AI Research Architect — Evidence-Driven Research

Research Architecture: Project-Management Platform Selection

Decision need

Select a platform for a 120-person remote design agency for the next 12 months. This is a

research workflow design, not a live comparison: no vendors, sources, prices, or performance claims have been researched in this example.

Primary question

Which candidate platform best fits the agency’s documented workflow, security, collaboration, integration, accessibility, and total-cost requirements over the decision horizon?

Subquestions

1. Which workflows must the platform support (intake, planning, design review, client approvals, resourcing, delivery, and reporting)?
2. Which capabilities are mandatory, preferred, or out of scope?
3. What current security, privacy, identity, retention, export, and administrative controls are documented?
4. Which integrations are required, and are they native, partner-provided, or dependent on custom work?
5. What is the 12-month total cost under realistic seats, tiers, add-ons, implementation, and support assumptions?
6. What migration, adoption, accessibility, reliability, and exit risks need testing?

Scope lock

Included: platforms named by the buyer; current product capabilities; documented commercial terms; relevant independent evaluations; workflow-fit tests.

Excluded until specified: legal certification, negotiated discounts, undocumented roadmap items, and claims about the agency’s internal systems. Geography, budget, and current stack are unknown and must be requested before final scoring.

Source and inclusion plan

Primary: official product documentation, security/privacy pages, status history, accessibility statements, and dated pricing/plan pages. Record the exact page, date checked, and claim supported.

Independent: credible hands-on evaluations, public incident reporting, and research with disclosed methods. Record author, date, method, sample, and conflicts where available.

Internal: authorized workflow requirements, integration inventory, seat estimate, test scenarios, and user feedback.

• Exclude anonymous claims and undated marketing summaries from decisive scoring. Treat vendor statements as vendor-reported, not independently verified.

Comparison method

First label each requirement

mandatory,

preferred, or

informational with an owner and evidence standard. Score each candidate 0–3 per requirement: 0 = no evidence or does not meet; 1 = partial or high uncertainty; 2 = meets with documented conditions; 3 = meets with strong, current evidence. Keep confidence separate from score. A missing answer is

unknown, never a zero-quality claim. Do not average away a failed mandatory requirement.

Contradictions and uncertainty

Maintain a claim ledger with claim, source, date, source type, evidence excerpt or location, candidate, confidence, and status. Preserve disagreements—for example, vendor documentation versus a dated independent test—and investigate version, plan, or test-condition differences. Mark stale evidence when product terms may have changed.

Next steps

Confirm candidate names, geography, budget range, current stack, mandatory controls, and research deadline. Then collect dated evidence, run approved workflow trials, document assumptions, and produce a comparison with unresolved gaps and a recommendation proportional to evidence strength.