Select for the workload, not the feature list

Vendor and technology selection is a traceable decision under uncertainty. The goal is not to name the product with the most features or the architecture the evaluator personally prefers. It is to select an option whose verified capabilities, operating model, economics and risks best satisfy the workload's agreed requirements and the organization's constraints. A defensible recommendation shows its evidence, assumptions, rejected alternatives and conditions for review.

Begin with outcomes and constraints before looking at products. Record required user journeys, service levels, data characteristics, integration boundaries, security and privacy obligations, recovery targets, geography, accessibility, delivery date, budget, skills and support model. Separate mandatory constraints from preferences and future possibilities. Weighting everything equally hides priorities; inventing weights after seeing vendor scores simply launders a preferred answer.

Create a shortlist from requirements, not marketing categories. Compare a credible build option, buy option and status quo where each is plausible. Disqualify an option only against a documented requirement. Normalize scope, service tier, region, support, discounts, traffic, retention and growth assumptions before comparing. Feature presence is insufficient: verify limits, failure semantics, identity boundaries, auditability, export behavior, API compatibility, version policy and the capabilities actually included in the contracted tier.

Evaluate quality across the whole workload. Reliability, security, performance, operational excellence, cost and sustainability interact. A managed service can reduce patching and on-call work while adding dependency, quota and exit risk. A self-managed component can offer control while transferring upgrades, capacity, backup, incident response and specialist staffing to the buyer. A familiar technology may reduce delivery risk; an unfamiliar but strategically important option may justify a bounded proof of concept and training plan. Name these exchanges rather than collapsing them into a single unexplained score.

Model build versus buy over an explicit horizon. Include acquisition or engineering, migration, integration, testing, licenses, infrastructure, support, observability, security, compliance, staffing, training, incident impact, upgrades and eventual exit. Use demand scenarios instead of one point forecast, and show sensitivity to the assumptions that can reverse the result. Discounts and vendor credits are temporary unless contractually durable. Sunk cost is not future value, and a cheaper line item is not lower total cost when it creates more operating work or risk.

Treat supplier and software supply-chain risk as part of architecture. Assess the supplier, product, dependencies and delivery chain. Require evidence appropriate to impact: security practices, vulnerability response, provenance, subcontractors, support and incident notification, business continuity, data handling, compliance, roadmap governance and financial viability. Contract terms should cover service levels, data ownership, deletion, audit evidence, material change, breach notification, support escalation and termination assistance. A certification is evidence for a defined scope and date, not a substitute for workload-specific controls.

Make lock-in concrete. Inventory proprietary APIs, data formats, identity models, operational tooling, skills, commercial commitments and managed semantics. Estimate export volume, time, fees, transformation, dual-running, validation and business interruption. Portability is not automatically worth maximizing: abstraction adds cost and can erase useful platform capability. Spend on exit options in proportion to concentration risk and the cost of being unable to move. Test the highest-risk exit assumptions before signing, and preserve ownership of data, configuration, schemas and decision evidence.

Use experiments for decision-critical unknowns. A benchmark or proof of concept should state a hypothesis, representative workload, baseline, versions, success and stop criteria, raw measurements and limitations. Use the same conditions for competing options and disclose vendor assistance or special configuration. Test negative paths such as throttling, dependency loss, restore, regional constraint and quota exhaustion when they can reverse the choice. A polished demonstration proves little about production fitness.

Govern the decision openly. Include engineering, operations, security, privacy, finance, procurement, legal, accessibility and affected users in proportion to risk. Manage conflicts of interest and keep vendor claims separate from independently observed evidence. Record the decision in an ADR or equivalent: context, requirements, candidates, method, evidence, tradeoffs, residual risks, owner, approval and review triggers. Dissent and uncertainty belong in the record.

Selection does not end at contract signature. Validate implementation against the decision, measure service quality, unit cost, support outcomes, security events, roadmap changes and concentration risk, and compare actuals with the model. Review on material price, licensing, ownership, capability, regulation, workload or risk changes. Prepare renewal and exit evidence early enough to preserve negotiating power. The best selection is not permanent; it remains appropriate while its documented assumptions stay true.

A solutions architect is accountable for the operating reality after the purchase, not only the comparison spreadsheet. Confirm that the contracted tier, region, support path and identity model match what was scored; a lower SKU that silently drops audit export, restore granularity or private connectivity is a different product. Keep a dated evidence pack with the benchmark configuration, residual risks and the named owner who will run the first renewal review. If that pack cannot be produced, the organization bought a slide, not a decision.

Worked example: Pro $2k scored, Enterprise demoed

Mandatory gate: 30-day restore and audit export. Sales runs the PoC on Enterprise. Procurement signs Pro.

SKU	list / month	audit export	restore window	vs the gate
Pro (what we signed)	$2,000	no	7 days	fail — different product
Enterprise (what we scored)	$8,000	yes	30 days	pass
Pro + add-on later	unknown	maybe	maybe	not a modeled option

The $6,000 gap is not a discount. It is the feature the matrix already required. That table is the interview.