How much can a firm pay for its AI: the ROI calculation almost nobody does

The conversation about AI in the legal sector has spent two years focused on what platforms can do: how fast they respond, how many sources they consult, how secure they are. A column by Leopoldo Hernández in El Confidencial argues that conversation is no longer enough. The relevant question is different: how much economic value does AI actually generate, and how much of that value can the firm capture?

From social validation to investment analysis

Harvey and Legora have become the visible references of the new generation of legal AI, and the usual question is which is better. Hernández quotes his partner Patricia Villa Berger to describe how some firms decide: they buy because the "cool guys" in New York use them. In an industry built on reputation and peer behaviour, social validation weighs. But it is not investment analysis.

The providers' economics also matter. Harvey has passed through approximate valuations of 3, 5, 8, 11 and, in September 2026, 15.5 billion dollars. Those valuations embed enormous growth expectations, and it is the client firms who will have to produce the revenues that justify them.

The calculation almost nobody makes

The column's exercise is simple and useful. Suppose a platform saves 120 hours per biller per year, ten per month. Those hours do not automatically become money: part will be less time worked, part will allow faster responses, and only a fraction will turn into billable capacity, better prices or lower costs.

If a firm economically captures 35% of those hours, that is 42 hours a year. At a contribution of 300 dollars per hour, the potential value is around 12,600 dollars per biller per year. Requiring a minimum return of three times the investment, the maximum defensible price would be about 4,200 dollars per biller annually. In Latin America, a reasonable zone would be closer to 2,000-3,000 dollars, depending on market, rates and real ability to monetise efficiencies.

The practical rule proposed: below 3,000 dollars per biller, the investment can be defensible; between 3,000 and 5,000, the ROI must be demonstrated; between 5,000 and 8,000, the technology must transform processes, staffing or pricing; above 8,000, it would have to replace relevant costs.

Total cost, beyond the licence

The column stresses a point management committees usually discover late: analysing the licence price is not enough. Total cost of ownership includes implementation, integrations, security, knowledge management, training, administration and the future cost of switching providers. Vendor lock-in may become one of the great problems of the next stage of legal AI, because nobody knows who will dominate the market in five or ten years: Harvey and Legora lead today, but Microsoft, Alphabet, OpenAI or Anthropic can reshape the competitive structure.

For the Spanish market the reading is direct: the capturable-value math per biller looks more like Latin America's than New York's, and firms should invert the usual process — first determine how much value they can generate and capture, then define how much they can pay, and only at the end negotiate with providers.

The productivity paradox

There is a deep contradiction: firms built their economics around hours and now buy technology whose main benefit is reducing them. If a ten-hour job drops to six, the firm created four hours of efficiency; but if it keeps billing by the hour, it may also have cut its revenue. The ROI will lie not only in working faster but in capturing the value of working better, which forces a rethink of pricing, leverage, staffing and value proposition.

And a less discussed consequence: for decades young lawyers learned by doing the work AI is starting to perform — due diligence, research, document review, first drafts. Who will train the next generation, and with what method, is a question the sector has not yet answered.

Internal links

• Business development
• From saved hours to ROI: AI in law firms
• Account planning and commercial pipeline