Interview
By
Tendem Team
“I Spent 10 Years Automating Humans Out. Here’s Why I Now Build Them Back In.”
A conversation about Tendem’s new MCP connector – and why the real fix for AI slop is a vetted human, one message away, without leaving your chat.
Dasha, product owner of Tendem’s MCP launch, sat down with Nikita, an ML engineer on the team, to pressure-test one idea: should there be a human inside an AI product at all? The exchange below is edited lightly for length and clarity.
Every chat assistant on Earth can now give you a fast answer. The hard part is trusting it. Look at the fine print under any of them and you will find the same line: this tool can make mistakes, please double-check responses. That sentence is doing a lot of heavy lifting. It means that the moment a task needs real judgment, taste, or expertise, you – the user – become the safety net.
Tendem’s MCP connector is built to remove that burden. It lets your AI tool call a vetted human expert to finish a job, right inside the chat you are already working in. To make the case honestly, we handed the microphone to the most skeptical person we could find: an engineer whose entire career has been about taking humans out of the loop.
Start at zero. What is Tendem in one breath?
Nikita: Tendem is a work platform where AI agents and human experts finish high-stakes tasks together. The AI does the heavy lifting – it understands what the customer actually wants, scopes the task, and breaks it into pieces. Then a real human expert finishes the part the AI could not, double-checks the sources and the quality, and signs off. So it is not a pure AI tool, and it is not a freelance marketplace. It is something in between, taking the best of both: the speed of AI and the quality of a human expert. The tagline is literally the job: less slop, more done.
You’re an ML engineer. Why would you deliberately put a human inside an AI product?
Nikita: That is the exact question I asked myself when I joined. I spent around ten years trying to automate humans out of everything, so I am a strange person to be defending this. But here is the answer. AI alone is fast and cheap, which is why it is so popular. It also still hallucinates and still breaks down on complex, judgment-heavy work. Bigger models did not make that go away – they made it subtler.
Dasha: Subtler how?
Nikita: The early language models, like GPT-3, sometimes produced almost incoherent text – words that did not fit together, so you could tell something was wrong. That is no longer the case. Today even the smaller models generate clean, well-formatted text. But it can now be wrong in a confident, nicely formatted way. It all sounds plausible, and the mistakes are much harder to catch. So a human in the loop is not a downgrade of the AI. It is what makes the AI’s output trustworthy. I stopped seeing the human as the weak part of the chain.
Paint the picture. Someone trusts AI to finish a task on its own – what goes wrong?
Nikita: You get a result that looks right but might not be right, and that is the trap. You cannot just trust it, because sometimes it is off. So you read it, verify it, fact-check it, fix the small bits, second-guess the parts you are unsure about. In the AI era, verification becomes the work. You delegate the easy 80% of the task and keep the hard, annoying 20% of double-checking for yourself – while being sold a finished product. You can even ask the AI “are you sure?” and it will say “yes, of course” and still be wrong.
In one line: AI without a human in the loop feels fast, but it is quietly slow.
Tendem now speaks MCP. For a non-technical reader, what does that even mean?
Nikita: The easiest way to think about MCP is a USB-C cable for AI agents. Years ago, every device had its own charger and it was a nightmare. Then it got standardized and life got easier. The same thing is happening in AI. External tools get exposed through MCP, and you connect them to your favorite AI agent so everything lives in one place.
Dasha: So in plain terms, what does it buy me?
Nikita: You keep a single chat open with the agent you already use – Claude, ChatGPT, whatever you prefer – and add connectors to it. With the Tendem connector, you hand off a task, a vetted human expert picks it up, finishes it, and the result comes back in the same conversation. No switching tabs, no copy-paste. That is the magic of being an MCP rather than a separate app.
Walk me through it. How does someone actually hand off a task?
Nikita: Three steps. First, you describe the task in plain language – English or another language is fine – the way you would explain it to a smart colleague. Second, when you submit it, the AI structures the task: it scopes, breaks it down, and organizes it. Third, on the Tendem side, a vetted human expert verifies the AI’s work, fixes it if needed, fills the gaps, and signs off before the result comes back to you. Calling in a human for help has never been easier.
