Agentic AI: Things Are Getting Serious

August has been an eventful month in agentic AI news.

Harvey announced Harvey II. Then later this week it released details about Harvey Tenet, its first post-trained open-weight model focused on long-horizon legal work. Thomson Reuters is also talking about the next generation of CoCounsel Legal and its move toward agentic legal work.

The pace creates a difficult problem for law firms and legal departments of every size. There is no shortage of impressive technology. The harder questions are what to buy, when to buy it, what to connect it to, and how much organizational work needs to happen before a major platform investment makes sense.

Those questions will not have one answer because legal organizations are starting from very different levels of readiness.

A global law firm may have innovation lawyers, knowledge-management professionals, legal engineers, security teams, data specialists, and procurement resources. A midsize firm may have one or two people carrying most of that responsibility. A corporate legal department may depend heavily on IT, information security, procurement, and outside vendors. A smaller firm may have no dedicated innovation function at all.

The technology is advancing at extraordinary speed. Most organizations cannot change their processes, governance, data architecture, training, and culture at the same pace.

Harvey II

Harvey II addresses context. Legal matters develop over months or years. Lawyers learn the players, the history, the documents, the strategy, and the significance of facts that may make little sense outside the matter. Most AI interactions have required lawyers to recreate at least part of that context each time they begin a new task.

With Harvey II a Space can retain the context surrounding a matter or project, including documents, participants, work history, permissions, tasks, and other information. Harvey’s agents can work within that continuing context and use information about how lawyers and teams’ work. (harvey.ai)

An AI system working on a litigation matter could potentially understand the pleadings, key documents, deposition testimony, prior assignments, and recurring instructions before the lawyer starts the next task.

The same concept applies outside litigation. A transactional team could maintain deal context. An employment group could preserve investigation history. An in-house legal department could maintain context around a major contract, regulatory issue, or business unit.

Tenet

Tenet is a Kimi K3 base model that Harvey post-trained with Fireworks AI for long-horizon legal work. Harvey says the effort is intended to improve legal performance and control the cost of increasingly complex AI workloads. Harvey has also discussed a longer-term goal of helping legal organizations develop more specialized intelligence around their own work. (harvey.ai)

Legal AI is becoming capable of retaining context, planning multiple steps, using different sources, carrying work across a matter, and producing something closer to a complete work product for lawyer review.

Remember how not so long ago it was all about "prompt engineering"? It seems, however, that the era of prompt engineering as the central skill is already giving way to workflow design, supervision, validation, and governance.

CoCounsel

Harvey is getting a great deal of attention, but Thomson Reuters has also been rebuilding CoCounsel Legal around agentic work.

The next generation of CoCounsel allows a lawyer to describe an assignment in ordinary language. The system can determine which research, analysis, drafting, and other steps are required, sequence the work, and carry it forward. Thomson Reuters describes the architecture as increasingly agentic rather than a collection of individual AI skills that the lawyer must manually string together.

CoCounsel also has an important structural advantage: Thomson Reuters owns Westlaw, KeyCite, and Practical Law. It can work from Westlaw primary law, Practical Law guidance, organizational documents, and matter context, with citations available for lawyer review.

That makes the Harvey versus CoCounsel discussion much more interesting than some of the early comparisons suggested. CoCounsel is no longer simply an AI research assistant and Harvey is no longer simply an advanced legal chatbot.

Harvey grew as an AI-native platform and has built agents, workflows, model orchestration, integrations, institutional knowledge, and legal research connections around that platform.

Thomson Reuters began with one of the world’s largest legal research and know-how ecosystems and has been building increasingly sophisticated AI into that environment. CoCounsel may also present a more accessible entry point at this time.

Keep that in mind when evaluating either product.

Procurement

The pace of product development creates an unusual procurement problem. A legal organization may spend three months:

1. identifying requirements;
2. evaluating vendors;
3. conducting security review;
4. negotiating contracts;
5. running pilots;
6. obtaining stakeholder approval.

During those same three months, the products may change materially. That does not mean procurement has become futile. It means the evaluation criteria need to become more durable. Organizations should be careful about buying primarily because one product has a feature that another product lacks today. That gap may disappear quickly.

More durable questions include:

1. Does the platform fit the organization’s core workflows?
2. Can it work securely with the organization’s information?
3. How does it handle permissions?
4. What authoritative sources does it use?
5. How much human review does the output require?
6. How well does it integrate with the existing technology stack?
7. How difficult is adoption?
8. Who will own administration and workflow development?
9. What is the total cost over the contract term?
10. Can the organization exit or adapt if the market changes?

Data sources and governance remain central considerations

For organizations implementing or modernizing a document-management system, knowledge platform, or Microsoft 365 environment, that foundational work has become even more important.

