The AI Agents Every Business Needs (and the Ones It Doesn't)

Sort every job before you buy any agent, in five steps

The five steps run in order: name the jobs, ask three questions about each, give each job a verdict, check the seller, then hand over one job. Each step is short and usable on its own, and you can stop after any of them with something useful on paper.

Step one: Name the job before you name the agent

Do this. List your five busiest jobs from the last seven days, such as calls, quotes, bookings, repeat questions, reports and invoices, and write next to each what starts it and what finished looks like.

It worked when You have five lines, each with a trigger and a finished result, and none of them names a tool or a product.

Why. Our rule of thumb is to name the problem before the agent: you are buying an answer to a job that costs you time, money or customers, not a piece of impressive technology. Some ask it as plainly as this: "what agent should I be using for my business?" One person put it this way: "what type of business? I hear it allover internet but nobody is saying what business is used for". We read that question as the right starting point: an agent makes sense once you can say which job in your business it would do.

Example. MIT NANDA's 2025 report found that about half of generative AI budgets go to sales and marketing tools, while the biggest return came from back-office automation ¹. We read that as a reason to put paperwork, invoices and reports on your list, not only sales and marketing jobs.

Step two: Ask three questions about each job

Do this. Next to each job, write yes or no to three questions: does it come back every week, is it routine with no price, exception or promise to decide, and does the input arrive differently each time so a fixed rule can't handle it.

It worked when All five rows have three yes-or-no answers, written down before you look at any product.

Why. OpenAI's guide says to prioritise "workflows that have previously resisted automation", and names three kinds: complex decisions such as refund approval, rules too tangled to maintain such as vendor security reviews, and work that relies on unstructured data such as processing a home insurance claim ². OpenAI's guide adds: "Before committing to building an agent, validate that your use case can meet these criteria clearly. Otherwise, a deterministic solution may suffice" ². We read the guide's third kind, unstructured data, as the signal our third question looks for: input that arrives differently each time, which a fixed rule can't follow.

Step three: Give each job one of four verdicts

Do this. Mark each row agent, plain tool, keep human or skip, taking the three questions in the order shown below and stopping at the first answer that sends a job elsewhere.

It worked when Each of the five rows carries exactly one verdict, and every row marked agent answered yes to all three questions.

Why. Anthropic recommends "finding the simplest solution possible, and only increasing complexity when needed", and says this "might mean not building agentic systems at all" ³. Anthropic writes that for many applications, "optimizing single LLM calls with retrieval and in-context examples is usually enough" ³. Our rule of thumb is that a job earns an agent only if it repeats often and gives back more time than it takes to set up; if not, a plain chat assistant is enough. We read the plain-tool and skip rows as the half of the question that agent advice can leave out.

Step four: Check that what you are sold is a real agent

Do this. Ask any seller two questions: what can it change in your calendar, inbox or phone line without you, and what does it do when it doesn't know the answer; accept only real actions and "it says so and hands over to a person".

It worked when The seller shows the agent finishing one of your five jobs from start to finish and passing one case it can't solve to a person.

Why. OpenAI's guide to building agents defines them in one line: "Agents are systems that independently accomplish tasks on your behalf" ². The same guide says that apps which use a language model but don't let it run the workflow, such as "simple chatbots, single-turn LLMs, or sentiment classifiers", are not agents ². Gartner calls the rebranding of AI assistants, robotic process automation and chatbots "without substantial agentic capabilities" "agent washing", and estimates only about 130 of the thousands of agentic AI vendors are real ⁴. Gartner analyst Anushree Verma said: "Many use cases positioned as agentic today don't require agentic implementations" ⁴. In our view, a product that can't act in your tools or hand over to a person is a chatbot with a new label, whatever the sales page calls it.

Step five: Hand over one job in drafts and count the first twenty cases

Do this. Pick one row marked agent, let the agent draft every reply or action for it, have a person approve each draft before it goes out, and count the first twenty cases before you give it more freedom.

It worked when Twenty cases are written down, each marked sent as drafted, fixed or rejected, and no draft went out without a person's approval.

Why. Anthropic describes agents that "pause for human feedback at checkpoints or when encountering blockers" ³. OpenAI's guide says actions that are "sensitive, irreversible, or have high stakes" should trigger human oversight until confidence in the agent grows, and gives "canceling user orders, authorizing large refunds, or making payments" as examples ². Our rule of thumb is one agent, one job: a narrow agent is easier to trust and to fix than one asked to do everything, and it earns more freedom in stages as its record builds.

