Recently I watched a neat pattern emerge, one that turned curiosity into repeatable internal progress instead of slack chat chaos and half-baked pilots nobody owns. Big companies ask the same two questions, over and over, in different variants. The first is, what can generative AI do for us? The second is, what exactly are your day to day tasks, because frankly I have no clue what most of you do. Those two questions collide in one practical answer, datathons, run the right way.
Run badly, a datathon is a lab experiment that dies on the vine. Run well, it is an internal innovation ecosystem, a pressure cooker where the company’s messy, siloed problems meet cross-functional talent and come out with actual solutions you can scale.
Here’s the simple model that works. Business units come with data, often messy, sometimes unloved. The datathon team acts as a consultant and triage unit, we “act as sort of like data science consultants” to decide if the data can actually yield value. Then we throw in a mix of data architects, cloud engineers, developers, domain experts and data scientists for a focused sprint.
The point is not to produce a glossy demo for investors. The point is to produce a path, a realistic route from problem to production, or at least to a pilot you can fund. After the event you have either a working prototype, a clear reason why the data won’t help, or a technical block that can be funded as a discrete project. That last outcome is still useful, because it converts vague curiosity into a concrete ask.
A nice side effect, if you care about that sort of thing, is culture. People see colleagues solving problems close to their own work, they get inspired, and ideas spread horizontally. Datathons are the internal show-and-tell that actually matters. The showcases that come out of them are the sparks that move someone from “what does this even do” to “how can I use this next week?”
There are two levers you can pull to change adoption, top-down and bottom-up. Both matter. A board directive that says employees are allowed to use AI is necessary, but not sufficient. The other lever, the human one, is legitimacy. If a high-ranking executive sits in a live demo and uses the tool, legitimacy spreads. I loved the phrasing from an internal playbook, “if the boss of the boss of the boss is using this technology, then maybe it’s okay if I use it too.” It’s blunt, and it’s true.
Pair that with the grassroots energy from datathons and you get a flywheel. People try things, internal teams build showcases, leadership endorses the right stuff, training scales that knowledge. Suddenly adoption is not a mystery, it’s a repeatable process.
People will not use a tool they do not trust. After those early data leak scares, a lot of employees went underground, using consumer LLMs to do work with sensitive documents. That is a toxic mix of risk and denial. The fix is sensible and boring: create clear guidelines about what data can be used where, build safe internal tooling, and run training that teaches not only how to get good output but also what is compliant.
Training matters in two ways. One, help people get useful results quickly. We once spent a lot of time on “prompt engineering” because it mattered. Today, tools and integrations have evolved, but knowing how to shape a task and what data to hand a model is still key. Two, keep people curious. The companies who succeed are the ones who make it easy for employees to re-engage and explore new uses, rather than putting AI behind an HR memo and a policy PDF.
And always keep asking the political question, the one nobody wants because it’s awkward, “how do you regain trust in a system?” That matters when past experiences with tools have been bad. The answer is incremental, transparent wins, not PR spin.
We get carried away and throw LLMs at everything. That’s a mistake. If you need deterministic arithmetic, use a deterministic tool. As one pithy line put it, “don’t use it to solve problems which you can already solve with existing technologies.” Two years ago it was reckless to paste your whole Excel workbook into a chat box and call it done. Today, LLMs can be used sensibly as a translation layer or orchestrator, calling deterministic functions, writing a Python helper, or triggering a workflow in your stack.
Think of the LLM as the brain that decides when to call the calculator, not the calculator itself. That yields better reliability and a path to compliance, and it turns “random answers” into predictable results.
When teams say they want to scale, engineers picture load balancers and autoscaling. That is necessary, but the harder part is adoption. You can deploy a universal internal assistant, but if people don’t know what to ask it, or they fear data leaks, nobody will use it. So run the datathons, publish real showcases, create short training and guidelines, and then iterate.
A practical sequencing looks like this
Those steps are boring to write, powerful to execute.
If you want AI to matter inside a big company, stop treating it as an abstract technology and treat it as a problem solving engine, governed and grounded. Datathons, done well, are the engine room. They translate curiosity into prototypes, stale processes into new workflows, and most importantly, they provide the human connection that makes technology stick. If you want to try one, pick a real pain point, invite a cross-functional team, and tell the exec team to show up and try it live. The rest will follow.
Commenting Rules: Being critical is fine, if you are being rude, we’ll delete your stuff. Please do not put your URL in the comment text and please use your PERSONAL name or initials and not your business name, as the latter comes off like spam. Have fun and thanks for your input.
Join a growing community. Every Friday I share the most recent insights from what I have been up to, directly to your inbox.
