MIT’s NANDA initiative went looking for proof that generative AI was paying off inside real companies. It interviewed leaders, surveyed employees, and analyzed public AI deployments. The number that came back, published this August in its report The GenAI Divide: State of AI in Business 2025: 95% of generative AI pilots at companies are failing. About 5% are seeing any real return at all.
The report’s lead author, Aditya Challapally, named the pattern behind the number: “Almost everywhere we went, enterprises were trying to build their own tool.” Homegrown AI succeeds roughly a third of the time. Tools developed with outside expertise and tailored to a company’s specific workflows may be more effective than generic solutions. Tools that cannot adapt to existing workflows may still answer questions effectively but have a more limited impact on broader operations.
Imran Tariq read that finding and recognized his own argument in it. He’d made the same distinction from the inside of his own company: “LLMs memorize, but agentic AI learns.” Memorizing everything and learning what actually matters are not the same skill. MIT just measured how much that gap costs.
Stop Asking AI Questions. Start Giving It a Job.
Tariq and Jun Xiong co-founded Prime Movers AI on the same premise the MIT data ended up confirming: most companies aren’t failing at AI because the models are weak. They’re failing because they’re asking a search engine to do a workforce’s job. “You don’t win deals with clever prompts,” Tariq has said. “You win by recognizing patterns and responding with precision.” A chatbot that only responds to prompts can still be useful, but its role may be closer to an information-retrieval tool than an active part of a company’s workflow.
The Trouble Begins When Growth Depends on One Person’s Memory
For many founders, growth can eventually become constrained by how much information, decision-making, and day-to-day oversight one person can reasonably manage.
Judgment calls that go undocumented and processes that remain concentrated in a founder’s memory can create dependencies that become harder to manage as a company grows, similar to the broader challenge of aligning AI tools with the workflows they are intended to support.
Their answer isn’t to dump every file the company owns into a model and hope something useful surfaces. That doesn’t create a smarter system. It creates a slower one, sorting through a pile it was never built to hold. Their answer is what they call a second brain: not a pile, but a structure. Organized the way a company organizes its own departments. Organized the way DNA organizes itself into genes, each one carrying exactly one instruction and nothing more.
They break their own approach into four distinct systems — the breakdown they built together, and the one they run their own company on. A second brain, holding the brand voice, the business context, years of notes a founder would otherwise carry alone. A teammate, a single agent doing one job well, the kind that runs a set of Meta ad accounts or an entire content pipeline end to end. A workflow, where one input triggers a whole process instead of a single task: hand it a video idea, and it researches the topic, writes the script, writes the caption, and hands back editing notes, no further prompting required. And an automation, built once to run without anyone ever opening an AI tool to trigger it. Morning briefings. Competitor monitoring. Weekly reports. Built once, running indefinitely.
Build a Teammate, Not a Tool
Two of those four systems are where Prime Movers AI does its most specific work, together, and both are direct answers to what MIT found broken.
The first is a teammate they call the SOP librarian, built to solve the integration problem: the disconnect between an AI tool and the way a company actually operates. Its whole job is to become the memory of the company. Every SOP, every checklist, every meeting note and decision, every policy, every vendor detail, where the logins live — never the credentials themselves; every client onboarding step, every question that keeps coming up, every piece of tool documentation, every scrap of training material the next hire will need. Solve a problem once, and it’s captured. Get asked the same question twice, and it gets documented so nobody asks a third time. Most shared drives are where documents go to be forgotten. This one is built to keep them current, and to cite their source instead of guessing.
Xiong brings a specific, lived version of that problem to the build. Before Prime Movers AI, she led a bilingual property and lettings operation in London that managed a large client base, giving her firsthand experience with how disconnected tools can create challenges at scale. She built the process the AI had to fit, not the other way around, and she does the same work now across both U.S. and Chinese AI model ecosystems alongside Tariq, rather than defaulting to whichever one is loudest.
Find the One Thing Actually Holding You Back
The second is an automation designed to work across multiple models rather than within a single one. It can draw on tools such as Claude and Codex to help identify the constraint that may be having the greatest impact on a company’s throughput, rather than focusing only on smaller operational issues. It runs on a set schedule, similar to the morning briefing, so the process does not depend on someone remembering to initiate it.
The Payoff Isn’t the Information. It’s That None of It Gets Lost.
Tariq’s own answer to why so many AI initiatives stall traces back further than Prime Movers AI. He co-authored The Pyramid of Trust with Aimee Tariq, Lisa Fei, and Tyler Wagner, a book about what has to be true before people act on what they’re told. MIT’s data is the same argument in numbers: a tool nobody trusts to touch the real workflow never gets the chance to prove itself, and 95% of the time, it doesn’t get one. That instinct, built over many years when he was running Webmetrix Group and recognized in many media appearances, as well as contributing to TechCrunch, is what shaped the four-system model Tariq and Xiong built together.
A second brain built this way can become more useful over time as refinements are retained, reducing the need to repeatedly revisit the same information or processes. A founder’s memory has natural limits, and some institutional knowledge may be lost when team members leave. Built this way, the company’s memory can remain more consistent and accessible across the organization rather than depending primarily on any one individual.
MIT put a number on the gap: 95% of AI pilots fail, mostly for reasons that have nothing to do with the model. Tariq and Xiong built Prime Movers AI around closing that specific gap.
Leave a Reply