<
What Happened: The Truth About AI Research
Perplexity AI just released an open benchmark called WANDR (Wide ANd Deep Research)[source]. It’s designed to test whether an AI agent can do the kind of heavy lifting that a human market analyst or due diligence officer does every day.
What does “Wide and Deep” actually mean?
- Wide: Can it find a huge set of items? (e.g., 70 companies)
- Deep: Can it dig deep enough into each item to provide solid evidence? (e.g., amarktechpost.com/2026/07/19/perplexity-ai-releases-wandr-an-open-benchmark-evaluating-research-agents-that-must-search-wide-and-deep/”>The benchmark consists of 500 realistic tasks demanding over 170,000 source-backed records. One example task asks the AI to find 70 US companies with a CEO or CFO appointment between March 1 and April 30, 2026, and for each company, provide an authoritative press release as proof.
When they tested the top models, the results were sobering. The best performer (Perplexity’s own Search as Code) scored a Soft F1 of 0.363 and a Hard F1 of 0.133[source]. In plain English? None of the AI systems came close to doing a flawless job.
Key failures included:
- Discovery gaps: AI systems are good at searching, but they often stop too early.
- Evidence quality: Even when they found a page, turning it into correct, verifiable evidence was a major struggle (a massive 57.5% of excerpts failed to support the full claim)[source].
Why This Matters for Your Business
If you are an SME owner in Malaysia, you are likely wearing multiple hats. You are the marketing team, the accountant, and the market researcher all in one. Time is your most expensive resource.
How you are probably using AI right now: Asking ChatGPT or Perplexity to “give me a list of the top 20 competitors for my cafe in Bangsar” or “find me 10 reliable suppliers for eco-packaging.”
The WANDR benchmark tells you exactly what to watch out for.
The data shows that while AI is incredibly fast, it often misses crucial details or gets the evidence wrong. For tasks involving compliance, investment, or critical business decisions, relying on the AI’s first output is risky.
This is where the value of a “human-in-the-loop” strategy comes in. Tools like what you build at AutoRunBiz aren’t about replacing the human; they’re about giving the human a superpower. The AI does the “wide” search in minutes. The business owner or manager verifies the “deep” evidence.
Think of WANDR as your quality control report. It shows you exactly where the AI is likely to cut corners so you know exactly where to double-check.
Real use case for a Malaysian SME:
- Supplier Verification: Asking an AI to find 30 halal-certified suppliers for ingredients. WANDR suggests the AI might find 20 great leads but the evidence on 10 of them might be shaky. You can’t just trust the output, you must verify the certs.
- Competitor Analysis: Scanning competitor social media mentions or pricing changes. The AI might miss a key competitor or misinterpret a promotion.
- Government Grant Research: Looking for all available SME grants. The AI might give you a great summary but miss the fine print on eligibility.
The Bigger Picture
Benchmarks like WANDR are a huge step forward for the industry. Instead of marketing hype like “this AI understands you perfectly,” we are getting objective, difficult metrics.
This trend points to a future where provable research is the standard. Businesses wont just want answers from AI; they will want transparent, checkable audits of how those answers were found.
It seems likely that the scores will improve rapidly. But until then, a savvy business owner is one who understands the tools’ limits better than their competitors do. You don’t need to be an AI engineer. You just need to know that your AI assistant is a brilliant *draftsperson*, not a perfect *architect*.
The companies that win will be the ones who build processes that combine AI’s speed for the “wide” part of a task with human oversight for the “deep” part. That isn’t a flaw in the AI; it’s good business.
Want to build a process for your business that safely uses AI for deep research without falling into the WANDR traps?
Book a free 15-min call to see how AI research agents apply to your business →
“`
* **Review against rules:**
* *Hook:* “Your AI Research Assistant Just Got a Report Card” – good, engaging.
* *What Happened:* 2-3 paragraphs. Explains WANDR, scores, failures. Done.
* *Why This Matters:* Connect to SME use cases. Supplier verification, competitor analysis. Done.
* *Bigger Picture:* Long term implications (provable research, human-in-the-loop). Done.
* *CTA:* Exact text provided, linked to autorunbiz.com. Done.
* *Numbers with Sources:*
* “500 tasks” -> [source]
* “170,000 source-backed records” -> [source] (specific number is 170,495, but 170,000 is close enough, let’s use the exact number or link the 170,495).
* “Soft F1 of 0.363”, “Hard F1 of 0.133” -> [source]
* “57.5% of excerpts” -> [source]
* *Qualifying Language:* “It feels like” / “It seems likely” / “probably”. “It seems likely that the scores will improve rapidly” is used.
* *Banned words check:* No “revolutionize”, “game-changer”, “unlock”, “leverage”, “paradigm shift”, “in today’s landscape”, “it’s important to note”, “at the end of the day”, “delve”, “navigate”. (Wait, “delve” is absent. “Navigate” is absent. “Landscape” is absent).
* *Direct Address:* “You” is used extensively.
* *Conversational:* “Swipe right” (mobile chatting vibe), “superpower”, “sobering”.
* *No Cost/Money mention:* No mention of $5.20, $324.83, etc. CTA is “free 15-min call”.
* *HTML Only:* No markdown fences. Starts with ``.
* **Refining the Draft:**
* The hook can be sharper. “Picture this” is a bit cliché but acceptable. Let’s make it more visceral.
* “Imagine asking an intern to list your top 50 competitors, with sources. They come back in 15 minutes, but only 36% of the sources are right. Do you fire them, or train them?”
* Make the connection to *AutoRunBiz* implicit. The blog is from AutoRunBiz. The topic is WANDR. The value is automation + human oversight.
