AI proposes, the analyst decides: governed AI in competitive intelligence

Generative AI can make competitive intelligence faster. It can also make it indefensible: a fluent paragraph with a number nobody can trace, presented to a portfolio committee as fact. The fix is not less AI. It is AI with a narrow job, a visible output and a human signature.

Three ways ungoverned AI breaks CI

Invented numbers. A language model asked how many Phase 3 trials are running in psoriasis will produce an answer. It may even be close. It has no query, no date and no source, so it cannot be defended or repeated.

Silent edits. Cleaning sponsor names and condition terms is real work, and AI is good at it. Done silently, it changes counts with no record of what changed, so last quarter's chart can no longer be reproduced.

Overconfident prose. Models write fluently in the language of certainty. "Competitor X is accelerating" reads like a finding; it is usually an inference from a date field.

One rule: AI proposes, the analyst decides, the system records

Flow · a governed questionDesign

Every step is visible; numbers come only from the query

A working example, on this website

The "Ask the registry" widget on this site is a small, real implementation of the rule. The AI does one thing: it reads a question into filters (condition, phase, status, sponsor class and breakdown), and those filters are shown to you. The count itself comes from the ClinicalTrials.gov API at the moment you ask, with the query time printed under the chart and a link to the same search on ClinicalTrials.gov.

Ask for recruiting Phase 3 psoriasis trials by country and you get a chart such as "23 recruiting Phase 3 psoriasis trials: China leads with 15" (queried 17 Oct 2026). Ask whether one drug works better than another and the widget declines: efficacy, safety, price and forecasts are not questions a registry count can answer, so it says so instead of improvising.

Boundaries · what the model is allowed to doDesign

Narrow job, visible output, human signature

AI may

• Read a plain-language question into explicit filters
• Propose a chart type that fits the data
• Propose a cleaned sponsor or condition name, with its source and confidence
• Draft a reading of a chart, marked as a draft

AI may not

• Supply a number
• Change a dataset without an analyst's approval
• Make a causal claim or an unsupported forecast
• Freeze anything into a pack without a signature

Three design choices that make it defensible

Numbers come from queries, never from the model. If a figure appears in a chart or a sentence, there is a query behind it with a date.

Every AI change is a proposal. A cleaned name arrives with its reference source and a confidence score. The analyst accepts or rejects it; accepted changes become rules that replay on the next refresh, and the raw record is kept.

Language is calibrated. Draft readings use estimative bands, from "almost certain" to "remote", in the tradition of the US intelligence community's analytic standards (ICD 203), and stay drafts until an analyst signs them.

Speed and defensibility are not a trade-off when AI does the reading and the registry does the counting. See how AI is governed in Comphas.

Run it yourself

Every number above came from a query you can repeat. Ask the registry the same question for your own indication; the answer is drawn live, with its source and query time.

Method and queries

The worked example uses the live Ask widget on this website (Cloudflare Workers AI for reading the question; ClinicalTrials.gov API v2 for every count). The psoriasis result was queried on 17 Oct 2026; repeating it today may give a different number.