Introduction
MIT's 2025 State of AI in Business report found that ninety-five percent of enterprise AI initiatives deliver no measurable value. About five percent produce real operational gains. MIT calls the split the GenAI Divide.
Law firms sit on the same divide. Most have tried a chat tool, many have been offered an AI feature inside something they already pay for, and few can point to an hour it gave back. MIT's explanation fits a firm as well as it fits a bank:
Most AI systems do not learn. They do not remember, adapt, or improve.
Since the report came out in 2025, the big chat tools have added memory. It changes less than it sounds. The report's three findings follow, each read from the firm's side of the desk, with what has and has not changed since.
The learning gap, seen from a law firm
In 2025, MIT found that the top barrier to AI adoption is not model quality or regulation. It is that the tools do not learn: they cannot keep context, follow a real workflow, or improve from being corrected.
The report named three problems. Each one has a version a paralegal would recognize, and the first has moved since.
The tool remembers you, not your cases
In 2025 a language model needed the whole story every time. The big chat tools have since added memory, so it is worth being precise about what it holds: a few facts and preferences about you, kept with your own account at that vendor. Not the lease, not the client, not which file is which, and nothing your paralegal can see. The matter is in there only if someone types it in, and what one person types in stays with that person.
The tool does not fit how the work is done
A case is not one workflow; it is forty small ones that change with every letter from the other side. A tool set up for the intake form breaks the first time a client sends photos instead of a PDF.
A smarter model is not a firm that learned
The models do get better, every few months, for everyone at once. That is not your firm's corrections sticking. "That is the landlord, not the tenant" is still said every week, because nothing the firm uses keeps it where the next person would find it.
MIT summarizes it plainly: most AI tools do not adapt, remember, or evolve, and that is what separates the organizations that get value from those that stall.
Users were asked to rate each issue on a scale of 1-10
Put the three together and the missing thing has a plain name in a firm: the case. Who the client is, what was asked for, what arrived, which file is which, and who is on the matter. A tool that keeps none of that starts every task from zero.
Why lawyers like ChatGPT and still don't trust it with the case
MIT found that employees prefer ChatGPT or Claude for daily tasks because these tools feel fast, flexible, and intuitive. People often trust their output more than the enterprise AI they were given. Lawyers are no different: a chat tool is a fine first reader of a demand letter.
The trust stops at the stakes. When the work is sensitive or stretches over weeks, close to ninety percent of the people in the study prefer a human colleague. In a firm that line is the case itself: the chat tool gets a paragraph to rewrite, the associate gets the matter.
"Would you assign this task to AI or a junior colleague?"
The reason is the same learning gap. A consumer tool keeps a few notes about you, not the case; it cannot follow how the firm likes things done, and needs the matter explained again every time. Trusting it with the case would mean doing the case twice.
MIT's conclusion is that AI helps individuals and fails to change organizations. A firm sees it every week: AI that helps with a sentence and never with a file.
What the firms on the right side of the divide do differently
MIT studied the small group that crossed the divide. They did not buy smarter models. They chose systems that:
- work inside the day's workflow instead of beside it
- keep memory from one use to the next
- take a correction and keep it
- get better with use
- change as the business changes
They avoided generic AI features and brittle internal builds. They put the AI where the work already was, with the context it needs already in place.
"What's missing is systems that adapt, remember, and evolve."
For a law firm, the place where the work already is has a name: the case. And in most firms, the case lives in an email inbox.
What this means for a law firm
An inbox is the pile MIT describes. An AI pointed at it has to dig past newsletters, spam, and forty other matters to find one client, and it starts from zero the next time. The tool is not the problem. The pile is.
This is why we built Kacti around the case, not around the AI. Every message and file a firm and its client exchange is kept by case and by sender, in the firm's own case folder, on the drive it already uses. An AI tool, your firm's own or Kacti's agent, starts inside the case: the client's messages, their files, and the people on the case, already in order. That is the memory MIT found missing, kept by the app instead of re-explained by the paralegal.
AI is optional. You choose when it works, and on which files and conversations. It can read the photos, screenshots and scans a client sent, put each one in its place in the case file, and draft your reply. Nothing it drafts is sent until a person at your firm approves it. Kacti's own agent works under a zero-retention agreement, so nothing it reads is stored or reused; a tool of your own works under your firm's agreement with that provider.
The five percent did one thing: they gave the AI the context and let the work stay where it was. For a firm, that is the case file, kept in order from the first message.
If you'd like to see how a case runs in Kacti, from the first invite to the day it closes, talk to us.
