For anyone with a data set and a question.
The wordmark, three prisms, and a verb. Pentagram tidied the blocks; the idea - that a human should still be doing the thinking - stayed exactly where it was.
Somewhere right now, a customer-support lead who has never opened a terminal is highlighting phrases in a spreadsheet of angry phone transcripts. By lunch, the model she is building will start flagging the calls that turn into cancellations. No data scientist was harmed in the making of this model. None was even invited.
That is the everyday scene Pienso is quietly engineering: the person who understands the problem is the same person who builds the AI for it. For a decade, the industry sold the opposite arrangement - the experts described the problem, and a separate priesthood of engineers translated it into code, usually slowly, occasionally wrongly. Pienso looked at that handoff and decided it was the bug, not the feature.
"Any domain expert, not just an AI engineer, should be able to train and fine-tune their model."
- Birago Jones, Co-Founder & CEOMost organizations are drowning in text. Support tickets, claims, transcripts, filings, moderation queues, intelligence reports. The knowledge to make sense of it lives in the heads of the people reading it all day. The tools to scale that knowledge lived somewhere else entirely - behind Python, behind GPUs, behind a hiring budget for talent that startups and agencies rarely win.
So the questions piled up unanswered. Not because nobody knew the answer, but because the answer needed a translator, and the translators were expensive, busy, and gone by Friday. The result was a strange standoff: the experts had the judgment, the engineers had the keys, and the data just sat there, accumulating.
"There's no replacement for human expertise, even in the world of AI."
- Pienso, company valuesThere was a second problem, quieter but heavier. The customers who needed this most - governments, broadcasters, pharma, finance - were precisely the ones who could not pour sensitive data into someone else's cloud and pray. Compliance is not a vibe. It is a wall. Any answer that required shipping the data out was, for them, no answer at all.
Birago Jones and Karthik Dinakar met at the MIT Media Lab, where the curriculum apparently encourages you to solve a real problem before graduating. Theirs was online bullying. They teamed up on a class project to help social platforms flag harmful content - and immediately ran into the standoff. The machine learning worked. The humans who understood the nuance of cruelty had no way to steer it.
The bet they placed in 2016, originally under the unglamorous name UBQ Labs, was this: keep the human in the loop, but give the human the controls. Not a wizard that hides the model. An interface that hands it over. Let the moderator, the analyst, the nurse, the investigator annotate the data themselves and watch the model learn from their judgment in real time.
"Pienso" means "I think." The product is the footnote: you think, the machine remembers.
- On the nameIt was a contrarian wager in an era that worshipped automation and treated humans as the thing to be removed. Pienso treated humans as the thing to be amplified. Lightly ironic, given the headlines, that the durable AI bet turned out to be the one that kept people in the room.
Birago Jones and Karthik Dinakar found the company (as UBQ Labs) out of MIT Media Lab research on content moderation.
$2.1M seed round brings in Eniac Ventures, Uncork, Indicator Ventures, and the MIT E14 Fund to help non-programmers touch the algorithms directly.
CogX recognition for NLP work; broadcaster Sky begins using Pienso to analyze customer-service interactions.
$10M round led by Latimer Ventures - with Gideon Capital, SRI, Uncork, Good Growth Capital, and GraphCore's Nigel Toon - pushes total funding past $17M.
A team of roughly 24 builds the "Garden of LLMs," scaling sales, customer success, and engineering for enterprise and public-sector accounts.
Pienso is a point-and-click interface for the whole arc of a model. You bring unlabeled text. You start annotating - teaching it, in your own words, what matters. It uses a semi-supervised approach, so a little human labeling goes a long way. You fine-tune. You evaluate. And then, the part the regulated buyers actually care about, you deploy: cloud or on-premise, your servers, your control, your data never leaving the building.
They call the model library a "Garden of LLMs" - a rare flash of whimsy in a field that otherwise names everything after plumbing. The pricing matches the philosophy: you pay for the models you actually deploy, which is a polite way of daring you to test before you commit.
Read calls, chats, and reviews at scale to find the sentiment and signals that move retention.
Train models on hate speech, misinformation, and harm - with human nuance built in, not bolted on.
Classify, summarize, and extract meaning from mountains of unstructured documents.
Surface the subtle signals that matter for compliance, security, and public safety.
The novelty isn't that the AI is powerful. It's that you can see what it learned - and why.
- On interpretabilityProof, in enterprise AI, is unglamorous and specific. At Sky, the UK broadcaster, Pienso reads customer-service calls to understand what is actually going wrong on the phone. At a US government agency, the same no-code platform helps monitor the tracking of illegal weapons. Different worlds, identical promise: the expert builds the model, and the data stays put.
Disclosed funding by round, USD millions
Bars scaled to the $10M Series A. The "total" bar laps the field because it counts rounds this chart politely declined to itemize. Sources: TechCrunch, FinSMEs, Crunchbase.
Behind the numbers sits a roster that reads like a vote of confidence from people who build chips for a living - GraphCore founder Nigel Toon among the Series A backers. When the hardware crowd bets on your software, it usually means the software is doing something the hardware can't fake.
Pienso's stated mission is almost suspiciously plain: put AI into the hands of the people who have problems to solve. No moonshot language, no promise to reinvent intelligence itself. Just a redistribution of who gets to use the most powerful tools - away from the few who can code them, toward the many who understand what they are for.
The culture underneath it is described in four words the company seems to actually mean: thoughtful, playful, purposeful, humble. It is a strange set of adjectives for an AI startup in 2026, an industry not famous for humility. Then again, naming your company after the verb "to think" sets a certain tone, and somebody has to be the grown-up at the party.
The bet was never that machines would think for us. It was that more of us would get to think with them.
- The throughlineAs models get bigger and the temptation to outsource judgment grows, the boring questions get louder. Who owns this model? Where does the data live? Can anyone explain what it learned? Pienso built its whole company around answering those before they were fashionable - ownership, sovereignty, interpretability - which positions it well for a market discovering that "trust us" is not a compliance strategy.
Go back to that support lead from the opening scene. A few years ago she would have filed a request, waited a quarter, and received a model she could neither inspect nor adjust. Today she built it herself before lunch, on her own data, on her own servers, and she can tell you exactly why it flagged the call it flagged. The translators are still welcome. They are just no longer the gate. That is the change Pienso is selling - not smarter machines, but more people allowed to think with them.