
Build vs buy for enterprise AI is decided one layer at a time. Buy the model and the platform. Buy a finished product when the job is the same in every company. Build only the workflow that carries your own rules and connects to your own systems. Almost no enterprise has a reason to build a model or an orchestration platform, the software that runs AI workflows. Almost every enterprise has a few workflows no product will run the way it needs.
The route you pick changes the odds of reaching production. For its report The GenAI Divide: State of AI in Business 2025, MIT NANDA interviewed 52 organizations. Tools bought or co-developed with external partners reached deployment about 67% of the time, against about 33% for tools built fully in-house. The authors note the figures are self-reported and show correlation only. They also found that only 5% of custom enterprise AI tools reached production, so a purchase order is no answer on its own either. In our AI consulting work, that call comes before any build is scoped.
With ordinary software, build vs buy is one decision about one product. An AI system is several layers, and each one has its own answer. The model writes, reads and classifies. The platform runs the workflow, calls the model and retries what fails. The workflow holds your rules: which records to read, what counts as a good answer, when to stop and ask a person. The integrations read from and write to your ERP, CRM and help desk. A finished AI product bundles all four. A custom build buys the first two and writes the last two.
Seen this way, most of a custom build is bought. The drawing below takes apart a customer service agent we built for Pinkcube, which sells branded merchandise to Dutch and Belgian businesses. It marks which parts were bought and which were written for Pinkcube.
Every layer except one came off the shelf. The layer we built holds what is specific to Pinkcube. It sets which order questions the agent may answer and how it reads a HubSpot deal and its sub-deals. It also sets which of seven ticket categories a chat lands in, and what the Gatekeeper blocks. The Gatekeeper is the second agent, which checks every draft. The Pinkcube case study shows the agent covering all 168 hours of the week, up from 45 staffed hours.
Buy when the job is the same in your company as in everyone else's, and a product already reaches every system the job touches. Meeting summaries, drafting help and search across documents in one suite all fit. So do the AI features inside a CRM you already pay for. You get updates and support without staffing a team, and you can switch off the subscription when a better product arrives.
The MIT NANDA study found that about 83% of pilots of generic chatbots such as ChatGPT and Copilot went on to implementation. Users praised those chatbots for flexibility and familiarity. The same users described custom and vendor-pitched tools as brittle or poorly matched to their actual workflows. A general-purpose product earns its place on general-purpose work. A product sold as a fit for your workflow needs testing on that workflow before anyone signs.
Buying fails in two predictable places. The product cannot reach a system the workflow needs, so the team copies data in and out by hand. Or the product runs the vendor's version of the process, and your team changes how it works to match. If either shows up in a trial, the job belongs on the build side of the line.
Build the workflow when one of four things is true. No product reaches all of the systems involved. The data must stay inside your own environment. Your team needs to change the rules without waiting for a vendor. The workflow is part of how you win and serve customers. Run each use case through the questions below. Each answer moves it toward one route.
An anonymous example from our own work shows the first condition. A Dutch Odoo implementation partner spent a year trying to automate appointment scheduling for a dental care group inside Odoo itself. We moved the scheduling rules into a separate engine and kept Odoo as the system of record, with five sync workflows running nightly. An AI agent now places about 1,500 appointments a week for 200 practitioners. Any appointment it cannot place comes with a report explaining why, as the AI agent development page describes.
The third condition is the one buyers underrate. When we built daily MI (management information) alerting for IES Limited, we put the alert thresholds in a control panel the business edits. IES made 34 threshold changes in six weeks with zero developer tickets, as the IES daily MI alerting case records. A bought product would have put each of those changes in a vendor's queue.
For most enterprise use cases the answer is a hybrid route. You buy the orchestration platform and the models, and build the workflow and the integrations for your systems. We build on n8n, where Ovidius is an official Select Partner. Teams standardised on Azure keep the same split by building on the Azure stack they already pay for, through Microsoft Foundry. n8n bills per workflow execution, so one run counts once however many steps it takes. Its standard version can be self-hosted in your own cloud account, and the workflows, credentials and execution logs stay there. When n8n is the wrong fit for a job, the AI agent platform comparison scores eight platforms against the same four jobs.
| Route | You own | Cost grows with | Main risk |
|---|---|---|---|
| Buy a product | Your configuration and your data exports | Seats or usage, at the vendor's prices | The product cannot reach a system, or runs the vendor's process instead of yours |
| Buy the platform, build the workflow | The workflows, rules and integrations, running in your account | Model usage and hosting; self-hosted n8n Community Edition has no fee per run | A workflow with no named owner after launch |
| Build everything in-house | All of it, including the upkeep | Your team's salaries, whatever the volume | Half the deployment rate of bought tools in the MIT NANDA sample |
The hybrid route still means building. You write the business logic, test it against real data and run it after launch. That takes the same discipline as any software project. It removes the work no enterprise should pay for twice: model training, job queues, retries, connectors to common SaaS tools and an editor your team can read. Our guide to how agents connect to your ERP, CRM and data covers the integration layer in more detail.
The routes differ less in total than in what the bill grows with. A bought product charges for seats or usage at prices the vendor sets, and the bill grows with headcount or volume. A build on a bought platform costs the build up front, then model usage and hosting every month. A fully in-house build adds the salaries of the people who keep it running, at any volume.
Running costs on a build can be small next to the work it replaces. GoKickflip's content pipeline runs on n8n, with about 33 executions per article. It spends $1.79 in AI tokens per article, or $12.11 including all infrastructure. Before the pipeline, a freelance article cost $700. The GoKickflip case puts the pipeline at $6,317 a year against $364,980 for the human process at equal volume.
Our breakdown of what AI consulting costs explains how a build is billed and who pays for an overrun under each pricing model. Our guide to what AI consulting is sets out what an engagement contains at each stage.
If the list is long, the AI Audit does this work across your processes in 4 to 8 weeks. It ends in a ranked 90-day roadmap that is yours whether or not you build with us.
Both, at different layers. Buy the model and the platform, and buy finished products for jobs that are the same in every company. Build the workflows that carry your own rules, connect to your own systems or set you apart from competitors.
Buying costs less to start, and the bill grows with seats or usage. A build on a bought platform costs more up front, then model usage and hosting each month. Those monthly costs can be small: GoKickflip's pipeline runs at $12.11 per article including all infrastructure. Compare the two over three years at your expected volume.
You buy the model and an orchestration platform such as n8n, and build the workflow and integrations on top. The workflows run in your own account, and nobody on your team maintains a model or a job queue.
About 33% of tools built fully in-house reached deployment in MIT NANDA's 2025 study of 52 organizations. For tools bought or co-developed with external partners, the figure was about 67%. The authors note the figures are self-reported and may reflect differences between the organizations rather than the route alone. If you take the partner route, our list of the top AI consulting companies for enterprise sorts firms by the job each one suits.
Yes, and it is often the right order. A bought product shows which parts of the workflow it cannot handle, and that gap becomes the scope of the build. Check before you sign that you can export your data and configuration, so the move stays open.
If you have a workflow that sits on the build side of the line, book a discovery call. A senior partner will ask which systems it touches and what it costs you today. Before we scope a build, they will tell you whether a product would do the job.