Translating the Tools & Trends That Shape the Legal Profession
    Rising Above the Noise

    One Agent, One Dataset

    Laura Kennedy & Chelsey Lambert 5 min read
    One Agent, One Dataset

    Monday, July 20, 2026

    Arnold & Porter named its first Chief AI Officer on Monday. Morgan & Morgan named its first on Tuesday. By Wednesday, Litera had rebuilt its entire brand around a single AI agent and Amazon had surfaced a legal product folded into a general-purpose assistant. Harvey bought its third company of the year on Thursday. Then on Friday, the chief executive of a data company published the sentence that gathered the whole week into one line: law firms don't have an AI problem, they have a data problem.

    Read in sequence, the week stops looking like five unrelated announcements and starts looking like an industry arriving at the same conclusion from five directions. The argument about which model or which agent wins is quieting down. In its place is a quieter agreement about where the value actually sits, and it sits below the agent, in the firm's own accumulated knowledge. In the language of Adaptive Legal Practice Systems™ — ALPS™, the framework Laura developed at Circle Management Group — AI has always been the connective tissue running through a firm's ten technology layers. This week the vendors agreed, out loud and in their product decisions, that the tissue is only ever as good as what it connects to. The agent is becoming a commodity. The dataset underneath it is not.

    July 15, 2026: Litera Bets the Brand on One Agent

    Litera relaunched on Wednesday around a single AI agent it calls Lito, and the relaunch is less a new product than a declaration of where the company thinks the market is going. Lito is meant to be the one interface across everything Litera does — drafting, document comparison, contract review, institutional knowledge, client development — collapsing the practice of law and the business of law into a single point of contact, embedded natively inside Microsoft 365 and Google Workspace.

    Chief executive Avaneesh Marwaha framed it as the end of an old division. "The practice of law and the business of law have run on separate systems for as long as firms have existed," he said. "That ends here: one agent, one dataset." The second half of that sentence is the part to hold onto. Litera isn't selling the agent as the differentiator. It's selling the unified dataset the agent runs on, and treating the agent as the way in.

    Chief product officer Adam Ryan drew the line even more sharply. "AI is table stakes. Trust is the moat," he said. "You can't retrofit years of deterministic accuracy." Litera's pitch leans on the three-decade-old, rules-based redline algorithm at the core of its comparison tools, the kind of deterministic reliability a probabilistic model can't manufacture on demand. That's a company telling the market that the model layer is not where it competes. It competes on the accumulated, verified data and logic it has built up over thirty years, and Lito is the door it puts on the front of that.

    July 15, 2026: Amazon Opens a Front Door

    The same Wednesday, Amazon Quick for Legal drew its first real attention as Amazon's entry into the legal vertical, an AI assistant aimed at legal research, contracts, and compliance. Coverage across Law.com and Artificial Lawyer landed on the same read, and it's an instructive one. Quick for Legal sits on top of Amazon Bedrock, routing queries to a range of third-party language models. It does not carry its own legal dataset. Artificial Lawyer described it, not unkindly, as mainly a fancy interface over models a firm could already reach.

    Set next to Litera's relaunch on the same day, the contrast writes itself. Both are front doors. One opens onto thirty years of structured legal data and deterministic tooling. The other opens onto a general-purpose model stack with no legal context of its own underneath it. Amazon now joins OpenAI, Anthropic's Claude for Legal, Microsoft, Palantir, and Perplexity in the crowded entryway of legal AI, and it arrives with the largest cloud business in the world behind it. What it does not arrive with is a proprietary body of legal knowledge, and this week made clear that the entryway is not where the contest is anymore.

    That's the useful thing about Amazon's move. It shows the floor. A company with Amazon's resources can stand up a competent legal front door quickly, which means the front door itself is no longer scarce. Whatever a firm is buying when it buys a legal AI product, it is not buying the interface, because the interface is becoming something any large platform can produce. It is buying, or failing to buy, what sits behind the interface.

