Updated November 2, 2026. The original publication date is retained above.
Start with the work
The build-vs-buy conversation in AI usually gets framed as an either/or choice. One option is a hybrid: you buy the foundation (an LLM API, a document processing service, a vector database) and build the integration layer that connects it to your specific systems, data, and workflows.
Start with the requirements your workflow must meet. Identify what existing tools cover and which gaps, if any, justify custom engineering.
Write one concrete job first: who does it, what they receive, what they produce, which systems they may use and who checks the result. Include the option to improve the process without AI. A reporting handoff with known fields may need ordinary workflow automation; a task requiring interpretation may justify an evaluated AI component.
When Off-the-Shelf Works
Evaluate a purchased tool when it meets the requirements that matter for your workflow:
- The problem is generic. Start by evaluating existing tools for tasks such as transcription, translation, and summarization. Test quality on representative examples before choosing a tool.
- Its data access fits. A purchased tool may already connect to your internal documents or business systems. Check its supported connectors, identity model, permissions, retention, and source update behavior against your requirements.
- Setup effort is manageable. Compare configuration, data preparation, security review, and staff training with the work needed for a custom option. Include the work needed to make the tool useful to your team.
- The vendor's roadmap aligns with yours. You're comfortable depending on their updates, pricing, and continued existence.
When Custom Makes Sense
Consider custom work when you can identify a requirement that available tools do not adequately meet:
- A required connection is missing. Start by checking supported connectors and exports for your data. For example, Microsoft documents Copilot connectors for external content with permission-based filtering. Whether a particular connector fits your sources and access rules still needs evaluation.
- Accuracy requirements are domain-specific. Test domain terminology on representative examples. Retrieval or fine-tuning may help, but the result still needs evaluation.
- A required control is unavailable. Map data flows, storage, access, audit needs, and responsible owners for each option. Compare the required controls with the supporting evidence for each option.
- The workflow fit is insufficient. Test the steps staff actually perform, including exceptions and handoffs. Estimate configuration and workaround costs before deciding whether a custom integration is justified.
The Hybrid Approach
A hybrid architecture can combine these layers:
- Foundation: Use a commercial model API or a model hosted on approved infrastructure.
- Data layer: Build a search layer of your own that lets the model look things up in your documents, databases and internal knowledge before it answers.
- Integration layer: Build the connectors to your existing tools — Slack, Teams, ERP, CRM, email.
- Interface: Build a custom UI or chatbot that fits your team's workflow, not a generic chat window.
Compare total costs across the options, including integration, hosting, model usage, human review, and maintenance. Use that full comparison to choose the architecture.
A Decision Framework
For each AI use case, document these questions and the evidence needed to answer them:
Source access
Check supported connectors, permissions, and source updates in existing tools. Consider custom work when a required source or access rule cannot be supported.
Workflow fit
Test representative tasks, exceptions, and staff review with existing tools. Consider custom work when a tested gap prevents the required workflow.
Operating controls
Check retention, auditability, hosting, and ownership for each option. Consider custom work when a documented requirement remains unmet.
Total cost
Include licenses, setup, review, support, and usage in the comparison. A custom alternative needs to justify its build and maintenance effort.
Investigate each unmet requirement, compare the alternatives and choose a focused scope. Use the build-versus-buy decision worksheet to record assumptions, tradeoffs and the questions to resolve before purchase.
What changes the implementation effort?
A license price or a model's usage rate is only part of the decision. Ask each option to account for the same work:
- Source preparation. What needs to be collected, cleaned, approved or kept current? Who owns each source? Resolve unclear permissions and contradictory documents as part of source preparation.
- The integration boundary. Reading one approved source differs from updating several business systems. Identify authentication, allowed operations, field mapping, exceptions and recovery for each connection.
- The consequence of a mistake. A draft for a person to review has a different acceptance plan from a system that sends a customer message or changes a record. Keep consequential actions behind an agreed approval boundary.
- Evaluation and review. Budget the effort to prepare representative examples, compare options, examine failures and measure the review work left for staff. Use a representative set of normal and difficult cases to judge quality.
- Deployment and ownership. Name the environment, source-code and account ownership, release process, documentation and responsible operator. Specify the deployment and handover work in the first delivery.
- Ongoing operation. Identify licenses, hosting, model usage, monitoring, source updates and support separately. Agree what happens when a vendor, model, source format or connected system changes.
Use these scope drivers to request comparable estimates. Ask for the deliverables, recurring charges, responsibilities and work covered by each option.
Compare a small purchase before a large commitment
For each plausible option, define the smallest delivery that would change the decision. That might be a configured tool trial using approved examples, an integration feasibility check or a focused prototype. Write its acceptance criteria before implementing it.
Agree which inputs are permitted, what a satisfactory result looks like, what must be refused and how a person reviews uncertain output. Record results that fail as well as those that pass. Compare the remaining staff work with the current process; do not count an automated step as a saving if the checking and correction consume the benefit.
Our owned software walkthrough shows an implemented rules-based example: a fixed request, an exact source excerpt and a visible review or stopping state. Follow the inputs and decisions directly.
Choose the next step, including defer
- Buy or configure when a tool meets the important requirements and its access, costs and exit arrangements are acceptable.
- Integrate when an existing product covers the core task but a defined connection or handoff is missing.
- Build when a demonstrated requirement justifies custom software and there is a viable ownership and maintenance plan.
- Defer or simplify when the data, responsible owner, expected benefit or acceptance evidence is not ready.
Use the findings to make the next decision. Adopt a useful option, refine a promising one or stop before an unsuitable project becomes a larger commitment.
Apply the decision to one workflow
For an internal knowledge assistant, compare tools on representative staff questions, source support, and access tests. For workflow automation, compare existing integrations and ordinary rules with AI on the same inputs, including review effort and exceptions.
If the decision is still unclear, an automation roadmap ranks up to three processes by time saved against cost and gives you a 90-day plan with a fixed price for the first build. If the workflow is already defined, explore AI implementation or custom software development directly.
The engagements page explains project work, ongoing technical capacity and reduced-upfront arrangements. To choose a structure for the work in front of you, tell us about your project.