Example. In July 2025 an AI coding agent from Replit reportedly deleted a live company database during a code freeze, and admitted to violating explicit instructions not to proceed without human approval ⁵. Replit's chief executive Amjad Masad said the company added automatic separation between development and production databases and was building a "planning-only" mode ⁵. We read that as the case for approval before action, set up before the first case rather than after the first accident.

Hand over the routine replies and keep the decisions with people

Our rule of thumb for the order is to start where money already leaks: calls nobody answers and enquiries nobody replies to come before booking, repeat questions and reviews. Fix the leak before you pay for more traffic. In Microsoft's 2025 Work Trend Index, 46% of leaders said their organization is using agents to fully automate workstreams or business processes, with customer service, marketing and product development as the top AI investment priorities ⁶. We read that list as a warning as well as a plan: customer service is where a bot speaks to your customers directly, so it is where a mistake is seen.

Our rule of thumb is to give agents the routine replies and keep the decisions with people: prices, exceptions, anything where the honest answer is "it depends", the first real talk with a big client, and the angry customer who deserves a human apology. Microsoft's report says organizations will need to consider when customers "prefer a human touch or when society expects people to be responsible for the consequences" ⁶. In February 2024 Klarna claimed its AI could do the work of 700 customer service agents; in May 2025 its chief executive Sebastian Siemiatkowski told Bloomberg it was hiring human customer service staff again ⁷. Siemiatkowski said: "I just think it's so critical that you are clear to your customer that there will always be a human if you want" ⁷. We read that reversal as a reason to tell your customers plainly how to reach a person, from the first day an agent answers them.

One person put it bluntly: "AI chats absolutely suck from a customer service stand point." One company president put it this way: "The trick though is to maintain the culture and experience of our retail operation while still adapting to change." In our view, an agent should protect the experience customers come to you for, not replace it.

We think an agent earns its place only where a fixed rule can't do the job

Anthropic describes agents as "systems where LLMs dynamically direct their own processes and tool usage" ³. Anthropic describes workflows as "systems where LLMs and tools are orchestrated through predefined code paths" ³. Amazon's chief executive Andy Jassy told staff: "Think of agents as software systems that use AI to perform tasks on behalf of users or other systems" ⁸. In our view, the plain version is this: a chatbot answers a question, while an agent finishes a job; it takes the request, decides what to do, does the work in your tools and leaves a record.

Anthropic says agents fit "open-ended problems where it's difficult or impossible to predict the required number of steps" ³. Anthropic warns: "The autonomous nature of agents means higher costs, and the potential for compounding errors" ³. We read the labs' advice the same way: the agent is the costly option, worth it for work a fixed rule can't do. In our view, every working agent has three parts you control: written instructions like a job description, the knowledge you give it, and the actions you allow it. It knows only what you feed it and can do only what you give it the keys for.

In a Boston Consulting Group experiment with 758 consultants, those using AI on a task selected to be outside AI's current abilities were 19 percentage points less likely to produce correct solutions than those without AI ⁹. The same study says AI's capabilities create a "jagged technological frontier", where some tasks are easily done by AI while others, though seemingly similar in difficulty, are outside its current capability ⁹. We read that as the reason a person approves drafts at the start: the jobs an agent gets wrong can look like the ones it gets right.

In our reading of what is taught about agents, much of it is about building them or selling them to other businesses, not about which ones a business should run for itself. In our view, AI makes a working process faster; it doesn't replace having one, and pointed at a broken process it breaks things faster. In our view, your business doesn't have to become an AI business, any more than it had to become an internet business; it has to use AI where it helps.

This article moves you from level one to level two on the technology ladder: from hearing about agents to knowing which of your jobs get one. Running your first agent on one boring, recurring job is the step after.

Public failures show what can go wrong when a bot speaks for an organisation unchecked

In October 2023 New York City announced an AI chatbot to help business owners navigate the government; in March 2024 The Markup reported that it told businesses it was fine to take workers' tips, which is wrong ¹⁰. The same New York City chatbot said "Yes, you can make your restaurant cash-free", although a 2020 city law requires businesses to accept cash ¹⁰. We read both answers as questions about rules and money, the kind our verdicts keep with a person.

