Start with an operating question, not a benchmark
A portfolio report should help an operating partner decide where to investigate, where to limit exposure, and where a management team needs support. Request and token totals cannot answer those questions on their own.
Start with facts the companies can compare: accountable owner, AI cost, budget status, service health, rules in force, intended result, and quality of proof. Label the source of every measure. Market research is context, not portfolio evidence. A result that changed during AI adoption does not prove that AI caused the change.
Keep four evidence classes separate
Each measure in the portfolio view belongs to one of four clearly labeled sources.
- Gateway records: requests, provider and model mix, tokens, cost basis, budget use, failures, fallback, latency, and control coverage.
- Finance data approved by the company: provider, reporting period, and USD statement subtotal compared with gateway records. The portfolio view does not expose a raw invoice or approve an accounting entry.
- Management evidence: project, business purpose, accountable role, baseline, target, approved spend, period result, and quality of proof. The value record does not contain an uploaded document or evidence URL.
- Independent context: published research labeled with its source, sample, geography, reporting period, method, and publication date.
Respect company boundaries while producing a rollup
A portfolio owner sees the combined operating picture and only the company details granted to that role. Finance checks and business results require separate approval from the company owner. Basic portfolio access does not reveal statement amounts, differences, baselines, current results, or the names of accountable roles.
Provider keys, raw invoices, prompts, responses, customer identities, and payment details stay out of the portfolio report. If a company revokes access, its protected figures disappear from the next report. A portfolio policy can propose a stronger control, but it cannot silently weaken or replace a company's boundary.
Turn measurement into an operating queue
Put the clearest exceptions first: workloads with no owner, no hard budget, too much spend on one model, no fallback, an unhealthy credential, repeated provider failure, AI spending with no current business result, or a result backed by weak evidence.
For each exception, record the recommended action, expected effect, approver, completed change, and before-and-after period. Do not rank employees or treat request volume as productivity.