- Why Build vs Buy AI Agents Is Rarely One Decision
- When Building Makes Sense, and When Buying Does
- Three Questions Every Path Has to Answer
- One Company, Three Different Answers
- Accountability Stays With You Either Way
- Settling the Questions Before the Choice
- Start With One Team
- Frequently Asked Questions
- How do you approach the build vs buy AI agents decision?
- When does it make sense to build an AI agent in-house?
- What costs come with buying an AI agent?
- Should you build or buy AI agents, or do both?
- Does buying an AI agent transfer responsibility to the vendor?
- Do agents employees build for themselves count as building?
Quick Answer: The build vs buy AI agents decision works best when it is made for each piece of work rather than once for the whole organisation. Buy where the work looks much the same in every company. Build where the work is distinctive, where data or regulation demands control, or where the know-how being encoded is itself the advantage. On either path, three things decide whether it pays: knowing which work the agent is for, keeping what it encodes, and measuring the result yourself.
In a survey of 495 US enterprise decision-makers carried out in November 2025, Menlo Ventures found that 76 per cent of AI use cases were bought rather than built, up from 53 per cent a year earlier. Ready-made products, it concluded, were reaching production faster.
That describes a market growing up. It does not settle the build vs buy AI agents question for any one organisation, and most build versus buy guidance reaches the same verdict because it is written by companies selling the buy side. GrowthNation sells agents too. Building still wins in some cases and buying in others, and the questions that decide the outcome are the same whichever way an organisation goes.
Why Build vs Buy AI Agents Is Rarely One Decision
Most advice treats build vs buy AI agents as a stance an organisation takes: a build culture or a buy culture, settled once at leadership level. The work does not arrive that way. Menlo found that organisations typically surface ten or more possible AI use cases, and the right answer for one need not be right for the next.
Older evidence points the same way. In a study of 61 IT sourcing decisions made by 40 US and UK organisations between 1991 and 1995, Mary Lacity and Leslie Willcocks found that selective decisions, outsourcing some functions and keeping others, succeeded more often than either total outsourcing or total insourcing. The technology has changed since then. The lesson has not: judging each piece of work on its merits beat picking a side.
A third path rarely appears in the comparison at all. A great deal of building already happens informally, as employees assemble agents in their own accounts, outside any approved tool. Menlo estimates that 27 per cent of AI application spend comes through individual users rather than central purchasing, possibly close to 40 per cent counting employees paying with their own cards. That is neither a deliberate build nor a purchase, but it still needs a decision.
When Building Makes Sense, and When Buying Does
Neither side of the build vs buy AI agents choice is a safe default. When Andreessen Horowitz surveyed 100 CIOs in 2025, it found a marked shift towards buying ready-made AI applications, partly because internally built tools proved hard to maintain and often gave no business advantage. The exception was regulated and high-risk industries such as healthcare, where privacy and compliance outweigh speed.
| Building an AI agent | Buying an AI agent | |
| Best fit | Work distinctive to how the business competes | Work that looks much the same in every company |
| What you control | The logic, data handling and choice of model | Configuration, within the limits the vendor sets |
| Speed to first use | Slower, and dependent on internal capacity | Faster for common use cases |
| Ongoing cost driver | Maintenance, model changes and the people who run it | Licence and usage pricing, and the vendor’s roadmap |
| Main risk | The builder or the team moves on | Lock-in, as switching costs rise over time |
| What stays if it ends | Code, prompts and data, if they were documented | Whatever the exit terms allow |
| Regulated data | Full control and full responsibility | Due diligence and contract terms carry the weight |
Where building earns its cost
Building is worth its cost when most of these hold, rarely just one:
- The work is distinctive to how the business competes, so a product built for every company would flatten what makes it valuable.
- Data or regulation requires control over how information is handled, stored and audited.
- The know-how the agent encodes is itself the advantage, such as a pricing judgement or a specialist assessment.
- There is capacity to run the agent after launch, not just build it: to maintain, test and update it as models change.
Where buying is the better call
Buying fits work that looks much the same from one company to the next, such as routine customer queries or standard correspondence. It also fits when speed matters more than a perfect fit, and when a vendor is better placed to keep pace with the models underneath. The same CIOs made that point themselves: a dedicated application team can keep re-evaluating models for cost and performance, which an internal team rarely has time to do.
The costs that arrive after launch
Both paths carry costs that a pre-launch business case tends to miss. In a 2026 survey by DataRobot, a vendor that sells an agent platform, 72 per cent of 413 AI practitioners and leaders in regulated industries said they had already exceeded their expected operating budgets, and 71 per cent said running agents cost more than building them. Buying has its own version: once prompts, guardrails and testing are tuned to one setup, the CIOs reported, switching becomes a job measured in engineering time. On either path, checking output costs time that tends to surface only when projects stall once the pilot ends.
Three Questions Every Path Has to Answer
Comparison checklists focus on the agent itself: features, speed, price and integrations. The questions that decide whether either path pays off sit before and after the agent, so they rarely make the list. They are worth answering before the build vs buy AI agents decision, not after it.
Do you know which work the agent is for?
Every comparison assumes the use case is already known. Often it is known only to the people doing the work. The repetitive tasks, workarounds and week-eating cases are rarely written down anywhere a buyer or developer would find them. Choosing which work goes first depends on drawing that know-how out before anything is bought or built, because an agent pointed at the wrong work fails on either path.
