The Reason LegalTech Said No: Your Product Reduces Their Billable Hours

Lisen Kaci taught himself to code backwards. He graduated in economics, went into sales for a year, did not enjoy it, and used the sales skills he had just acquired to win website-building clients — before he fully knew how to build websites. Then he had to learn fast enough to deliver. Necessity first, competence second.

That path led to engineering work in generative AI, including a lead engineering role building AI capabilities for medical document review, which is where he first saw the shape of the problems in document-heavy industries.

His first attempt at building on that insight failed, and the reason it failed is the most valuable thing in this conversation.

In Episode 129 of InsurTechTalk, Lisen and I covered why the legal industry quietly rejected his product, how Discrepancy AI generates hundred-page reports without anyone writing a prompt, what AI tool chaining actually means, and why he now sells before he builds.

About Lisen Kaci

Lisen Kaci is the co-founder and CEO of Discrepancy AI, which automates report generation for insurance companies and insurance consultancies. The platform ingests hundreds or thousands of pages of source documents — policies, price tables, drug formularies — and produces structured reports that can themselves run to dozens or hundreds of pages. Previously a lead AI engineer building generative AI for medical document review, Lisen studied economics and became a self-taught developer. At the time of recording the company had roughly three large customers and about $10K MRR, having pivoted out of legal tech.

The Failure That Redirected the Company

Discrepancy AI originally targeted legal professionals. It did not work, and the reason was structural rather than technical.

• Lawyers were reluctant to adopt a tool that would inevitably reduce their billable hours
• Nobody would say this openly — it is not something a firm admits to a vendor, since it amounts to saying they would rather not save the client money
• Lisen only uncovered it by talking to family friends and acquaintances inside the industry
• His generalized lesson: your product has to solve the problem for the company and for the individual employees who will actually use it. If it threatens the people whose adoption you need, the technology quality is irrelevant

I pointed out that this pattern is not unique to law. The same dynamic appears with TPAs in insurance, where efficiency gains cut directly against a billable-hours model — the incentive is to put someone on a plane and clock every minute, not to compress the work.

What Discrepancy AI Builds

The product targets a specific, unglamorous, enormously time-consuming workflow.

• Customers are insurance companies and insurance consultancies — the latter often working on a model where they negotiate renewals and get paid from the savings they generate
• Both produce standardized reports repeatedly: renewal reports, benefits reports, employee renewal rates
• The pattern is always the same — take an existing report format, apply new data, produce the new version
• Until recently this was done 100% manually, because the documents run to hundreds of pages

The technical obstacle Lisen describes is real and often understated: current LLMs have limited input context, so they cannot ingest very large document sets, and limited output context, so they cannot produce very long outputs. Real business work routinely involves thousands of pages in and an eighty-page report out.

The Insight: Stop Making People Write Prompts

The differentiating feature came directly from talking to the people using the software, not the executives buying it.

Customers told him something deflating: instead of spending an hour locating data and assembling a CSV, they were now spending 45 minutes fiddling with prompts to get output in the right format, shape, and table structure. A little time saved, but not remotely what they had expected — they were just doing a different chore.

Discrepancy AI’s answer removes the prompt entirely:

• Upload an old report as the example, plus the new source files
• The agent replicates that report’s style and structure with the new data — no prompt required
• Roughly 98% of companies producing reports already have exactly these two ingredients

Lisen’s analogy for why this works better than prompting is the game of telephone: every step of translating what you want into instructions introduces drift. Showing the AI the actual target artifact eliminates the drift.

The platform pairs this with a collaborative canvas — comparable to Figma or Canva in spirit — where multiple people and the AI work on the report together, with human-in-the-loop review throughout. Every extracted figure references the exact point in the source document it came from, which is how they keep hallucination in check.

AI Tool Chaining

Lisen is refreshingly deflationary about AI terminology. His definition of an AI agent: a model running in a loop. When people say a model is “thinking,” what is happening is that its own output is fed back into it so it can look at what it just did and go again.

Discrepancy AI’s own approach extends this across models:

• Multiple AI models are used internally, with an internal ranking of which model is best at which task
• Responses chain together — one model extracts data from documents, another standardizes it, another analyzes it
• His analogy: different people are good at different things, and working together they exceed the sum of their parts
• The system works through the report one section at a time until complete

Notably, they are not building on OpenAI, using a combination of other models instead.

On fine-tuning, he was candid that it is genuinely contested right now: you can spend months collecting data to fine-tune a model, and within that window a new general model may be released that outperforms your fine-tuned one. He also noted, with appropriate humility, that nobody fully understands why running a model repeatedly over its own output produces better results.

Selling Into Companies That Guard Their Data

Insurance buyers are cautious about data, and Lisen’s approach is to minimize what he holds rather than argue about security.

• Store as little customer data as possible, and do not retain it unless required
• Integrate with the cloud document providers customers already use, rather than becoming their storage
• Offer on-premise deployments for customers who need that additional layer
• Many customers have explicit agreements that their data will not be used for training

He is straightforward with those customers about the tradeoff: he has no objection to not training on their data, but the system then cannot improve on their data. That is an honest framing rather than a sales dodge.

Importantly, none of this touches the proprietary underwriting or actuarial data that actually differentiates a carrier. The system learns report formats and structures, not risk models.

Founder-Led Sales, and Building With Customers Present

The process change that followed the LegalTech failure is the practical takeaway for other founders.

• Get a customer to pay first, then build the solution for them
• Because he can both run the sales conversation and write the software, he can sit on a call, understand what someone needs, and go build it
• Iteration cadence with early insurance customers was every three days to a week: show what exists, watch them use it, take feedback
• His view is that this is the only viable approach for an early startup, because internal bias is the biggest killer of startups
• Unless you are Steve Jobs, building from your own head with no customer engagement fails

He also described the B2B sales reality plainly: find your champion — usually not the decision maker, but the person whose job your product makes easier — and that person carries the solution into the organization, followed by a six-to-eight-month cycle to close. In legal that was brutal; in insurance it has been substantially easier, and most growth has come through referrals, because it is a small community where people talk.

On the AI Naming Question

Asked whether a company needs “AI” in its name, Lisen had a sharp analogy. In the late 1990s every new company was an internet company and said so. Today every company is an internet company and none of them mention it. He expects AI to follow exactly the same path: universal, and therefore no longer worth advertising.

Key Takeaways

• A product that threatens the economics of the people who must adopt it will fail regardless of quality — billable hours killed the LegalTech version of this company
• You must solve the problem for the buying organization and the individual end user, and those are different problems
• Prompting can silently consume the time savings automation was supposed to deliver; replicating an example artifact removes the prompt from the workflow entirely
• An AI agent is a model running in a loop, and chaining specialized models beats relying on any single one
• Minimizing stored data and integrating with existing cloud storage defuses most enterprise data objections — at the cost of not improving on that data
• Sell first, build second, and iterate with the customer in the room every few days
• Referrals dominate growth in insurance because it is a small, talkative community