MMOOZZGG.com

How to build an AI agent

AI use case selection: the question that decides an agent project

6 min read5 sources

AI use case selection decides the success of an agent project more than the tools you build it with. According to MIT NANDA for 2025, 95% of generative AI pilots show no measurable effect on profit, although the tools are open to everyone. The gap between the 5% that succeed and the rest lies in the choice of task and in the metric.

On this page · 5 sections

A team and a head of AI are needed

Why building is no longer the problem

Not long ago, a support agent was custom development. Today it is built from ready-made blocks: the n8n, Flowise and Dify builders, the LangChain and CrewAI libraries, and the open DeepSeek and Qwen models, which run on your own server through Ollama.

Each tool has its own license terms, but for a company’s internal agent almost all of them can be installed on your side with no license fee. For personal use, a manager can build a simple agent in an evening, and an agent with a knowledge base in a week.

In December 2024, Anthropic advised starting with the simplest solution and making it more complex only when needed. Sometimes that means not building an agent at all.

Hence a practical rule: a typical agent is built from open blocks. If someone charges for it as for unique development, ask what the price is made of.

So the question “how to build” is solved with documentation and a few weeks of work. The question “what to build” is not solved with documentation: the answer depends on the processes and money of a specific company.

Why pilots show no effect

In its report “The GenAI Divide” of August 2025, MIT NANDA found that 95% of generative AI pilots show no measurable effect on profit. In June 2025, Gartner predicted that more than 40% of agentic AI projects will be closed by the end of 2027.

McKinsey in November 2025: 88% of organizations use AI, 62% experiment with agents and 23% scale them. 39% see an effect on EBIT, the company’s profit before interest and taxes.

Gartner names the reasons for closure directly: rising costs, unclear business value and weak risk control. All three are about the choice of task and project management, not about the quality of code.

Gartner gives one more reason: of thousands of “agent” vendors, only about 130 are real, and the rest renamed chatbots and process robots. A company that has not decided what it needs easily buys an old product under a new name.

A detailed review of the numbers and reasons is in the article on why agent projects get closed.

How to choose a task with money behind it

According to Fortune’s summary of the MIT report, more than half of generative AI budgets go to sales and marketing tools, while the researchers found the biggest return in automating internal operations. The money often lies not where people look for it first.

Choose the task by calculation, not by fashion:

1. List the processes where people do repetitive work by rules: requests, booking, document checks, answers to typical questions.

2. For each one, calculate the losses in money: employee hours, lost requests, errors and fines.

3. Check whether the process has data: correspondence, a knowledge base, access to the CRM or the accounting system. Without data, the agent will answer in general words.

4. Set a metric that you can see in your systems without manual counting, and measure it before the start.

5. Start with a process where the losses are large, the data exists and the metric is simple. Queue the rest by the same calculation.

A task without a metric and without a cost of losses is not fit for a pilot, even if you can build an agent for it in an evening.

Who decides this in a company

A contractor knows how to build an agent. It sees the economics of your processes only through the specification, and it is responsible for the build, not for whether the agent earns its money back.

What to build should be decided by someone who looks at the company’s processes and economics as a whole. In a small or mid-sized company, this is a fractional AI director, with us from $999 a month, 5 hours a week.

In a large company, this is an AI department: people and rules inside the company that choose tasks, accept the work of contractors and measure the effect.

An AI director does not write code. The director goes through processes with department heads, calculates losses, chooses the first process, sets the metric and accepts the contractor’s work against it.

MIT NANDA notes that buying from specialized vendors and partnerships succeed in about 67% of cases, and in-house builds a third as often. The same report names, among the conditions of success, the involvement of division heads, not only a central AI team: the person who is responsible for the process must understand the task.

AI solutions map

The AI solutions map is a list of the company’s processes with their losses in money, the solutions that fit and the order of rollout. It answers the question “which agent to build first” before the money goes into building.

A map for the whole company takes us 7 weeks, with a price from $93,000. We fix the exact price after a non-disclosure agreement and a check of the scope.

After the map, building becomes a technical task. Each agent in the queue goes into a pilot with a metric, and the contractor gets a specification with a number, not a wish.

If you don’t have a map yet, start with one process by the steps above and a pilot: our pilot costs from $1,499 and takes 14 days.

Questions and answers

Why doesn’t the IT contractor decide which agent to build?

A contractor sees a process through the specification and is responsible for building on time. It does not see the losses, data and economics of all the company’s processes, and that is not its area of responsibility. So the task is chosen by whoever is responsible for the company’s result: an AI director or an AI department.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. Anthropic, 19 December 2024 (anthropic.com)
  2. Gartner, 25 June 2025 (gartner.com)
  3. Fortune via Yahoo Finance, 18 August 2025 (finance.yahoo.com)
  4. IT Brief, 11 November 2025 (itbrief.co.uk)
  5. Fortune, 18 August 2025 (fortune.com)

Found a mistake? Tell us

Publications

Contact us