- Every AI app can use the same frontier models. The model alone does not decide how good an AI tool is.
- The harness is critical: the data, tools, checks, and design built around the model.
- The user experience is critical: how humans and AI interact, collaborate, share information, and build trust.
- Life science needs a strong harness more than other fields: the data is specialized and complex, and the questions are too scientific for normal search engines to handle, like which labs use a technique in a particular way or which companies are working on a specific kind of assay.
When ChatGPT came out, I was blown away, like everybody else. But as the years (four! an eternity in AI!) went on and AI apps multiplied, one thing kept nagging me. Any developer can give OpenAI or Anthropic a credit card and get the same raw intelligence the biggest companies use. So why are some AI apps so much better than others?
Here’s a simple test. Ask a general AI chat app (Claude, ChatGPT, Copilot, or whatever your favorite is) for the work email of a proteomics core director at a mid-sized biotech. You’ll often get a polite refusal or a guess built from the company’s email format. Ask Tnuki, which runs on models from the same AI labs, and you’ll get an address from a real source, or a plain “not found”. ChatGPT and Claude are incredibly valuable, and I use them myself for hours straight every day (working on Tnuki!). But being great at one job takes more than raw intelligence.
Picture a genius locked in a room with no library, no pen and paper, and no internet. They won’t do real work as well as someone less brilliant who has the right resources and right tools, like books, data, training, a computer, etc. A frontier model, on its own, without access to the right resources, simply won’t perform as well. Especially if your industry demands specialized tools and data beyond your typical internet search.
What engineers call the harness
The equipped room has a name. Engineers call it the harness: everything that surrounds the model and decides
- what it can see: which databases, documents and search results it can reach,
- what it can do: which tools it can use, and how those tools are shaped,
- what gets checked before you see an answer, and by whom,
- how the result is shown to you, so you can check it too.
The vast majority of applications use off-the-shelf models, but the harness is where each team is building something custom, and in my experience it’s where nearly all of the difference comes from. Nearly all of our engineering at Tnuki goes into the harness.
Why life science sales needs a strong harness
Plenty of work goes fine in a general AI chat app: drafting a memo, summarizing an article, brainstorming. Life science sales is harder on AI, for three reasons.
The data isn’t on the open web. A lot of what a rep needs is private or licensed. Verified emails and phone numbers aren’t something a model learned in training, and clinical trials, grants and publications live in specialized databases with their own search rules. Tnuki connects to more than 40 of them, from PubMed and ClinicalTrials.gov to NIH RePORTER and openFDA, plus contact data checked across more than ten providers.
A wrong answer is expensive. A guessed email bounces, and enough bounces hurt your sending reputation. And if your email gets a scientist’s research wrong, they’ll notice right away. Nobody knows their work better than they do.
Each question has a right source. It’s a bit like keeping up with the news. If you only read one outlet, you end up in an echo chamber; someone who’s really well read knows which source to trust for what. How many trials a company is running is a question for the trial registry, not a web search, which tends to undercount. Who funds a lab is a question for the grant databases. We’ve built our own in-house search engine, index, and tools specific to life sciences, designed from the ground up for AIs to use, so that they are naturally routed to the right data sources, with the results presented in the best way for AI models to consume.
Catching mistakes makes the AI smarter
The figure at the top shows one of the checks we built. Before an email, a LinkedIn link or a phone number lands in a Tnuki table, plain code confirms it really appears in something Tnuki found. That takes no judgment, so we don’t ask an AI to do it.
When the check blocks something, it doesn’t just throw it away. It tells the AI what went wrong, so the AI can go back and copy the real address exactly, or leave it out.
Something that’s stuck in my mind from all this: when you design an AI system (model+harness+data) that reduces hallucinations, you’re directly increasing its effective intelligence. AI systems (and humans) perform better when they get feedback and reflect on where they went wrong.
People stay in the loop
No set of checks is perfect, so people watch the checks. Flags from Tnuki’s table research go to our team several times a day, with an instant alert if they spike. A reviewer AI from a different lab re-reads finished sessions looking for mistakes, and we read what it finds.
And nothing goes out without your approval. I don’t think AI should replace salespeople, and I don’t think that’s the right direction anyway. The best technology has always worked the other way: it takes the dull work off your plate, so you can spend more of your day on what people do best, which is understanding customers and helping them solve problems.
The model will keep changing. The harness is the product.
New models will keep arriving, and every app will get them on the same day. That’s good news, because everything built on top gets smarter. It also means the model won’t be what sets one tool apart. The harness will, and that’s where we spend our time.
Whatever AI tool you’re considering, test it the way our most careful users test Tnuki: like a new coworker. Ask it questions you already know the answers to, and check its work.
If you want to see Tnuki for yourself, try it free, no signup needed.