Research method

Demandbase account evidence, compared by operating model

Use a dated rubric to compare account research without turning feature lists into performance claims. The workflow keeps identity, observation, inference, and activation as separate states.

Account research operating model

Compare the operating model

Dimension Agent-native research Seat SaaS Manual analysis
Setup Install in working agent Provision workspace Write checklist
Task model Natural-language brief Filters and dashboards Analyst interpretation
Evidence Requested per result Provider-dependent Depends on notes
Human review Explicit activation gate Configuration-dependent Inherent but variable
Cost Runtime and connected sources Seat, credit, or platform Analyst time

Checked 20 July 2026. Unknown means public evidence did not confirm the dimension.

Task, result, evidence

SAMPLE

Structured result

Domain identity, ICP fields, dated signals, committee hypotheses, and exclusions remain individually reviewable.

Data behavior ledger

Sources

Config-dependent. Record which first-party or connected source supports each observation.

API keys

Verified practice. Use environment secrets, minimum permission, and separate test credentials.

Runtime

Inspect locally. Review package prompts, network calls, retries, outputs, and errors.

Retention

Config-dependent. Define access, correction, suppression, and deletion for every destination.

Run a known-account evaluation

Prepare confirmed accounts, exclusions, ambiguous identities, stale signals, and expected committee roles. Measure false matches, unresolved fields, source recency, review time, and correction work. A completed command confirms runtime progress; it does not validate data quality, intent, or permission to contact anyone.

For each account, record the domain resolution decision, observation timestamp, signal source, alternative explanation, committee-role evidence, and reviewer outcome. Compare methods on accepted evidence and operational effort rather than database size alone. If a field cannot be confirmed, preserve Unknown instead of substituting a confident inference.

npx -y @okki-global/okki-go-taroball