We conducted a series of conversations with in-house legal counsel at regional companies over the second half of 2025, asking a specific question: what would an AI research tool need to do for you to actually rely on it in your daily work? The responses were more consistent than we expected, and they centred on a concern that does not appear in most AI legal tool marketing materials: the ability to explain the answer to someone who is not a lawyer.
The business stakeholder problem
In-house counsel do not give advice into a vacuum. They answer questions from business teams, finance, commercial, operations, who need to understand not only what the legal answer is but why it is the answer and what the consequences of a different course would be. The in-house lawyer is a translator between the legal framework and the business decision.
An AI tool that returns a correct answer without a source puts the in-house lawyer in a difficult position when the business team asks "how do you know that?" The lawyer cannot say "the AI told me" with any credibility. They need to be able to say "this is Article 88 of the UAE Companies Law, which requires X, and the reason this matters for your transaction is Y." The citation is not just a professional verification requirement; it is the raw material for the explanation that the stakeholder conversation requires.
This came up in almost every conversation we had. One Legal Director at a regional real estate group put it directly: the time saving from AI matters because his team is small and the query volume is high. But if the AI output is not explainable, he has to re-research the question anyway before he can brief the CFO. The tool saves him nothing except a first draft he cannot use.
The trust threshold question
Several counsel articulated a threshold concept: there is a level of verification overhead below which they would adopt an AI tool into their workflow, and above which they would not. The threshold varies by person and by query type. For low-stakes internal orientations, a rough answer without sourcing might be acceptable. For anything that goes into a formal advice, a decision memo, or an approval request, the answer needs to be verifiable.
The threshold concept is important because it maps to a specific product requirement. A tool that produces cited outputs with article-level specificity sits below the threshold for most professional use cases. A tool that produces uncited prose summaries sits above it for most formal use cases, regardless of how accurate the prose turns out to be. The decision about whether to require citations is therefore a decision about which side of the threshold the tool lands on for professional in-house work.
One counsel described it as the difference between a capable junior associate and a capable consultant who has never practiced law. Both might give you the right answer. The associate can tell you where they found it and walk you back through the reasoning. The consultant can give you a confident summary. For internal orientation, either is useful. For anything that carries your name, you need the associate.
Speed matters, but not unconditionally
The time-saving argument for AI legal research tools is real. Routine research questions that previously took an hour of database time can in some cases be answered in minutes with a well-designed tool. The counsel we spoke with did not dispute this. The qualification they added was: the time saving only accrues if the output is usable without significant rework.
A tool that saves an hour of research time and creates forty minutes of verification overhead is a twenty-minute saving, not an hour. Whether that is valuable depends on the cost of the forty minutes of verification. If verification requires the counsel's own senior review, the cost is higher than if a junior associate can do it. The overhead is not constant across the team.
For in-house teams in the Gulf, where legal team sizes are typically smaller relative to the companies they serve than in comparable Western markets, this overhead calculation matters more than it might for a large law firm with a broad associate base. The General Counsel who is personally reviewing a significant portion of the team's research output will feel the verification overhead more acutely than a practice group head at a firm with fifty associates available.
Language and jurisdiction specificity
Regional counsel consistently raised language and jurisdiction coverage as requirements that most available tools do not meet. For a company operating across the GCC, research needs to cover not only UAE law but Saudi, Bahraini, and Kuwaiti frameworks, at minimum in the areas the company's operations touch. Arabic-language statutory coverage is important because official texts in Saudi Arabia are Arabic-first, and English translations are frequently not current with the most recent amendments.
The jurisdiction specificity requirement connects to the citation requirement. A generic AI tool that produces confident answers about "Gulf commercial law" without distinguishing UAE from Saudi from Bahrain is producing output that looks like research but is not. The specific article numbers that constitute a citation also force the tool to commit to a jurisdiction, because different jurisdictions have different article numbering. A tool that cites vaguely is likely summarising at a level that elides jurisdictional differences.
What the data layer looks like from the in-house perspective
Several counsel described their ideal workflow in similar terms. They would ask a question, receive an answer with citations to the specific provisions, verify the key citations with a quick source check, and have a document trail for the file. The AI tool is responsible for the retrieval and the initial synthesis. The counsel is responsible for the verification and the judgment. The tool is faster than a database search for retrieval; the counsel's verification check is faster than reading the full statute because the relevant passages are already identified.
This workflow depends on the citations being accurate and specific enough to verify quickly. A citation to an article number that is wrong, or a section heading that requires the counsel to read several paragraphs to find the relevant sentence, adds back the time the tool was supposed to save. The quality of the citation layer is not secondary to the quality of the answer; it determines whether the answer is usable in the professional workflow described above.
What counsel said they would not accept
We also asked about deal-breakers. The responses were short. Fabricated citations, where the tool invents a provision number, were described as immediately disqualifying. Not just for the tool, but for AI legal research more broadly, in the short term. "If I rely on a citation that does not exist in a formal document and that is discovered later, the credibility damage is mine, not the tool's." Hallucinated citations were the single most frequently mentioned concern.
Closely related was the concern about citations to outdated versions of legislation. UAE law in particular has been amended significantly over the past several years across commercial, employment, and data protection frameworks. A tool that is not current with the most recent amendments and does not indicate that its source is potentially outdated is described as less useful than no tool, because it creates a false confidence that a second check would need to unwind.
These are the design requirements we built Qanooni around: accurate, article-level citations to current legislation, with the ability to follow each citation to the primary text. The counsel conversations confirmed that this is not a feature preference. It is the threshold requirement for professional in-house use.