AI becomes much more useful when it can operate against the organization’s own information.

NetDocuments is moving aggressively in this direction. Its ndConnect program supports Harvey and Thomson Reuters and provides connectivity for tools including Claude, ChatGPT, and Microsoft Copilot Studio. NetDocuments says those connections operate within its existing permission and governance structure. (netdocuments.com). The same is true for other household name platforms like iManage and others.

An organization does not necessarily need to identify the one AI platform it intends to use for the next five years but it should be building a governed information environment that supports several competing tools over time.

Your technology strategy needs to be anchored in good, clean, usable, governed data. The DMS, Microsoft 365 environment, security model, permissions, matter structures, metadata, and knowledge organization are no longer merely operational systems. They are an integral part of the AI infrastructure. If those systems are weak, every AI platform connected to them starts with the same disadvantage.

Permissions become visible

AI also changes the practical consequences of excessive permissions. A document that a lawyer technically can access, but would rarely discover manually, becomes easier to surface once natural-language search and agentic tools are connected to the repository.

Before AI, that problem might remain invisible for years. Once the repository becomes searchable through natural language, a user can ask a simple question and surface information that previously required considerable effort to find.

The AI exposes permission issues, therefore, organizations need to think ahead about access and permissions around:

1. Compensation;
2. HR records;
3. personnel information;
4. board materials;
5. financial information;
6. privileged investigations;
7. restricted matters;
8. confidential business information;
9. ethical walls and need-to-know restrictions.

This is not glamorous AI work, but we cannot say this enough: this may be some of the most consequential AI work organizations do.

Microsoft

For all these reasons, Microsoft deserves a prominent place in the discussion because so much legal work still happens in Word, Outlook, Teams, Excel, and PowerPoint.

Microsoft 365 Copilot will not replace Westlaw, CoCounsel, Harvey, or other legal-specific platforms but it could address a different set of work.

Lawyers and legal professionals spend significant time:

1. reviewing email;
2. drafting routine communications;
3. revising documents;
4. summarizing meetings;
5. organizing information;
6. preparing presentations;
7. analyzing spreadsheets;
8. searching internal information.

Those activities occur every day. Creating efficiencies around that create meaningful cumulative savings. It should also be pointed out that, right now, Copilot is still a relatively affordable tool.

Copilot Studio adds another layer by allowing organizations to build agents around repeatable processes. That gives legal organizations another path to experimentation before, alongside, or instead of a major legal-AI platform deployment.

Change Management

The key to all things AI related is of course adoption. Purchasing licenses does not create adoption.

People need training, examples, guidance, support, and opportunities to use AI in work they perform.

It bears repeating that the competitive landscape keeps broadening. Features that once distinguished one platform are spreading quickly, which means AI strategy should be broader than feature-by-feature comparisons.

Contracting

Pricing models are likely to evolve as agents become more capable. Traditional legal software has largely been sold by seat. Agentic AI makes that less predictable because one user can initiate large volumes of automated work involving significant compute, document processing, and research.

Legora announce a move to task-based billing. Microsoft already uses consumption credits for some agent activity. Harvey increasingly supports matter-level tracking of usage and cost with their new Spaces. While Harvey has not yet moved to token billing, the economics of AI are clearly moving beyond a simple seat-count model. In fact, this is one of the reasons for Tenet.

Organizations considering multi-year contracts should ask:

1. What exactly is included in the subscription?
2. Are agentic workflows included?
3. Are there usage limits?
4. Can the vendor introduce credits or consumption charges during the term?
5. Are large-document workloads treated differently?
6. Are new generally available features included?
7. Can seat counts change?
8. What happens if the vendor changes the product architecture?
9. Is there a renewal cap?
10. Can the organization exit if the product changes materially?

A three-year discount may be attractive. A three-year commitment in a market moving this quickly also has risk. Negotiations and contracting should address both.

Conclusion

The importance of Harvey’s recent announcements is not only about Harvey. They signal that legal AI is moving quickly toward persistent context, agentic workflows, specialized models, deeper integration with organizational knowledge, and, likely, more task-based pricing. Law firms and legal departments should respond with a strategy that is continuously evaluated rather than treated as static.

Resist the temptation to make technology strategy based entirely on the newest feature release.

The better approach is to build an environment that allows the organization to evaluate and adopt new capabilities as they mature.

That means getting the information foundation right, cleaning up permissions, implementing governance, training users, identifying valuable workflows, and measuring results.

Karta Legal advises law firms and legal departments on vendor-neutral AI strategy, governance, workflow design, Microsoft 365 readiness, technology evaluation, and responsible adoption. We begin with the organization’s work, risks, data, and people, not with a predetermined product.