Microsoft’s researcher now grades its own homework with rival ais
Microsoft just handed its enterprise research agent a red pen and told it to mark itself—using the same models it competes against. The new twin-engine setup inside Researcher, part of 365 Copilot, pits OpenAI’s latest against Anthropic’s Claude in a closed-loop duel where one model writes the first draft and the other tears it apart before any user sees a sentence.
The feature, codenamed Critique, quietly rolled out to Frontier testers last week. It works like a peer-review session on amphetamines: Model A maps the query, scrapes internal SharePoint lakes, and spits out a raw memo. Model B—sometimes Claude, sometimes GPT-4o, depending on workload—then interrogates every claim, checks citations against the original corpus, and either stamps “verified” or forces a rewrite. The cycle repeats until both sides sign off. Microsoft claims the loop lifts factual accuracy by 18 % and cuts hallucination rates almost in half, according to early telemetry.
Two models enter, one report leaves
Enterprise customers have seen this movie before: vendor promises “deeper insights,” delivers a bloated summary larded with footnotes nobody asked for. The difference here is the friction. Critique adds latency—up to 45 seconds for a 20-page brief—but that delay is the product. Microsoft is selling hesitation as a service, betting that compliance officers will trade speed for an audit trail that shows exactly which model objected to which sentence and why.
If that still feels like black-box roulette, there’s Council: a split-screen mode that runs the identical prompt through Anthropic and OpenAI in parallel. The pane spits out two full reports, side-by-side, with deltas highlighted in real time. The first time I saw it, the Claude branch unearthed a 2021 antitrust filing the GPT branch missed; GPT countered with newer EU regulatory guidance that Claude hadn’t ingested. Users can pick either, blend both, or—more likely—forward the cage match to legal and let them sort it out.
Microsoft won’t say which model wins more often. Internally they track “challenge rate,” the percentage of times the reviewer forces a rewrite. Last month Claude challenged GPT 62 % of the time; GPT pushed back on Claude 57 %. The numbers are close enough to keep product managers awake—proof that model diversity isn’t academic, it’s defensive.
Autonomy is the next battlefield
While Researcher polishes its own prose, the company is already beta-testingCopilot Cowork, an agent that doesn’t just answer questions—it executes. Think: scan your calendar, cross-check attendee tenure, auto-generate a pre-read, Slack the deck, book the follow-up, and file the expenses. Early adopters inside Accenture claim the tool shaved 40 minutes off each client-prep cycle. The catch: Cowork demands write access to SAP, Workday, and Salesforce. IT departments are suddenly being asked to trust an AI intern with the keys to the ERP kingdom.
Microsoft’s pitch is that the same dual-model safeguard runs under the hood. Every autonomous action gets a “shadow pass” from a second model before execution. If the models disagree, the workflow pauses and pings a human. It’s a clever insurance policy, but one that only works if the humans still read their alerts.
The subtext is hard to miss: Google’s NotebookLM may generate slick audio summaries, but Microsoft is weaponizing disagreement. Redmond is betting that in regulated industries—finance, pharma, energy—the winning pitch isn’t speed or even accuracy; it’s cover-your-ass granularity. When the SEC comes knocking, you’ll want a PDF that shows Claude flagged paragraph 3, footnote 7, and the exact timestamp.
Microsoft hasn’t priced either feature yet; both remain gated behind the Frontier program for “select” E5 subscribers. Expect a 30 % surcharge when it hits general availability this fall. For now, the company is happy to let rivals chase chatbots that sound smarter. While they do, Microsoft sells the only AI that argues with itself—and keeps the receipts.