AI has changed the economics of research. Reading a hundred documents used to cost a week; now it costs a prompt. What it has not changed is where credibility comes from. A competitive intelligence deliverable is trusted because a named person verified the facts, made the inferences, and signed the labels. No model can carry that accountability, and pretending it can is how CI functions will embarrass themselves this decade.
So the useful question is not "should we use AI" or "can AI do CI." It is a staffing question: for each step of the workflow, what gets delegated to the machine and what stays with the analyst. This guide walks the full CI cycle, triage, scoping, collection, synthesis, drafting, and QA, and draws that line at every step. One principle governs all six:
AI gets breadth, speed, and mechanical checks. The human keeps definitions, labels, inferences, and anything that will be signed.
If you are new to CI, three frameworks appear throughout and take one sentence each. The Signal Triage Matrix decides which incoming signals deserve analysis, by impact and urgency. The Fact vs. Inference Ladder labels every claim by how much weight it can bear, from Verified Fact [F] down to Speculation [SP]. The CI Report Pyramid says a deliverable must climb from data to insight to implication to a recommended action.
Step 1: Triage. AI clusters the noise; the human places the quadrants
Hand to AI: the volume problem. Feed the day's inbound, alerts, filings, press releases, forum chatter, into a model and have it deduplicate, cluster related items, and produce one-line summaries per cluster. This is the work that used to eat the first ninety minutes of every morning, and models are genuinely good at it.
Keep human: the two triage judgments, impact if true and time-to-relevance. These depend on knowing your company's exposure, your quarter's decisions, and your leadership's blind spots. None of that is in the model's context, and a model asked to rate "importance" will rate coverage volume, which is precisely the bias triage exists to cure.
Example. A pump manufacturer's CI inbox catches 40 items overnight. The model collapses them to nine clusters and flags that six articles about a competitor's "major expansion" all trace to one press release. The analyst reads nine lines instead of 40 items, then places the expansion story herself: high impact, low urgency, scheduled deep dive. The machine compressed; the human decided.
Step 2: Scoping. The human owns the definition; AI attacks it
Keep human: the definition of the question. Market boundaries, units of measure, time windows, whose revenue counts. Every downstream number inherits these decisions, and they encode judgment about what the client actually needs. Delegating the definition is delegating the deliverable.
Hand to AI: the stress test. Once you have written the scope, have a model generate edge cases and ambiguities: does the definition include retrofits, does it double-count distributors, what adjacent categories could a reader assume are in. Models are excellent at producing the annoying questions a good reviewer would ask, in seconds instead of at the review.
Example. An analyst scopes "the European market for warehouse robotics, 2026, manufacturer revenue." The model returns twelve boundary questions, including whether software subscriptions count as manufacturer revenue and whether a robot built in Asia but deployed in Poland is in scope. Two of the twelve force a real revision. Ten minutes, one avoided rebuild.
Step 3: Collection. AI finds and extracts; the human opens the source
Hand to AI: breadth. Locating candidate sources, extracting every mention of a product line from five years of transcripts, pulling segment tables out of filings, summarizing a 300-page regulatory decision to find the ten relevant pages. Extraction and location are where AI multiplies a researcher most.
Keep human: verification and tier judgment. Two hard rules. First, nothing enters the evidence base as a Verified Fact until a human has opened the underlying document; a model quoting a filing is a Reported Claim from a fallible narrator, not the filing. Models fabricate citations that look immaculate, and the failure mode is not frequent, it is quiet, which is worse. Second, the human assigns the source tier. A model will paraphrase a template-factory market report and a specialist research house in the same confident register, laundering tier 5 content into your notes with the tone of tier 2.
Example. Asked for a competitor's stated capacity plans, the model returns five quotes with dates and document names. The analyst opens all five sources: four check out exactly, one attributes a real-sounding sentence to an earnings call where it does not appear. Four verified facts, one caught fabrication, and the check took twenty minutes. That ratio, high yield with a nonzero fabrication rate, is exactly why the opening-the-source rule is unconditional.
Step 4: Synthesis. The human makes the inference; AI red-teams it
Keep human: the inference itself, and its Ladder label. Connecting separate facts into a finding is the analyst's core act of judgment and the thing a signature vouches for. A model handed four data points will produce a fluent narrative connecting them; fluency is not corroboration, and a model's confidence level carries no information about truth.
Hand to AI: the adversarial pass. Give the model your facts and your draft inference and ask for competing explanations that fit the same evidence, plus what evidence would distinguish them. This is the cheapest red team ever built, and it directly generates the promotion tests that weak claims need.
Example. Four facts about a logistics competitor: warehouse leases in two new regions, a hiring spike in cold-chain roles, a partnership with a grocery chain, new temperature-monitoring patents. Draft inference: entry into grocery fulfillment. The model offers two rivals: serving one large pharma contract, or building capacity to resell. It suggests the discriminating evidence, whether the leases include food-grade certification. The analyst checks, confirms, and labels the finding a Corroborated Inference with a better foundation than before the challenge.
Step 5: Drafting. AI compresses; the human writes what gets signed
Hand to AI: structure and compression. First-pass ordering of the evidence, tightening a 90-word paragraph to 40, generating six candidate action titles for a slide, converting a report section into speaker notes. Mechanical writing work, high volume, low stakes.
Keep human: the top two Pyramid layers. Implications name what the finding means for your company specifically, which requires context the model does not have. Actions name an owner, a move, and a deadline, which requires authority the model does not carry. And every claim label in the final text is placed or confirmed by the analyst, because the labels are the signature.
Example. The analyst hands the model a verified evidence block and asks for five action-title candidates. Four are grammatical restatements of the topic; one contains a usable falsifiable sentence. She rewrites that one, sharpens the number in it, and writes the decision request herself. The model contributed a draft; the argument is hers.
Step 6: QA. AI runs the sweeps; the human owns Band A
The 10-point pre-delivery checklist splits cleanly along the same line.
Hand to AI: the mechanical bands. Sweep for template ghosts, placeholder text, and internal notes. Check that every figure carries an as-of date. Flag any chart whose axis does not start at zero. Reconcile units across the document and flag any number that appears twice with different values. Models are tireless at exactly the checks humans skip when the hour gets short.
Keep human: the career-killer band. Tracing facts to opened sources, rebuilding the arithmetic, verifying entities, auditing the claim labels. These are the checks where the checker's name is the point; a QA step exists to attach accountability, and accountability cannot be delegated to a system that cannot be fired, deposed, or embarrassed.
Example. Before a sizing ships, the model flags that a growth figure appears as 14 percent on page 2 and 12 percent on page 9, and that one chart implies a 2025 data point the text dates to 2024. Both are real catches. The analyst then rebuilds the central calculation herself in a fresh sheet, because that check was never the machine's to sign.
The pairing table, compressed
Across all six steps the same verbs sort cleanly. Delegate: cluster, summarize, extract, locate, stress-test, red-team, compress, sweep. Retain: define, place, verify, tier, infer, label, recommend, sign.
If a task's verb is in the first list, automate it without guilt. If it is in the second and you catch yourself delegating it, you are not augmenting the workflow, you are outsourcing the accountability while keeping the byline.
Two closing rules make the whole system safe. First, the asymmetry rule: AI failures in the delegate column cost you minutes; undetected AI failures in the retain column cost you the credibility discount that never comes off. Price the delegation accordingly.
Second, the disclosure habit: note in your methods what the machine did, the same way you note your sources. Analysts who hide the AI use the AI badly; the ones who declare it have usually built the checks this guide describes.
The machine reads faster than you ever will. It cannot be responsible for anything. Build the workflow on both truths at once.