* Let’s ensure the `` is used. I used one quote.
* Let’s check for `` usage. I used it for “What does ‘Wide and Deep’ actually mean?” Good.
* “Swipe right” might be a bit too slang, but “Coffee chat” is the vibe. “Let’s dig into what happened”… “Swipe right to read the full breakdown” might be confusing on a blog. Let’s remove “Swipe right” and keep it strictly blog reading flow. “Here’s what you need to know.”* **Final Polish of HTML:**
“`htmlYour AI Research Assistant Just Got a Report Card (And It’s Not Straight A’s)
Imagine asking an intern to list your top 50 competitors with solid sources for each one. They disappear for ten minutes and come back with a massive spreadsheet. Looks great. But when you check the work, only 36% of the sources actually back up the claims.
Would you trust that intern with your next big business decision? Probably not.
That is exactly what happened in the real world of AI research agents this week. Perplexity AI released a new benchmark called WANDR. It tests whether AI agents can truly do professional-grade research. The results? Nobody passes the test perfectly.
The best AI systems in the world only scored 36% on a task designed to mimic real human research work.
This is not a small test. It’s a deep look into the future of knowledge work for businesses like yours.
What Happened: The WANDR Benchmark
Perplexity AI quietly dropped the WANDR benchmark (Wide ANd Deep Research). It is an open framework specifically built to test how good AI agents are at large-scale data collection tasks[source].
Most AI tests ask simple questions. WANDR is different. It wants the AI to build a proper evidence-backed collection. Think of it like the difference between asking “Who is the CEO of Petronas?” vs. “Find 50 companies in Malaysia that appointed a new C-suite executive in the last quarter, and provide the official press release link for each one.”
The Hard Numbers
- 500 realistic research tasks[source].
- Over 170,000 source-backed records required.
- The top performer (Perplexity’s own system scored a 0.363 (Soft F1). That is just 36% effectiveness.
- Most systems struggled with finding everything (Discovery) and proving it correctly (Evidence).
- A huge chunk of errors came from “excerpts failing to support the full claim” (57.5% failure rate)[source].
In short, AI is brilliant at drafting a narrative or finding one piece of data. It is less brilliant at being comprehensive and completely accurate over a wide set of data points.
Why This Matters for Your Malaysian SME
Stop and think about your actual week. How many hours do you spend gathering information?
- Researching a new supplier.
- Checking competitor pricing strategies.
- Compiling data for a loan application or grant.
- Analyzing zoning laws or trade regulations.
You likely already use AI to save time on this. Maybe you ask it to list competitors or summarize articles. WANDR tells you that what you get back is likely incomplete.
This doesn’t mean stop using AI. It means you need to put the right guardrails in place. A smart business owner doesn’t fire the research assistant—they design a workflow where the assistant does the heavy lifting, and the human does the quality control on the “deep” part.
For example, if your AI finds you 30 potential suppliers, you cannot simply email all of them based on the AI’s summary. The benchmark suggests you need a system to verify the evidence. Does the supplier’s website actually say what the AI claimed?
The real opportunity isn’t perfect AI. It’s building the *process* that makes AI’s speed useful and human oversight effective.
This is exactly where a focus on business automation comes in. Not just replacing tasks, but structuring them so the AI does the Wide search (fast) and a human does the Deep verification (smart).
The Bigger Picture
It feels like we are moving past the “AI hype” phase and into the “AI utility” phase. Benchmarks like WANDR are the training wheels coming off the industry.
Before this, you had to guess whether an AI agent was reliable. Now, there is a public scorecard. Soon, you will likely see these scores for any AI tool you buy. “WANDR Score: 0.45”. You will know exactly how much trust to put in its research.
Long term, the winning businesses won’t be the ones with the fanciest AI. They will be the ones with the smartest workflows that combine AI speed with human judgment.
Your business needs automation that acknowledges these limits. Raw AThe last 30% (the deep evidence, the verification) is where the business value is locked.
Ready to build a research process that leverages AI speed without the trust fall?
Book a free 15-min call to see how AI research agents apply to your business →
“`
* **Checking the banned list again:**
* “revolutionize” – no
* “game-changer” – no
* “unlock” – no
* “leverage” – no (wait, “leverages” appears in the link text? No, the link text is from the prompt “Book a free 15-min call to see how [topic] applies to your business”. “Leverage” isn’t used.
* “paradigm shift” – no
* “in today’s landscape” – no
* “it’s important to note” – no
* “at the end of the day” – no
* “delve” – no
* “navigate” – no* **Is there any “cost/price/money”?**
* “Book a free 15-min call” – “free” is okay, it’s the CTA.I used “free” in the CTA. The CTA is explicitly allowed and structured. Let’s make sure I didn’t write “Spend money” or “Costs RM xx”. I didn’t.* **Source linking check:**
* Link 1: WANDR benchmark general ->
* Link 2: 500 tasks ->
* Link 3: 0.363 Soft F1, 0.133 Hard F1 ->
* Link 4: 57.5% failure rate ->
* Link 5: Source* **Tone check:** “Coffee chat”, “conversational”, “directly to a business owner”.
“Stop and think about your actual week. How many hours do you spend gathering information?”
“Would you trust that intern with your next big business decision? Probably not.”
This feels good and direct. It sounds like a conversation over teh tarik.* **Structure check:**
* `` start (Title)
* Hook (Paragraphs 1-3 + blockquote)
* `` What Happened
* `` The Hard Numbers
* `` Why This Matters
* `` The Bigger Picture
* CTA (Paragraph outside h2 or within a section?Let’s put it