Build, buy or partner

Three routes, one portfolio

Every AI use case reaches the company by one of three routes:

• Buy. A product off the shelf, or an AI feature inside software the company already pays for.
• Build. A custom agent built by an internal team.
• Partner. A custom agent built with an outside team, which the company owns and runs, alone or with that partner.

Most companies end up with all three. The question is which route fits which use case, and the answer changes as the market moves.

The evidence favours outside help for early builds. MIT’s 2025 study found that solutions bought from specialised vendors or built with partners succeeded about 67% of the time. Purely internal builds succeeded about a third as often. The lead author put the problem plainly: almost everywhere the researchers went, enterprises were trying to build their own tool.

When to buy

Buying makes sense when the process is common across companies and the vendor’s product already does it well. Meeting notes, writing assistance, coding assistants, common customer service features and AI inside existing CRM or ERP suites all fall here.

Three cautions apply.

• Check that it is really an agent. Gartner estimates that only about 130 of the thousands of vendors marketing “agentic AI” offer real agent capability. Many have relabelled existing chatbots, assistants or automation tools.
• Check what the product can reach. A bought agent is only as useful as the data and systems it can connect to. A product that cannot read the company’s own documents or write to its own systems will stay a side tool.
• Check the exit. What happens to the data, the configuration and the history if the company leaves?

When to build

A custom build makes sense when the process is specific to the company, runs on its own data, or is a source of advantage.

Signs that a use case needs a build:

• the process crosses several systems that no single vendor covers;
• the decisions depend on rules, history or judgement that only the company holds;
• the work touches customers in a way that should reflect the company’s own voice, pricing or policy;
• doing it better than competitors is worth real money.

The cost of building has fallen sharply. Stanford’s 2025 AI Index found that the price of running a model at a fixed level of capability fell from about $20 per million tokens in late 2022 to about $0.07 by late 2024, more than 280 times cheaper. Engineering time, integration and evaluation are now most of the cost of a custom agent, and they are the parts a good team can control.

When to partner

Partnering combines a custom build with outside experience. It fits companies that want an agent built around their own process and data, but do not yet have a team that has built and run agents before.

A good partnership leaves the company stronger, with:

• the code, prompts, rules and test cases in the company’s own repositories;
• the agent running in the company’s own cloud, or in an environment it controls;
• internal people trained during the build, ready to run and extend it;
• no dependency on the partner to keep the agent working.

PepTalk, a speaker booking marketplace, shows what a partner-built system can replace. Each speaker request touched at least five systems: an inbox, a speaker database, a document editor, a CRM and a calendar. Staff copied data between them by hand. Twistag built four chained agents in place of middleware between the old tools: one to turn any request into structured data, one to match it against more than 3,000 speaker profiles, one to draft the proposal with guardrails checking it before it goes out, and one to handle follow-ups. Five tools became one platform, and staff review every output in a structured interface where their corrections improve the agents over time.

The questions that decide it

For each use case, five questions usually settle the route.

When most answers point to buy, buy. When they point to build, decide between an internal team and a partner based on the skills the company already has and how fast it needs a result.

Avoid lock-in

Models change every few months. A model that leads today may be overtaken next quarter, or change its price. Two practices protect the company:

• A provider layer. Agents call models through an internal layer, so the model behind each task can change without rewriting the agent. The European RegTech platform from Chapter 4 swapped its model provider twice during the engagement without changing the product.
• Owned test cases. A set of real cases with known right answers, owned by the company, makes it possible to test a new model or vendor in days.

Twistag learned this on its own product. Refraction, an AI code generation tool Twistag built in 2022, routed each request to the model best suited to the language and task. That design carried it through provider outages, price changes and new model releases, and it served more than 20,000 engineers. In 2024, as large vendors brought similar tools to market, Twistag shut it down on purpose. A product that was worth building in 2022 had become something to buy. The lesson holds for every company: review the build or buy choice every year.

Common mistakes

• Building what can be bought. An internal team spends six months on a meeting-notes tool that already exists.
• Buying what should be built. A generic product forced onto a process that sets the company apart, until it becomes a side tool nobody uses.
• Contracts without ownership. The partner or vendor owns the prompts, the test cases or the data, and leaving means starting again.
• One model everywhere. Every agent locked to a single provider, with no way to switch when prices or quality change.
• Deciding once. A route chosen in year one and never revisited as the market changes.

What to do this month

1. Mark each use case on the short list as buy, build or partner, using the five questions.
2. List the AI features already included in the company’s main software, and check which short-list items they cover.
3. Review every current AI contract for ownership of code, prompts, test cases and data.
4. Agree a rule that every custom agent calls models through a provider layer.

Questions for the leadership team

1. For each use case on the short list, is the process common across the industry or specific to this company?
2. Which existing software vendors already offer AI features that cover part of the short list?
3. What would it cost the company to switch model provider or vendor in two years' time?
4. Who owns the code, the prompts, the test cases and the data in each current AI contract?
5. Which skills does the company want to build in-house over the next two years, and which will it keep buying?