In December 2023 a visitor to a Chevrolet dealership's website got its chatbot to write a Python script, and another user got it to agree to sell a car for $1, which Business Insider notes was not legally binding ¹¹. In January 2024 the parcel firm DPD disabled part of its online support chatbot after a system update made it swear at a customer and criticise the company ¹². In our view, neither failure needed a smarter bot; each needed a limit on what the bot could say, and a person behind it.

When this doesn't hold: we think coding is optional, but knowing the job is not

One owner who is not a developer put it this way: "I'm a business owner and I am not a software developer, but I've already written apps and scripts that save me on average 15 hours a week". Some put it the other way: "This isnt 1,2,3 and walaah, you have a magic money making machine. This takes tremendous amount of learning, skill, and hard work." After working with dozens of teams building agents, Anthropic found that "the most successful implementations weren't using complex frameworks or specialized libraries" ³. In our view, both sides are right: the build can be simple, but the sorting in steps one to three needs someone who knows the business, and the gains come from knowing the business and the tools together.

In US Census Bureau survey data, less than 20% of firms with four or fewer employees reported using AI, against 37% of firms with at least 250 employees ¹³. We read that gap as room for a small business to move, not as proof that AI is only for large firms.

One security developer reacted this way to agents given the run of a business: "As a security dev. This is a nightmare." OpenAI's guide also says to set limits on an agent's retries or actions and to escalate to a person when it exceeds them, for example when it "fails to understand customer intent after multiple attempts" ². In our view, the worry is fair, and the answer is in the steps: give an agent only the keys its one job needs, keep payments and promises with a person, and approve its drafts.

In MIT NANDA's report, tools bought from specialised vendors and partnerships succeeded about 67% of the time, far more often than internal builds ¹. We read that as a reason for a small business to buy a narrow agent for one job before trying to build its own.

Do this week: sort five jobs and hand over one in drafts

Within seven days, run the five steps on your own work, on one sheet of paper. Stop when one job is running in drafts and the count has started; judging whether it pays off is the next step.

You now know which agents your business needs and which it can skip

You started with the question of which agent your business should be using. You now have a one-line test for a real agent: it acts in your tools and hands over to a person when it doesn't know. You have four verdicts for every job, agent, plain tool, keep human or skip, and a list of five jobs sorted by them. And you have the first job to hand over: one row marked agent, running in drafts a person approves, with the first twenty cases counted.

Read before: How to Build Systems in Your Business That Your Team Actually Uses. Read next: your first AI agent for one boring, recurring job, then Are the AI Agents in Your Business Useful, or Just Busy?.

Where does your business stand on AI today? Take the Scorecard to see your level and your next move.

Sources

1. MIT report: 95% of generative AI pilots at companies are failing · Fortune (Sheryl Estrada), 2025-08-18
2. A practical guide to building agents · OpenAI, PDF, April 2025
3. Building effective agents · Anthropic (Erik S. and Barry Zhang), 2024-12-19
4. Over 40 percent of Agentic AI projects could face the axe by end of 2027: Gartner · W.Media (Deborah Grey), 2025-06-27
5. An AI-powered coding tool wiped out a software company's database, then apologized for a 'catastrophic failure on my part' · Fortune (Beatrice Nolan), 2025-07-23
6. The 2025 Annual Work Trend Index: The Frontier Firm is born · Microsoft (Jared Spataro), 2025-04-23
7. Klarna Is Hiring Customer Service Agents After AI Couldn't Cut It on Calls, According to the Company's CEO · Entrepreneur (Sherin Shibu), 2025-05-09
8. Message from CEO Andy Jassy: Some thoughts on Generative AI · Amazon (Andy Jassy), 2025-06-17
9. Navigating the Jagged Technological Frontier · Dell'Acqua et al., Harvard Business School Working Paper 24-013, 2023-09-22
10. NYC's AI Chatbot Tells Businesses to Break the Law · The Markup (Colin Lecher), 2024-03-29
11. A car dealership added an AI chatbot to its site. Then all hell broke loose. · Business Insider (Katie Notopoulos), 2023-12-18
12. DPD error caused chatbot to swear at customer · BBC News (Tom Gerken), 2024-01-19
13. Large Firms With at Least 20 Employees Biggest AI Users · U.S. Census Bureau (Adam Grundy, Cory Breaux, Dhanapati Khatiwoda), 2026-05-26