance how and when model outputs need a human in the loop. In practice, give each agent the fewest permissions it needs, require approval before anything is paid, sent externally or deleted, and treat any agent that reads email, documents or web pages as exposed to prompt injection, the hidden instructions we describe in our best AI agents guide.
A 90-day plan for your first business agent
This is the sequence we recommend for a first deployment. Adjust the timing to your size; the order matters more than the dates.
- Weeks 1 to 2: pick one task. Use the five-question test, write down today's baseline (volume, time per task, error rate, cost) and name one accountable business owner.
- Weeks 3 to 4: choose the tool. Check whether software you already pay for has an agent feature before buying a new platform. Run the vendor questions above.
- Weeks 5 to 8: pilot in shadow mode. The agent drafts or proposes; people approve every action. Track completion rate, error types and review time.
- Weeks 9 to 10: rerun the ROI numbers with real completion and review figures instead of the vendor's.
- Weeks 11 to 12: go live with limits. Let the agent act alone on the lowest-risk slice, keep human approval for the rest, and review the action log weekly.
Should you build instead of buy? Build when the task touches proprietary systems no vendor connects to, or when the agent is part of your product. Our tutorial on how to build an AI agent shows the core loop in tested Python. For everything else, buying or switching on an agent feature in your existing software is usually faster and cheaper to maintain.
The bottom line
AI agents for business pay off when they take on frequent, rules-bound tasks with checkable results, inside a process you have redesigned around them. The 2026 data shows adoption climbing and value lagging, so start narrow, measure completion rate on your own data, price failed attempts into your ROI, and set permissions and human checkpoints before you scale. For more guides on agents, models and tools, browse our AI hub.