MMOOZZGG.com

How to create an AI agent

Build or buy an AI agent: three paths compared

5 min read3 sources

Whether to build or buy an AI agent depends on the process, not on what a contractor can build. The MIT NANDA report of 2025 found that buying from specialist vendors and partnering succeed in about 67% of cases, and in-house builds three times less often. A third path sits between them: assemble the agent from open building blocks.

On this page · 4 sections

A team and a head of AI are needed

Three paths

A company that needs an agent has three paths. They differ in price, time, and who is responsible for the result.

1. A ready solution on a subscription. The service can already answer customers or sort documents, and you connect it to your channels. Launch is fast, but the settings and the dialog history stay with the vendor while you pay.

2. Assembly from open blocks. A builder such as Flowise under the Apache 2.0 license, or the LangChain and LangGraph libraries under the MIT license, an open model, and your rules. The software is free; you pay for setup and connections.

3. In-house development. A team writes the agent for your process from scratch or almost from scratch. It takes longer and costs more, but the agent does exactly what the company needs.

A subscription wins on launch time, open blocks on cost of ownership, and in-house development on fit to the process. No path wins on all three at once, so the choice starts with the question of what matters more for this process.

The difference between the second and third paths is often not visible in a sales proposal. A contractor calls assembly from ready blocks development too, so ask what exactly is written from scratch and what is taken ready-made.

For a manager, the paths also differ in who holds the system a year later. A subscription stays with the vendor; assembly from blocks and custom development stay with you, if the contract says so.

What research says

The MIT NANDA report “The GenAI Divide” came out in August 2025. Its main finding: 95% of generative AI pilots have no measurable effect on profit.

For choosing a path, another number from the same report matters more. Buying from specialist vendors and partnering reach a result in about 67% of cases, and in-house builds three times less often.

Our explanation is simple: a vendor has already made the typical mistakes with other clients, while an internal team makes them for the first time and at your expense. This does not mean your own programmers are worse; they just repeat the same task fewer times.

These are survey data from companies, not a precise measurement. But a gap of several times says that you should build your own only where nothing ready exists. Why a contractor often offers development where a ready solution would do is covered separately.

When in-house development makes sense

In-house builds reach a result three times less often, so keep them for tasks where you cannot do without them. For agents, there are four such cases:

1. The agent must act in your systems, while ready services can only answer in a chat: post documents in 1C (a Russian accounting system), change deals in the CRM, set tasks in the warehouse.

2. The agent’s work is your advantage: your own calculation method, pricing, or production logic. No ready solution exists, because nobody else has such a process.

3. Customer data or trade secrets cannot go to an outside service, and no ready solution installs on your server.

4. At your volume, a subscription costs more than your own development plus support. You check this with a three-year calculation, not with a contractor’s opinion.

You can check whether something ready exists before you get a development estimate. Ask two or three vendors for a pilot on your requests from last month and compare what share of cases each closed without a person.

If none of these points applies to you, start with a ready solution or with open blocks. Even when development is needed, it is built from ready parts: you take an open model and tune it with instructions and a knowledge base, rather than train it from scratch.

Open blocks as the middle path

Open blocks give what a subscription lacks and cost less than development from scratch. The code is open, you install it on your own server, and you do not depend on the vendor’s pricing.

Flowise builds an agent with a mouse: the model, instruction, knowledge base, and tools are connected on a diagram. The project has 55,000 stars on GitHub, that is, marks of interest from developers, so it is not hard to find an engineer who has worked with it.

LangChain and LangGraph are needed when the agent is written in code in Python or JavaScript: complex logic, many steps, and a check of the result at each one. Both libraries are under the MIT license, whose terms allow the code in commercial projects.

Assembly from blocks does not mean “free”. You pay for the connection to the CRM and 1C, for instructions and the knowledge base, for testing before launch, and for support. But a year later you have your own system, not a subscription.

This path suits most typical agents: taking leads, answering from a knowledge base, booking customers, sorting documents. A review of open tools with their licenses is on the page about tools for creating agents.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. Fortune via Yahoo Finance, 18 August 2025 (finance.yahoo.com)
  2. GitHub, 16 October 2026 (github.com)
  3. GitHub, repositories langchain-ai/langchain and langchain-ai/langgraph, 2026 (github.com)

Found a mistake? Tell us

Publications

Mike Mozg

Agentic AI use cases at AT&T: nearly 500 processes now run on agents

AT&T no longer uses AI only to help staff work faster: it is rebuilding whole business processes around agents. SiliconANGLE reported on 14 November 2026 that nearly 500 agentic AI use cases at AT&T are already in production. Here is how the company rebuilds them, where a human still checks the agent, and what other large companies can take from it.

News4 min
Contact us