AI in Legal: Between Hallucinations and Hard Reality

Anupam Kundu
Oct 14
3 min read

In early 2025, Deloitte made headlines for the wrong reasons.One of its government-commissioned reports contained fake legal citations and references to academic work that didn’t exist — all generated by AI. The firm had to refund the contract partially.

Embarrassing? Absolutely. But more than that, it was a warning.

If one of the largest consulting firms in the world can be blindsided by hallucinations, what does that mean for startups, law firms, and corporate legal departments that are eager to plug AI into their daily work?

The scale of legal spending

Legal isn’t a side function. It’s one of the big-ticket cost centers in any large organization.

In the US, companies spend about $2.6 million per billion dollars in revenue on legal. For a Fortune 500 company, that means tens — sometimes hundreds — of millions of dollars a year.
In Europe, CAC 40 and FTSE 100 companies report similar proportions, with annual legal budgets often stretching into the hundreds of millions.
In France, while data is patchier, sector studies suggest that large companies typically spend between 0.2% and 0.6% of revenue on legal.

For CFOs and boards, this isn’t a rounding error. Legal costs are significant — and they’re under pressure.

The pressure cooker inside legal departments

The 2025 Legal Department Operations Index highlights a paradox:

45% of legal teams say they are under-resourced.
55% report flat or shrinking budgets.
Yet 73% say they plan to adopt advanced tech to automate tasks.

In other words, legal departments are being asked to do more with less — exactly the conditions where AI looks like a savior, but can just as easily disappoint.

Where AI is actually helping

Despite the hype, there are clear areas where AI has added value:

Contract Review and Due Diligence

Tools like Kira Systems and Luminance can scan thousands of contracts, highlight unusual clauses, and cut due diligence from weeks to days.
Lawyers still need to verify, but the time savings are undeniable.

Legal Research

In France, Doctrine.fr makes jurisprudence searchable through AI.
In the US, Casetext — bought by Thomson Reuters for $650M — uses large language models to accelerate case research.
These tools succeed because they work in bounded domains where citations can be verified.

Litigation Analytics

Companies like Predictice (France) or Jus Mundi (international arbitration) provide data-driven insights.
For insurers and corporates, predicting case outcomes isn’t theoretical — it has real bottom-line impact.

Where AI is failing

The weak spots matter just as much:

Consumer “Robot Lawyers”

DoNotPay promised to fight parking tickets and handle small claims. Instead, it faced lawsuits itself, accused of unauthorized legal practice and sloppy advice.
Courts rarely accept filings generated by these tools.

Unsupervised Drafting

Generating full briefs or memos with AI sounds tempting — until a fake case or misquoted precedent slips in.
Deloitte’s misstep proves the point: even seasoned consultants didn’t catch the hallucinations. For law firms, that’s a non-starter.

The bottlenecks

Why is AI struggling to break through in legal?

Trust: One bad citation can sink a filing and expose a firm to liability.
Regulation: Bar associations in France, the UK, and the US remain cautious about AI tools giving legal advice.
Data Access: While legislation is public, annotated case law and commentary are often behind paywalls — expensive for startups.
Business Models: Consumers won’t pay much for “access to justice” apps, so scaling without B2B customers is hard.

Why innovation still matters

And yet, legal innovation isn’t optional.

Costs are unsustainable.
Workloads keep growing — more regulation, more compliance, more disputes.
Budgets remain flat.
Young lawyers burn out doing repetitive reviews.

Without new tools, the legal function risks becoming a bottleneck on corporate agility. With the right tools, it can become a driver of efficiency.

The AI Pathway lens

At AI Pathway, we look at adoption through three layers:

Vision: Cut costs, speed up processes, but maintain trust.
Strategy: Start with high-volume, lower-risk tasks — contract review, e-discovery, research — before touching sensitive litigation.
Enablers: Clean data, strong governance, and human lawyers in the loop. Without these guardrails, ambition becomes liability.

The Deloitte fiasco shows what happens when enablers are missing.

The reality check

So, is AI “working” in legal?

Yes, in document intelligence, research, and analytics.
No in consumer apps and uncontrolled generative drafting.

The winners are using AI to augment lawyers, not replace them.

The Deloitte case illustrates the risks. But the growing cost and pressure inside legal departments show why change is unavoidable.

The truth lies somewhere between hype and doom: a slow, uneven transformation.