AI investment creates a familiar strategic choice with new technical uncertainty: build a system, buy a product or partner with specialists. The right answer depends less on general AI ambition than on how distinctive the workflow is and what the organisation must control.

Buy commodity AI capability, build the layer that differentiates the business and partner when the capability is strategic but speed or specialist depth is missing.

Many organisations will combine the three. A commercial model and cloud platform can support a custom workflow delivered with an engineering partner.

Define the capability

Name the user, workflow, decision and outcome. “Deploy an enterprise assistant” is too broad. “Help support agents draft grounded answers from approved policy with citations” can be evaluated.

Ask whether the capability changes why customers choose the company or how it operates uniquely. Strategic differentiation increases the value of custom control.

Map data sensitivity, integrations, latency, volume and consequence of failure.

Buy when the workflow is standard

Commercial products fit common tasks such as meeting notes, generic productivity and established platform features. They offer speed, support and shared improvement.

Evaluate real workflows, not the demo. Review data use, retention, residency, model changes, export, identity, administration and integration. Calculate total licences and change cost.

Avoid paying for broad suites when only one feature will be adopted.

Build when control creates value

Custom development fits distinctive workflows, proprietary data advantage, unusual integration or strict control requirements. Build the differentiating orchestration rather than training a foundation model by default.

Include product ownership, evaluation, security, monitoring and maintenance. A prototype is not the full cost.

Use modular providers and interfaces where portability matters. Do not create theoretical abstraction that delays user learning.

Partner for capability and speed

A partner can add AI architecture, data, evaluation, security and product delivery without waiting to hire every role. The organisation should retain business ownership, data control and decision rights.

Evaluate how the partner discovers the workflow, handles failure and transfers knowledge. Require transparent source, accounts, documentation, tests and exit plan.

ValueCoders is Vinove’s engineering company, while Agentra focuses on practical AI products.

Score six dimensions

  1. Differentiation: how unique is the capability?
  2. Workflow fit: how much adaptation is required?
  3. Control: what data, model and roadmap ownership matters?
  4. Capability: can the organisation build and operate it?
  5. Speed to evidence: which option tests value fastest?
  6. Lifetime economics: what is the three-year cost and switching risk?

Weight them by consequence. A low-risk internal tool and a customer-facing decision system should not use the same threshold.

Run a bounded proof

Test the highest-risk assumption with representative data and users. Measure task quality, adoption, complete cycle time, review load, cost and severe failure.

Avoid a proof that ignores integration and access; those may be the difficult parts. Decide in advance what evidence triggers scale, redesign or stop.

Preserve optionality honestly

Negotiate data export and termination. Version prompts and evaluations outside a vendor interface where practical. Use standards and modular boundaries at real change points.

Do not let fear of lock-in create a complex platform before product value exists. Optionality has a cost too.

The Vinove model builds focused companies and stays accountable for the long term. Apply the same lens to AI sourcing: choose the structure that can still be owned and improved after the first launch.

Build, buy and partner are not identities. They are allocation choices. Protect the distinctive layer, reuse the commodity layer and make operating ownership explicit whichever route you take.

Watch for hidden switching costs

Switching risk can live in stored prompts, proprietary workflow builders, embedded vector indexes, provider-specific tool formats and employees trained around one interface. During evaluation, ask how data, evaluations, logs and workflow definitions can be exported.

Test one realistic migration path before a large commitment. Can a representative prompt and evaluation set run on another model? Can knowledge sources be rebuilt from the system of record? Can identity and permissions move without recreating them manually?

Do not pursue perfect portability. Preserve the assets that encode business knowledge and quality evidence. Those are usually more valuable than the model endpoint itself.

Procurement questions specific to AI

Ask how model changes are communicated, whether customer data trains shared systems, which regions process data, how outputs and logs can be exported, and which security evidence is available. Confirm rate limits, outage behaviour, support and price changes at realistic scale.

Bring technical and workflow owners into procurement. Contract terms matter, but they cannot reveal whether the product actually fits the work or whether employees will trust it.