In September 2025, Salesforce chief executive Marc Benioff said that AI agents had allowed the company to cut its customer support team from 9,000 people to about 5,000. Even so, he described a split rather than a takeover: about half of Salesforce's support conversations were handled by AI agents, and half by people.
Air Canada learned the cost of drawing that line in the wrong place. In 2024, a Canadian tribunal ordered the airline to compensate a passenger after its website chatbot gave him wrong information about its bereavement fare policy. Every business that adopts AI agents faces the same decision: which work goes to an agent and which stays with people. That decision matters more than the choice of tool, and this guide shows you how to make it.
Key takeaways
Gartner predicts that 40% of enterprise applications will include task-specific AI agents by 2026, up from less than 5% in 2025.
Gartner also predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls.
The safest first tasks are high in volume, easy to check and cheap to undo.
Keep people in charge of decisions that are irreversible, regulated, emotionally sensitive or brand new.
Launch in stages: test on past cases, run in shadow mode, then go live with approval gates and clear metrics.
Why this decision matters right now
Adoption is moving quickly. McKinsey's State of AI 2026 survey, with 1,719 participants in 97 countries, found that 40% of respondents from large organisations now report scaling AI agents, up from 27% a year earlier. Among smaller organisations, that figure stayed flat at 22%.
Returns are moving more slowly. In the same survey, 37% of respondents attribute at least some EBIT impact to AI, about the same as the year before, and AI high performers still account for just 6% of respondents, unchanged from 2025.
Gartner's warning explains part of that gap. It describes most agentic AI projects as early experiments driven by hype and often misapplied. The technology is rarely the deciding factor. Task selection is. Choose the wrong first task and an agent becomes an expensive demo. Choose the right one and it pays for the next project.
What an AI agent actually is
An AI agent is a large language model that can do more than answer. It can use tools, such as your CRM, helpdesk, database or email, and work through a task step by step. It plans, takes an action, checks the result and decides what to do next, until the goal is met or it hands the task to a person.
That makes it different from two technologies it is often confused with:
A chatbot answers questions. It does not take actions in your systems.
Robotic process automation (RPA) follows a fixed script. It can break when a form, layout or rule changes.
An agent goes further than both. It can read messy inputs, decide which tool to use, take action and adapt when something unexpected happens. That flexibility is its value, and also its risk.
Anthropic, the company behind Claude, draws a useful distinction in its guidance on building agents. Workflows follow predefined code paths, with AI handling specific steps. Agents decide their own process and tool use. Its advice is to find the simplest solution possible and only add complexity when it is needed, because agentic systems often trade speed and cost for better performance.
Be careful with vendor claims, too. Gartner calls the practice of rebranding chatbots and RPA tools as agents "agent washing", and estimates that only about 130 of the thousands of agentic AI vendors are real.
The 5-question test: what to automate first
Score each candidate task against these five questions. The best first projects score well on all of them.
Is it frequent and repetitive? A task done hundreds of times a week returns far more than one done twice a month. Multiply frequency by the minutes each one takes to see the real size of the prize.
Is the input digital and the data reachable? Agents need data they can read and systems they can access through an API. If the information lives in someone's head or on paper, fix that first.
Can you define "done" and check it? If you cannot describe a correct result, you cannot measure whether the agent produced one. Good first tasks have clear success criteria.
What does a mistake cost, and can it be undone? A wrongly tagged ticket costs seconds. A wrongly approved refund or a wrong promise to a customer costs money and trust.
Would someone care that a human did it? Some work carries value precisely because a person does it, such as a difficult conversation or a personal apology.
The rule that falls out of this test is simple. Automate first the work that is high in volume, easy to check and cheap to undo.
Workers broadly agree. A Stanford study of 1,500 workers across 844 tasks and 104 occupations found that workers were positive about AI agents automating 46.1% of the tasks studied. The most common reason, given in 69% of those responses, was to free up time for higher-value work. Among workers who raised concerns, the top worry was not job loss. It was a lack of trust in the accuracy and reliability of AI, cited by 45%.
The same study found a mismatch that every buyer should note. Of the Y Combinator company-task mappings it analysed, 41% fell into zones where workers do not want automation or where it is a low priority. In plain terms, many of the start-ups building agents are targeting work that people would rather keep.
Where AI agents pay off first: examples by department
McKinsey's 2026 survey found that respondents most often scale AI agents in IT, knowledge management and software engineering. These are areas with plenty of digital data and clear success criteria. Here is what strong first projects look like across a typical business.
Customer support
Triage incoming tickets by topic, urgency and customer value, then route them to the right team
Draft replies from your own knowledge base, for a support team member to approve
Summarise a customer's full history before a person picks up the case
Answer routine questions such as order status, delivery times and account changes
Sales and marketing
Research and enrich new leads before the first call
Update the CRM after meetings, with notes, next steps and follow-up tasks
Draft personalised follow-up emails for a salesperson to review
Turn webinars and calls into first drafts of content
Finance and operations
Extract data from invoices, purchase orders and contracts into your systems
Match payments to invoices and flag the ones that do not reconcile
Check expense claims against policy and highlight exceptions
Monitor transactions for anomalies and alert a person
HR and internal operations
Answer employee questions about policies, leave and benefits
Schedule interviews and coordinate onboarding checklists
Search internal documents and return answers with sources
IT and engineering
Handle routine access requests and password resets
Search runbooks and past incidents during an outage
Generate tests and review code changes for a developer to approve
Notice the pattern. In almost every example, the agent prepares, drafts, matches or routes, and a person approves anything that matters.