Who keeps what the agent encodes?
An agent encodes how a piece of work gets done: the prompts, the rules, the exceptions and the judgement calls. On the buy side, the question is what the organisation can take with it if the contract ends: prompts, configurations, logs and data. On the build side, it is what happens when the builder moves on, and whether anyone else understands the agent. The legal side of who owns an agent an employee builds is separate. The practical question is whether the capability belongs to the organisation or to whoever set it up.
Who measures whether it paid off?
A vendor’s dashboard is not a baseline. The CIOs Andreessen Horowitz surveyed were uneasy with outcome-based pricing for this reason, unsure how outcomes are set, measured and attributed. Whichever path is chosen, measurement should start from a baseline set before launch by the organisation itself, so the result is a figure a sceptic can reproduce rather than a claim taken on trust.
One Company, Three Different Answers
Consider a regional logistics firm working out where agents fit. It needs an answer for each piece of work, not a single policy, and one answer concerns an agent nobody officially commissioned. It helps first to look at the whole process each task sits inside, and whether a step needs an agent or fixed automation at all.
| Work | Likely path | Why | What to settle first |
| Customer delivery-status queries | Buy | Common across the industry and well served by existing products | Exit terms, and how pricing grows with volume |
| Quoting for non-standard loads | Build | Relies on two senior estimators’ judgement, which is the firm’s edge | Capturing that judgement before either estimator retires |
| Weekly carrier performance report, built by a planner in a personal account | Bring in-house and run centrally | Useful, but held in one person’s account with nobody else maintaining it | Who owns and maintains it from now on |
| Customs paperwork checks | Buy, with careful terms | Specialist suppliers exist, but the data is regulated | Requirements written down before procurement starts |
The firm answered the build vs buy AI agents question four times and reached three different answers. The row most organisations miss is the third, because nobody decided to build it.
Accountability Stays With You Either Way
Buying an agent moves the work to a supplier. It does not move the responsibility. The Information Commissioner’s Office states in its guidance on AI and data protection, currently under review following the Data (Use and Access) Act, that an organisation buying an AI system should evaluate its trade-offs independently, set out its requirements at the procurement stage, and be able to modify the system or switch provider if it stops being compliant. The same guidance says senior management cannot hand these issues to engineering teams.
The Lacity and Willcocks study found similar habits paying off a generation ago: decisions made jointly by business and technical leaders, internal bids weighed alongside external ones, and shorter contracts all succeeded more often. Before committing to either path:
- Write down requirements before talking to suppliers, so options are judged against the organisation’s needs, not a feature list.
- Cost an internal option alongside the external bids, rather than comparing bids only with each other.
- Keep the first term short, with an exit route that names what the organisation takes with it.
- Name a business owner for each agent, not only a technical one, and make a security review part of the decision rather than an afterthought.
Settling the Questions Before the Choice
GrowthNation is one of the buy options for part of this problem, and it is built around the three questions above rather than around any single build vs buy AI agents choice. Every person in every team is interviewed, by voice or text, so the list of work worth automating exists before anyone decides whether to build or buy AI agents for it. The resulting agents are served centrally, so the work built for the organisation stays with the organisation rather than sitting in an individual’s account. Hours saved are tracked per agent from the first day of use, against the organisation’s own starting point, and rolled up across the organisation.
Some of what surfaces will suit an in-house build or a specialist product instead. Knowing which is part of the point of mapping the work first.
Start With One Team
Start with one interview. GrowthNation will map a single team, show you the scorecard and the list of work it surfaces, and hand over a working agent. No deck.
Frequently Asked Questions
How do you approach the build vs buy AI agents decision?
Approach build vs buy AI agents as a decision for each piece of work rather than one choice for the whole organisation. Buy agents for work that looks much the same in every company, and build them where the work is distinctive, the data needs tight control, or the know-how is the advantage. Before either, establish which work the agent is for, who keeps what it encodes, and how its results will be measured.
When does it make sense to build an AI agent in-house?
Building makes sense when the work is distinctive to how the business competes, when data or regulation requires direct control, when the knowledge being encoded is itself a competitive advantage, and when there is capacity to maintain the agent after launch. If only one holds, buying or adapting an existing product is usually the better start.
What costs come with buying an AI agent?
Beyond the licence, buying brings usage-based pricing that grows with adoption, switching costs that rise as prompts and testing are tuned to one product, and the time people spend checking output. Responsibility for how it handles data stays with the buyer, so due diligence and contract terms matter.
Should you build or buy AI agents, or do both?
For most organisations, both. There are usually many possible uses for AI agents, and build vs buy AI agents has a different answer for each. Research on technology sourcing has long found that selective decisions, made piece by piece, succeed more often than committing everything to one approach.
Does buying an AI agent transfer responsibility to the vendor?
No. Under UK data protection guidance, an organisation that buys an AI system remains accountable for how it handles personal data. It is expected to evaluate the system’s trade-offs independently, set its requirements at procurement, review the service regularly, and be able to change or replace it if it stops meeting its obligations.
Do agents employees build for themselves count as building?
In effect, yes, although nobody chose to build them. Agents that employees assemble in their own accounts often solve real problems, but they depend on one person and sit outside central oversight. The useful ones are worth bringing under central ownership, where they can be maintained, measured and kept when the person who built them moves on.
