Democratizing AI is more than a login

Democratizing AI is often described as giving more people an account on a popular chatbot. That is access of a kind. It is not the same as making useful capability, control over data and knowledge, and human decision-making available beyond organizations that have staff and infrastructure.

This is not a new argument for Reclassify AI. It is a position about meaning, shared knowledge, and who gets to do serious work with information.

What we wrote then

In August 2019, in Smart Data and Smart Management (opens in a new tab), I wrote that I had founded Reclassify AI in response to how large organizations handled data: too much of it inaccessible, unreliable, and tangled in technical debt. One position in that note was that “engagement from a broad intersection of knowledgeable and skilled folks is absolutely essential for AI accuracy.”

In April 2020, in Knowledge Graph Sweet Spots (opens in a new tab), we argued that knowledge graphs could turn data systems into “flexible platforms for generating new information and even knowledge,” because their schemas “encode meaning formally.” One payoff was simpler querying: a business user or analyst could write questions closer to ordinary language than a traditional database query would allow.

In November 2020, in Semantic Data Platforms (opens in a new tab), we described what a semantic framework was not — not ripping out every database, not boiling the ocean overnight — and what it was. Two of the eight steps were about access: “find ways to make the data in your graph accessible to all users in the firm,” and “find ways to equip all data practitioners in the firm with the ability to add new taxonomies and data sources,” including educating people on the tools.

Those were early positions about knowledge and access rather than a forecast of chatbots. The through-line still holds: people need to know what their information means, what is authoritative, and how their expertise stays in charge.

Two halves of the same problem

Reclassify AI’s consulting work is the enterprise half of that view. Human-led AI starts with deciding what terms mean, which sources count, and who owns a judgment. Knowledge graphs and careful data work make those decisions visible. A fluent model that cannot tell an approved policy from a leftover draft is not a more open tool. It is a faster way to sound sure.

The other half is more practical. Useful AI should be available to people who do not have a development team or a private cloud stack. A consultant or accountant should not have to send a client’s documents to someone else’s servers simply to summarize them or organize working papers. When a workflow is configured so the model and the files stay on that computer, inference and document processing can remain on the device.

It is not automatic. Not every desktop application keeps the work local. Some fall back to a remote service when a task gets hard. Updates, licensing, or diagnostics may still use the network. A computer can be lost, copied, or subject to lawful access. Running AI on your own machine does not create a privacy or compliance guarantee, and it does not make every task worth automating. The person doing the work remains responsible for the output.

What this looks like in practice

Take a small accounting practice preparing a client’s year-end package. Some of that work may belong in a professionally appropriate hosted service, depending on the client, the engagement, and the terms. Other material is better kept closer. If that accountant can run a capable model on her own computer, she can ask for a first pass over documents she has decided should stay in the office. She still checks every figure and decides what goes to the client.

That example is illustrative, not a client story. The same pattern holds for an independent consultant reviewing drafts they already understand. The knowledge is theirs. The question is whether AI supports it or quietly relocates it.

A login and a privacy toggle are not the same as control. In October 2026, Max Eddy wrote in Wirecutter (opens in a new tab): “Opting out from model training should prevent human reviewers from seeing your chats, but your conversations will remain accessible to the company and to law enforcement.” Temporary chats, the piece notes, do not necessarily keep a company from accessing what you typed. Information shared with a chatbot also does not automatically receive the legal or medical protections people often expect when they speak with a professional. Turning off training is not the same as controlling who can see the material.

Hosted tools remain useful. Plenty of work does not require sending sensitive files anywhere. Democratizing AI should mean having a real choice, not pretending every hosted service is unusable or that every local setup is simple. A computer that can run a useful model costs money. First-time setup usually means downloading software and model files. Working without a constant connection still depends on that setup having been done. Digital literacy varies, and not everyone wants to configure tools. Local AI is not automatically the affordable path for every person or community.

What should be in reach

Communities and individuals should be able to bring their own knowledge — their terms, their records, their judgment — and choose how AI supports that work. Resource-rich organizations will continue to buy hosted platforms, and they should still decide what their data means. Smaller practices should have a path that does not require the same infrastructure, or the same willingness to hand documents over.

We use AI where it helps. Access is still incomplete if it only means an account and someone else’s record of what you typed. Capability, control of knowledge, and human decision-making have to travel together. That is what we were getting at in 2019 and 2020, in the language of data platforms and knowledge graphs. The tools have changed. The requirement has not.