Watch it happen: a real landing-page review
During the interview, Nikita ran a live task. A friend had built a landing page for a travel-itinerary plugin – put together with AI, aimed at an AI-native audience – and could not tell whether it was actually launch-ready. It looked fine. That is exactly the problem.
Inside Claude, with the Tendem connector already added, the request was simple: start the message with “Tendem,” attach the pinned design, and ask for a review of the design and copy plus a few draft strings for the Product Hunt launch. Claude asked permission to use the tool, uploaded the file to Tendem so a human expert could see it, negotiated a price with Tendem’s matching agent – under $30 for a design review, a copy review, and Product Hunt launch descriptions – and put the task in progress.
On confidentiality: your Claude conversation stays yours. Tendem only sees the parts Claude decides to share, and Claude is strict about sharing only what is relevant to the task – nothing more. A human expert then takes it from there, typically returning finished work in a few hours rather than seconds. As Nikita put it, if you want it back in seconds, use plain AI – getting it right is what takes a real expert an hour or two.
Once my task is running – how do I know the human on the other end is any good?
Nikita: It is definitely not a random person. The expert network is vetted, with a large quality-control machinery that is both automated and manual – people checking people, AI checking people. Do a bad job and you are removed from the platform, so the ones left are the ones doing it right. For each task there is a mechanism that picks the best performer, and none of this was built from scratch. Tendem is a product from Toloka, a company that has been in the data-labeling market for more than a decade and works with frontier AI labs that come to us for training data. Those labs hold very high quality standards. That is what we now offer everyone else through MCP.
Flip it around. What should you not trust AI to do alone?
Nikita: A simple rule: anything that needs human judgment, taste, or expertise benefits from a human. Research where the sources have to be right. Real work with data. A marketing strategy. Building a website. A presentation for an important talk. STEM problems where the answer is either correct or it is not. Any high-stakes work where being wrong has a cost.
Dasha: So if I don’t need perfection, I can run it with AI alone – but the second “this needs to be right, this needs to be verified” crosses my mind, that’s the line for handing it to an expert?
Nikita: Exactly. And if you are not sure, you can still create the task. The agents will talk to each other and come back with a scope. If the scope surprises you – if it turns out bigger than you thought – that is often a sign it is worth delegating, because it may be a field where your own expertise runs out.
More people are building their own agents. Why plug in a human as an MCP instead of writing more prompts?
Nikita: Because nobody actually wants to spend their time on configuration, heavy prompting, and debugging yet another agent. We pretend we do. We do not – we want the work done. And here is the gap: a reliable human expert is the piece almost every agent stack is missing. You can chain models together and spend days optimizing prompts and still have quality issues. With a human expert behind the MCP, you can guarantee a higher quality than any combination of models would give you.
Where is all of this heading?
Nikita: We are building a universal human layer – something callable by agents, or by people, in any situation where human judgment is needed. It is not really a website you visit, though a web version exists too. It is a set of tools your favorite AI agent can reach. You delegate a task in minutes, the price is known in advance so there is no surprise invoice – from a few dollars up for a more complex task – and execution runs from about an hour up to a day depending on the work.
Dasha put the finer point on the “why”: as long as every chat box still warns that AI can make mistakes and asks you to double-check, the goal is a future where the result simply gets things right. The human layer is what AI and agents call on for judgment, quality control, current facts, and taste – wherever real expertise is needed, across domains including STEM.
Your task: bring the one you wouldn’t trust AI to finish alone.
The task where you would normally roll up your sleeves and check it yourself, or go hunting for a second opinion. Hand it to a vetted expert instead and see what “actually done” feels like – without leaving your chat.
The rule is simple: when being wrong has a cost, or quality and originality actually matter, hand it to a human instead of babysitting your AI or managing freelancers. Everyone from everyday AI users to agent builders can plug the human layer into their tools today through an MCP.
