Posted by Catherine Carroll, head of legal operations at Legal Futures Associate iManage
Carroll: Get the information architecture in order first
Suppose a general counsel wants the total liability exposure on every client contract signed in the past three years, or a law firm wants to review a client’s entire real estate portfolio for risk in the leases.
Not long ago, that meant a multi-week project. Today, first-pass answer can arrive in minutes. Connectors built on open standards like the model context protocol (MCP) now let AI agents query governed document stores directly, so a question that once triggered a lengthy, in-depth project gets answered instantly.
There is a potential pitfall to be aware of, however: AI does not tidy the repository it is pointed at – it simply works faster with whatever it finds. The result? Teams that already invested in classification, data hygiene and permissions will scale their AI efforts more readily, while those who have not will struggle.
Where the mess comes from
In-house, the mess often hides in sales inboxes, subsidiary drives and files inherited through acquisitions. Law firms face this challenge on two fronts: their own engagement letters and client guidelines, and the client material they manage.
Picture a client whose paperwork sits under its registered name, a brand name, a shorthand, the name it traded under before an acquisition, and an overseas affiliate. A lawyer sees at once they belong together. An agent reads five unrelated counterparties, and the amendments come adrift from the master agreement.
Version control causes similar trouble. A document is agreed, tweaked at the eleventh hour and then signed, with all three iterations saved as ‘final’. Nothing tells the agent which one binds, so it treats each as equally authoritative.
Scattered storage and permissions make matters worse. Key files might be spread amongst multiple drives and locations rather than in a single, centralised repository, and permission structures might be designed for people navigating folder hierarchies, not for AI agents that can surface any content they can access with a single query.
For firms housing privileged client data, that is a regulatory, contractual and reputational risk, not just a housekeeping issue.
Groundwork before gadgets
The fix is unglamorous: get the information architecture in order first, then add AI tools. The effort divides into four areas:
- A single source of truth. Choose one governed repository for signed agreements and make data standards non-negotiable at upload, for instance by rejecting any file without a properly formatted counterparty name.
- A classification scheme machines can follow. Record a document’s status (draft, signed, lapsed, replaced) in its own field, not the file name, and settle how contract types differ, such as where a consultancy agreement ends and a master services agreement begins.
- Cleaning up the back catalogue. New intake rules only help new documents. Legacy material stays messy until someone deliberately fixes it.
- Permissions built for agents. Limit each agent to what the person asking may see, enforce matter restrictions and ethical walls inside the repository, and log what every agent retrieved and on whose behalf.
A note on the third item: faced with a huge archive to clean up, many teams simply put it off. Instead, rank the work by exposure: high-value contracts first, then live relationships, auto-renewing agreements, and anything under a riskier governing law.
An automated enrichment pass can beat hunting through disparate systems, provided a person reviews the results before anyone relies on them.
The payoff comes rapidly. With sound foundations, AI can compare negotiated terms across thousands of agreements and flag patterns unprompted. A US company’s sales agreement might state Delaware as the standard governing law, while the legal team commonly agrees to the laws of England and Wales for UK-based customers.
Without the need to track this deviation manually, the team can stay on top of trends and go straight to the commercial judgement: is the clause worth fighting over? If not, change the template.
Obligation tracking also becomes easier, covering anything requiring action, from renewal notice periods to customer-specific security commitments that could limit future IT changes. For law firms, this even becomes a capability to sell to clients, provided clients agree how their data may be used.
So, how best to reach this promised land?
A practical 90-day plan
Try a self-assessment. Could you say today, without commissioning a review, how many live agreements you hold with your 10 largest counterparties, which copy of a given contract is the signed one, and which people and AI tools can access it?
If not, choose a single question the business or a client needs answered and build a three-month pilot around it.
Month one: name an owner, start security and data protection review, select a single risk tier, locate every relevant agreement and flag duplicates and obsolete drafts, having first checked retention rules and legal holds.
Month two: introduce one naming convention, reconcile counterparty names, require a status field on upload, and audit access rights.
Month three: connect an agent, put your question to it and check its answers against the source documents, having agreed in advance who checks and what error rate is acceptable. An accurate answer in minutes, to a question that used to demand a project, is the business case for the next tier.
An old adage tells us that the best way to eat an elephant is one bite at a time. Steady, deliberate progress turns contracts into assets that AI can use effectively– and over time, legal organisations will find themselves able to quickly answer key questions, surface valuable insights, and drive better business outcomes, thanks to contracts that are ready to be questioned.
Leave a Comment