AI Agents & Automation in Practice
Build or Buy an AI Agent? Decide Workflow by Workflow
By Mudabbir Hasan, Co-founder · Published · Last updated
Decide per workflow, not per company. Buy for commodity work that any competitor could automate or while you are still testing value. Build where the capability is part of your edge, integrates deeply with your systems, faces strict compliance, or runs at a volume where per-seat pricing becomes a tax. If the answers split, buy the commodity layer and build the core, with one governance layer over both.
Why is build vs buy for AI agents hard to get advice on?
Nearly every guide on the topic is written by a company that sells custom AI. BotsCrew and KGT Solutions, the two we read in full, both sell AI builds and both end with an offer to book a consultation. We build custom AI too, so weigh this article the same way. What is worth keeping from them is the shared logic, and what is worth ignoring is the statistics.
Which statistics should you not rely on?
Both guides lean on figures they do not link to. BotsCrew quotes that over 80 percent of AI projects fail, attributed to RAND, with no link, and says external-partner builds succeed more often than internal ones, citing MIT without naming a study. KGT attributes several figures to Gartner, Deloitte, McKinsey and Forrester, but its reference list does not clearly identify the reports. BotsCrew also gives build costs as $10K to $500K+ in the body and $30K to $500K+ in the FAQ. Cost and success-rate numbers from these pages are not a basis for a decision. Use your own quotes.
Figures in the guides we could not verify
>80%
AI projects fail
Attributed to RAND, with no link
$10K+
Build cost, in the body
The same guide's FAQ says $30K to $500K+
72%
Needed major customisation
Attributed to Gartner, weak reference
?
Partner builds succeed more
MIT cited, no study named
What do the guides agree on?
Underneath different scorecards, the criteria are the same. BotsCrew's procedure counts how many "buy" and "build" statements are true for your hardest use case. KGT scores six factors and says four or more in one column decides it, with a three-three split pointing to a hybrid. The scales differ, which is a reason not to treat either as a standard. The criteria they use are close:
- Buy when the work is commodity, you are still experimenting, you have no in-house AI capacity, or speed is urgent.
- Build when the capability is part of your competitive edge, needs deep integration with legacy or many systems, falls under regulation that demands explainability and data control, or runs at a volume where per-seat pricing becomes a tax.
- Go hybrid when the answers split: buy the commodity layer and build the differentiated core.
Buy or build?
Buy when
- The work is a commodity
- You are still experimenting
- You have no in-house AI capacity
- Speed is urgent
Build when
- The capability is your edge
- It needs deep integration with your systems
- Regulation demands explainability and data control
- Volume makes per-seat pricing a tax
How do you run the test on one workflow?
Take the hardest workflow first, because it decides whether an off-the-shelf tool will cope. Answer these seven questions about it. They are our adaptation of the criteria above, not a published scorecard:
- Could a competitor automate this the same way, with the same tool?
- Does it need to read or write several of your own systems, not one vendor's?
- Does regulation require you to explain or control how each decision is made?
- Is the data it uses unique to your business?
- Does the agent have to own the process from start to finish?
- Do you have someone who can specify, supervise and govern the system?
- Are you still unsure the workflow is worth automating at all?
Yes to questions one and seven points to buy. Yes to two through five points to build. If question six is a no, BotsCrew's warning applies to both paths: without internal ability to oversee an AI system, an off-the-shelf tool only hides the gap for longer.
Seven questions for your hardest workflow
Could a competitor automate this with the same tool?
Does it touch several of your own systems?
Must you explain or control each decision?
Is the data unique to your business?
Must the agent own the process end to end?
Can someone specify, supervise and govern it?
Are you still unsure it is worth automating?
What should sit over both a bought and a built agent?
BotsCrew recommends one governance layer over the whole estate: access controls, audit trails, role-based permissions and observability. It also suggests treating autonomy as a dial, starting with a human approving actions and moving to human oversight as reliability is proven, per task rather than per agent. We cover how to decide where the human steps in in a separate guide.
What is our own view?
Our AI service page makes two commitments that apply here. We assess the specific problem and the available data together, so feasibility is established before any model work starts. And we do not take on a project that wants AI as a marketing label with no business problem behind it. Whether the answer is a custom model, a language-model integration or an existing AI API is decided from the problem and the data, not from a default.
What should you do next?
- List your candidate workflows and run the seven questions on each one separately.
- Start with a narrow commodity workflow if you are still unsure, and build only where you hit a ceiling.
- Get real quotes for any build and for the vendor option, and ignore the headline statistics.
- Decide who will own governance before you pick a tool.
Sources
- Off-the-Shelf or Built for You: A Decision Procedure for AI Agents, BotsCrew (accessed 2026-10-10)
- Build vs Buy AI Software: The Enterprise Decision Framework for 2026, KGT Solutions (accessed 2026-10-10)
Related guides
Talk to an engineer about your project
Start with a free project call. We learn what you are building and recommend the right next step, with no pitch.
Book a call