Research that holds up

No model looks up a market size the way you do. It generates the shape of a market-size statistic - a currency, a magnitude that fits the industry, a growth rate that rhymes with everything else it has read - and fills the shape plausibly. That is why invented numbers are not detectable by inspection: they are optimized, in a literal sense, to look right. The tell is never in the sentence. It is in the absence behind the sentence, which only a source-check reveals.

The invoice for skipping that check already exists. In October 2025, Deloitte Australia partially refunded the Australian government on a A$439,000 report after an academic found fabricated references and a misattributed court quote - errors Deloitte attributed to generative AI use it later disclosed. The refund was the small cost; the headline was the real price, because a consulting firm's product is the client's ability to trust what it hands over. Note what the incident did not require: bad strategy, wrong conclusions, or a hostile client. Unverified production details were enough.

Do not let the rule curdle into refusing AI research help - that forfeits real gains for imaginary safety. Ten leads an hour, three of which survive the gate, still beats two hand-found facts in the same hour. The rule changes the accounting, not the tool: leads are free, facts are expensive, and only facts enter deliverables. Even retrieval-based tools that return citations get the same treatment - a citation is a lead to a source, not proof you have seen it. Click through. Every time.

Not all sources earn the same confidence. Primary - filings, transcripts, papers, the client's own data room - is what "verified" means; the log's first worked row ("consultants with GPT-4 were 25.1% faster - HBS working paper, verified") points at the study itself, not at coverage of it. Reputable secondary - major press - is acceptable with the outlet named. Vendor and self-reported numbers get logged, used, and labeled in the deliverable: "McKinsey claims ~30% time saved - vendor claim, label it" is the second worked row. The claim may be true; the point is the client can see whose claim it is. Estimates - yours or AI's - carry the same obligation.

Inside that discipline, AI has real research jobs. Coverage sweeps: "here are my 8 claims for the Cadence engagement - what angles am I missing?" catches the branch nobody staffed. Summarizing documents you supply: feeding it the actual filing and asking for the pricing section is inside the frontier, because the source is in your hand. Drafting search strategies: where to look, what terms, which source class each claim probably needs. And retrieval tools that return citations still get the log treatment - the citation goes in the source column only after a human has clicked it and seen the page.

This session is the consulting discipline: what counts as a fact, who owns it, and how it is labeled when it reaches a client. The machinery of deep research - grounding techniques, retrieval pipelines, multi-source synthesis, research-agent workflows - lives in learn-ai-research-with-phoebe. Take that course to get faster at producing leads; keep this session's log to decide what survives the gate. The two compose: better machinery fills the log faster, and the log keeps the machinery honest.