Who runs your AI?
Firms that are currently trialing specialist legal-AI platforms alongside frontier AI tools need to decide which tools are worth keeping.
On the surface, the platforms look remarkably similar. Each offers agents, workflows, research, document review, Word integration and access to the firm’s knowledge. At the same time, the underlying models are improving. Frontier AI platforms today are capable of handling most of the legal firm's operations, at least on paper.
Legal AI is moving from pilot budgets into ordinary operating costs. Unsurprisingly, partners want to know which of these tools is worth paying for? What the specialist AI premium buys? Which capabilities remain distinctive and how much of the system they can operate without the vendor standing beside them?
What Gartner is predicting
Gartner predicts that, by 2028, 70% of enterprises will abandon agentic AI built through vendor forward-deployed engineering, or FDE. In this model, engineers from the vendor work directly with the customer to build and deploy the solution.
Gartner’s concern is operational dependence. The customer has a system, but cannot maintain or evolve it without costly external help.
Legal technology makes this easy to miss. What starts as a software purchase can become a continuing implementation project, with the vendor holding knowledge the firm never acquires.
When implementation never ends
Legal work gives standard software plenty of reasons to need configuration: different document collections, access rights, taxonomies, precedents and approval processes. A tool that fits neatly into one firm may require significant work at another.
Vendor involvement is not the problem. It becomes a problem when each new requirement becomes another piece of customer-specific work.
Gartner predicts that fewer than 20% of FDE engagements will turn recurring customer requirements into capabilities in the vendor’s core product by 2028. The repeated request from one customer may remain a bespoke feature, even when several customers need the same thing.
That is the uncomfortable question behind a premium AI licence. How much of the fee buys a repeatable product? How much pays for continuing access to the people who know how to make it work?
The answer will not be obvious in a demonstration. It becomes clear when a workflow changes, an integration breaks, or the firm wants to extend the system to another practice.
What the firm can support
For a smaller firm with a defined problem, a specialist legal tool can still be the sensible choice.
Consider document review. A specialist product may already contain the workflow, document handling, permissions and interface required for the task. The licence may cost more than access to a general model, but the firm does not have to assemble those elements itself.
That matters because smaller firms often lack the time, technical knowledge and integration capacity needed to turn a flexible platform into a dependable operational process. SRA research identifies barriers including cost, fear of rapid obsolescence, limited confidence and the difficulty of choosing, setting up and integrating technology.
The cheaper licence can become the more expensive system once somebody has to design the workflow, connect the data, test the output, train users and support it.
Larger firms are more likely to have knowledge teams, security specialists, engineers and enough users to spread development costs. A frontier platform can support several use cases and give the firm more control over data, integrations and model choice.
That does not mean a larger firm can run AI well. Fragmented knowledge, unclear ownership and weak evaluation can create the same dependency at a much greater scale.
What the premium should buy
Specialist legal AI can justify a higher price. The premium may buy authoritative content, tested workflows, reliable integrations, access controls, evaluation against legal tasks and lower internal support costs.
Those are product capabilities. Access to a capable language model is only one component.
The premium is harder to defend when the offering still needs extensive custom engineering, and when work created for one customer does not become part of the product available to others.
A frontier platform brings more flexibility and more work. Someone still has to design workflows, connect knowledge, test results and support users.
Choosing software divides responsibility for running it. Some comes with the product. The rest stays with the firm.
Ask who will make the changes when the work changes. If the answer is still the vendor’s engineering team, the firm is buying a continuing project, not a product.