A quick purchase can create a long commitment

This article uses an illustrative scenario and explicit calculation assumptions, not a CYTIZEN client result. Thomas, applications director, receives three proposals for contract analysis: an AI feature added to an existing service, a specialist platform and an internally developed service using a purchased model. Demonstrations look similar. Operating effort, data rights and the ability to change supplier do not.

The choice is not simply buy versus build. An organisation can buy the model, build orchestration and retain corpus responsibility. It can buy the application while developing SAP and CRM interfaces. Management must choose, layer by layer, what it will own, operate, control and replace. “Our solution is proprietary” is too vague to justify a budget or contractual dependency.

Separate five decision layers

Thomas describes the workflow: receive a contract, identify relevant clauses, suggest analysis, obtain legal approval and retain the approved version. He separates model, document retrieval, orchestration, business interface and existing-system connections. A matrix associates each layer with supplier, accessible data, support team, costs, export rights and outage consequences.

The model is not always hardest to replace. A clause taxonomy, approval history and accumulated integrations can create stronger dependency. Legal must preserve the distinction between suggested and approved text. Introducing an assistant must not implicitly change the signed-contract system of record. Each candidate explains what it can export and what cannot be reconstructed outside its product.

Differentiation needs evidence. If a product already covers the standard workflow properly, a wish for autonomy alone does not justify building an entire application. If specific rules create an advantage or essential control, the company must know where those rules live and who can change them. Management can buy the application while retaining an independent, documented rules layer.

Compare the same quality scope

Solutions receive authorised, de-identified contracts with business expectations. Candidates do not choose only documents on which they excel. The set includes poor scans, contradictory annexes, unusual clauses and out-of-scope material. Expected legal analysis provides the reference. Comparison covers meaningful errors, review time, correct citations and recognition of limitations.

The number of extracted clauses is insufficient. Extensive extraction that misses an important exception may increase legal workload. Thomas measures time to accepted analysis, corrections included. A slower generator that is easier to verify may be preferable. Results state test conditions, versions and limitations so that a pilot score does not become a universal promise.

Tests also cover what demonstrations omit: deletion, user revocation, result export, model change, connector outage and restoration of the last version. A supplier requiring exceptional intervention to export routine data must explain how exit will work. Migration feasibility cannot rely only on an assistance clause to be negotiated later.

Build total cost without invented savings

Three-year cost combines implementation, integration, licences or consumption, monitoring, support, recurring evaluation, security and exit. Scenarios distinguish low demand, expected demand and growth. Per-user pricing may suit frequent use but become expensive for a broad occasional-user population. Consumption pricing may be sensitive to long documents, retries and background processing.

A purely educational calculation uses option A: EUR 40,000 implementation, EUR 30,000 annual subscription and EUR 10,000 annual support, totalling EUR 160,000 over three years before exit. Option B uses EUR 80,000 development, EUR 12,000 annual infrastructure and EUR 22,000 annual maintenance, totalling EUR 182,000. If B requires unavailable internal capacity or A excludes interfaces, these amounts change. They are neither quotations nor market prices.

The comparison does not artificially monetise hypothetical benefits. Saved time is tested on comparable tasks and converted to capacity; finance distinguishes reassigned capacity from spending actually avoided. Sensitivity analysis identifies what could reverse the choice: volume, change frequency, human-review cost or internal-team availability. A transitional purchase can be reasonable if building becomes attractive only beyond an uncertain threshold.

The contract must reflect the architecture

Negotiation covers data, results, configuration and logs. Who owns a configured clause library? In what format can human corrections be exported? Is training on documents clearly excluded or bounded? Subprocessors, processing locations and deletion mechanisms are examined by competent functions. A commercial confidentiality promise does not replace reading the terms of the actual service.

Service commitments distinguish product availability from workflow completion. An accessible product with a failed document connector does not provide the promised outcome. Support response, model-change communication and regression procedures must fit the business. Where the offering and risk allow it, Thomas asks to retain a version or evaluate significant changes before release.

AI Act classification is separate from economics. The Commission publishes an updated timetable and differentiated obligations; purchasing a component does not automatically transfer all deployer responsibilities. The record identifies roles and refers legal decisions to specialists. It also preserves business continuity: during an outage, legal must work on authorised documents and versions without relying on the assistant.

In practice: the hybrid choice has a specific reason

The specialist platform adequately covers standard review, but the company must retain approval rules and SAP data. Management chooses a purchased application, documented integration and an internal register of validated decisions. This is not compromise for its own sake. Tests showed that rebuilding the document interface would cost more, while handing over the approval register would make exit too risky.

The decision includes triggers for review. Excess consumption, unusable exports or supplier changes degrading the reference set require correction or an alternative. The choice is not announced as permanent. A comprehensible architecture, checked contract and retained evidence allow it to be revised without reconstructing the entire argument.

What actually varies

In pharma, an application affecting a controlled process needs evidence and change management proportionate to its use; document extraction must not be treated as a quality decision without assessment. In finance, third-party and continuity requirements can strongly affect contracting. In smaller organisations, absent maintenance capacity may make buying wiser than apparently economical development. The choice must reflect what the organisation can operate, not merely what it can commission.

Keep the decision verifiable

The committee retains the compared proposal versions and commercial exclusions. Scope changes require a fresh comparison, preventing the original price of one option from later being compared with the expanded coverage of another.

Sources, method and limitations

Primary sources checked on 4 October 2026. Figures are hypothetical and options refer to no named supplier. The reasoning requires additional security, contractual and applicable-law assessment.