Most AI used at law firms today reads and answers. It summarizes a long document, drafts a clause, finds the precedent somebody half-remembers. When it gets something wrong, an attorney catches it before the work leaves the building, and the record of what happened sits in that attorney’s own review.
A quieter change is now underway, and it shows up in product announcements rather than in security bulletins. On November 5, DealCloser announced that its connector for Harvey is generally available, and the announcement is unusually specific about what the connection does. It “supports both read and write capabilities, enabling lawyers to access DealCloser transaction data and take approved actions directly from Harvey,” and it was “Built using the Model Context Protocol (MCP).” In practice, that means an assistant in one application can create a closing checklist in another, update an existing checklist, and create a task with an assigned owner. The company’s own example is Harvey noticing a missing approval, then adding a task and assigning responsibility.
That is a different kind of software than a chatbot, and it raises a different kind of question for whoever runs the firm’s systems. The question is no longer only whether the output was accurate. It is what the tool was allowed to touch, who agreed to the change, and whether anyone can reconstruct it afterward. Firms should not have to work that out under pressure after something has already been written into a client matter, and at Innovative Computing Systems we treat that as our job rather than yours: an environment that is safe, secure, and reliable has to account for software that acts, not just software that answers.
New York’s bar association has now written this down
Five days before that product announcement, on October 31, the New York State Bar Association’s House of Delegates approved its Legal Ethics Guidelines for Lawyers’ Use of Artificial Intelligence, prepared by the association’s Committee on Artificial Intelligence and Emerging Technologies. The full report runs to 20 guidelines. One of them, Guideline 9E, addresses AI systems that can take action, and it reads much more like a technical control specification than an ethics statement.
It asks a lawyer deploying such a system to define in advance the tasks, matters and data the system may access. It asks for human approval before consequential actions, and it names three of them directly: “changing access permissions,” “modifying or deleting client records,” and “disclosing confidential information.” It asks for “maintaining a record sufficient to determine what the AI System did, when, and on whose instruction.” It asks for monitoring, with use suspended when a system “acts outside its authority,” and for periodic testing of safeguards “including measures designed to prevent unauthorized instructions from causing external actions or disclosure.”
Two things are worth clarifying. NYSBA is a voluntary bar association, so these are guidelines interpreting the New York Rules of Professional Conduct rather than disciplinary rules, and firms outside New York are not bound by them. And the DealCloser announcement says nothing about permissions, access scoping or audit logging, which is not the same as saying the product lacks them; a press release is not a security datasheet.
What the two documents have in common is more useful than either on its own. A capability that writes into firm systems is now shipping in the tools firms already license, and a state bar association has set out, in detail, what someone is expected to be able to show about it later. Guideline 2A allows that assessment to be made at the firm level, and says a lawyer “may reasonably rely on it.” Section 3.3.7 puts the limit plainly: “Automated capability does not shift accountability away from the lawyer.” Somebody has to do the configuring and the record-keeping. In most firms, that somebody reports to the administrator, which is precisely the kind of work we take on so it does not land on an already full plate. You deserve better than discovering a connector’s reach after the fact.
What we do about it
Access scoping comes first, because it is the control that makes everything after it manageable. Innovative Computing Systems inventories the AI tools and connectors operating in a firm’s environment, including those enabled within software the firm already owns, and we scope what each one can reach so an assistant’s access matches the matters and data it actually needs. That means a tool connected for transaction work cannot quietly read across unrelated matters, which is the same discipline we apply to mailbox and system access generally.
Approval steps come next. We configure environments so a system does not change permissions, alter client records or send client information outside the firm without a person agreeing to it, which is the control Guideline 9E describes and the one most easily skipped when a connector is enabled in a hurry. Alongside it, we retain and review audit logs across identity, document and practice systems, so there is a durable record of what acted, when, and under whose account. That record is what turns a difficult question into an answerable one, and it builds on the same thinking behind treating AI interactions as firm records and tracking what AI actually changed on a matter.
Because these connectors authenticate as a user, identity controls carry more weight than they used to. An assistant acting with someone’s credentials inherits that person’s reach, which is why phishing-resistant authentication on privileged accounts and a clear picture of which vendors hold firm data belong in the same conversation, separate projects. Managed Intelligence is how we bring this together: we handle the governance work, the configuration and the review cycle so your firm can adopt these tools deliberately instead of discovering them in a log.
None of this asks your firm to become expert in agent architecture. It asks somebody to own the configuration, the approvals and the records, and to keep owning them as the tools change — work we have been doing for law firms since 1989 so your team can stay focused on clients and matters. When a capability like this arrives in a tool you already license, we would rather have the controls in place first than treat a preventable problem as an incident later. We’ve got your back.
If you would like to talk through how AI tools and connectors are currently configured in your environment, and what it would take to scope them properly, schedule a consultation. A Solutions Consultant will follow up to understand your firm’s setup and walk you through the options