What to keep human, for now
Some work should stay with people even when an agent could technically attempt it.
Decisions you cannot easily undo. Payments, large refunds, contract terms and anything that commits the business to a cost. Let the agent prepare the case and a person approve it.
Anything that creates a legal commitment. Air Canada's chatbot told a passenger he could claim a bereavement discount after booking, as long as he applied within 90 days of the ticket being issued. That was not the airline's policy. When the case reached a tribunal, Air Canada argued that it could not be held liable for information provided by its chatbot. In February 2024, the tribunal rejected that argument and ordered the airline to pay. The lesson is clear: you own every word your agent says.
Emotionally sensitive conversations. Complaints, cancellations, bereavement, financial hardship and vulnerable customers. These are the moments when customers most want a person, and when the wrong tone can cost you the relationship.
Hiring and people decisions. Shortlisting, rejecting, promoting and dismissing affect careers. The EU AI Act classes AI systems used to recruit and select candidates as high-risk, and similar scrutiny is growing elsewhere.
New problems without a playbook. Strategy, new markets and one-off crises need judgement, context and accountability that agents do not have.
Human in the loop: how to design agents that stay safe
Keeping humans involved does not mean doing the work twice. It means designing the system so that people step in at the right moments. These are the patterns we recommend:
Approval gates. The agent proposes and a person approves any action above a set threshold, such as refunds over a fixed amount.
Confidence-based escalation. When the agent is unsure, or a customer asks for a person, it hands over with a full summary.
Least-privilege access. Start with read-only access. Grant write access one tool at a time, and only where it is needed.
Evaluations before launch. Test the agent on a set of real past cases and measure how often it gets them right.
Audit logs. Record every action, input and decision, so you can explain what happened and fix it.
Cost controls. Track model usage per task and set limits, so a looping agent cannot run up a bill overnight.
Open standards help here. The Model Context Protocol (MCP), for example, gives agents a consistent way to connect to tools and data, which makes it easier to control exactly what an agent can see and do.
A 30-day plan for your first AI agent
Week 1: choose and define. List 10 to 20 repetitive tasks and score them with the 5-question test. Pick one. Write down what "done" means and the metric you will improve, such as handling time, accuracy or backlog.
Week 2: build the simplest version. Map the data and systems involved. Start with a workflow that has one AI step before you build a fully autonomous agent. Collect 50 to 100 real past cases as a test set.
Week 3: run in shadow mode. The agent works in parallel with your team, but people still make every decision. Compare its output with theirs and fix the gaps.
Week 4: go live in a controlled way. Hand over a small share of real volume, with approval gates on anything risky. Expand only when the numbers hold.
How to measure whether it is working
Two of the three reasons Gartner gives for cancelled projects are escalating costs and unclear business value. You can avoid both by measuring from the first day.
Cost per completed task, including model usage and the time people spend reviewing
First-time accuracy, checked against your test set and spot checks
Escalation rate, meaning how often the agent hands work to a person, and why
Time to resolution for customers or internal teams
Customer and employee satisfaction, because a faster process that people dislike is not a win
If cost per task is not falling, or accuracy is not stable, pause and fix the design before you scale.
Start small, measure everything, keep people in charge
At DibHub, we believe the businesses that get the most from AI agents will not be the ones that automate the most. They will be the ones that automate the right work first, prove the value in numbers and keep people in charge of the decisions that define their reputation.
If you are weighing where AI agents fit in your business, book a free 30-minute consultation with a senior DibHub engineer. We will help you score your workflows, choose the first one worth automating and plan a safe, measurable first version.
And one question to leave you with: which task in your business would you hand to an AI agent tomorrow, and which one would you never let go of?
Faique Ali Raza
Business Development Lead
DibHub
FAQs
What tasks should a business automate with AI agents first?
Start with tasks that are frequent, repetitive and easy to check, where a mistake is cheap to undo. Good examples include ticket triage, drafting replies from a knowledge base, updating the CRM after meetings, extracting data from invoices and matching payments. Keep a person in charge of payments, legal commitments and sensitive conversations.
What is the difference between an AI agent and a chatbot?
A chatbot answers questions. An AI agent can also take actions in your systems, such as updating a record, sending an email or processing a request. It plans the steps, uses tools and checks the results along the way. That makes agents more powerful than chatbots, and it is also why they need approval gates and clear limits.
Do AI agents need human oversight?
Yes, especially at the start. The safest approach is human in the loop: the agent prepares and drafts, and a person approves anything costly, irreversible or sensitive. As accuracy is proven on real cases, you can widen what the agent handles on its own, while keeping escalation to a person available at all times.







Comments
Loading comments…