    July 16, 2026: Harvey Buys the Knowledge Layer

    Harvey made the third version of the same argument on Thursday, and it made it with an acquisition. The company acquired Benchmark, a Y Combinator-backed decision infrastructure platform for asset management, its third acquisition of 2026. Benchmark's whole function is to capture the institutional knowledge investors build up across deals and apply it to the next one, and it's trusted by firms managing more than two trillion dollars in assets. Benchmark's co-founders, Alec Dunn and Connor Janson, are joining Harvey's product and engineering teams.

    Harvey chief executive Winston Weinberg said the deal came down to a shared conviction, and his phrasing could have served as this issue's thesis. "Benchmark is built on a conviction we share: a firm's edge is its institutional knowledge, and AI should help firms capture that knowledge and apply it." Harvey didn't buy a better model. It bought a system for turning accumulated expertise into a usable, queryable asset, and it bought it on the back of a quarter it says produced more than a hundred million dollars in net-new annual recurring revenue.

    The strategic logic tracks with where Harvey already sits. The company says it serves around fifty asset management firms on work like investment due diligence and deal document review, and Benchmark extends that coverage across the full deal lifecycle, from first screening through the investment committee decision. Benchmark's founders described the fit in the same terms Weinberg did, calling Harvey one of the strongest application-layer companies in AI and noting the two already shared customers. What's telling is the direction of the acquisition. A model company would have bought capability. An application company chasing scale might have bought market share. Harvey bought the layer that captures how experienced people actually decide, which is the layer that doesn't come out of a foundation model no matter how large it gets.

    The pattern across the three moves is hard to miss once you're looking for it. Litera is unifying its dataset and putting one agent on top. Amazon is offering an agent with no dataset of its own. Harvey is spending real money to acquire the machinery that captures a client's institutional knowledge. Three companies, three strategies, one shared premise: the agent is the easy part, and the knowledge underneath it is the part worth owning.

    In Other News This Week

    The moves around the edges of the week reinforced the same trajectory rather than cutting against it.

    Two firms named their first Chief AI Officers within a day of each other, and the pairing rewards a look side by side. Arnold & Porter appointed Roger Maeda, previously its director of enterprise applications, to lead a dedicated AI team translating emerging technology into deployable tools for client work. Morgan & Morgan appointed Kaushik Chakraborti, who spent more than a decade in senior engineering leadership at Amazon and helped launch Alexa+, to run its AI strategy across operations and client experience. One firm promoted from inside its own data and applications function. The other hired a big-tech builder from the outside. Both are staffing the role that owns the layer this week was all about, and the fact that a plaintiffs' powerhouse and a global defense firm made the same structural decision in the same week says the CAIO seat is no longer a novelty.

    Legatics launched a Model Context Protocol server connecting its transaction management platform to external AI tools, another instance of a vendor choosing to expose its data to whatever agent a firm prefers rather than trapping it behind a proprietary assistant. It's a small move with the same logic as the big ones. The data layer is the asset, and making it reachable is the point.

    Cooley's fund formation platform moved to flat-fee pricing on certain transfers, a modest billing-model data point that continues the pressure on the hour this newsletter tracked last week with Norm. And Artificial Lawyer's Friday roundup raised a point that ought to sit underneath every one of these announcements: whether all the firm AI investment of the last two years has actually reduced what clients pay. So far most firms can't show it, and that gap between spend and demonstrated savings is itself a data problem.

    The CTO Gap Zone™ of the Week: The Dataset Underneath the Agent

    The last few Gap Zones have circled the same asset from different angles. The model and the content are rented; the firm's own context is the layer it owns. Owning that layer is not the same as having encoded it. And once it's encoded, someone has to be able to read the meter on what it costs and returns. This week sharpens the point one more turn. Every major vendor move confirmed that the firm's data is the prize, and most firms still don't have that data in a form any agent can actually use.

    Kevin Walker, chief executive of Centari, put it plainly in his Friday piece. "The most valuable thing a firm owns is its accumulated and ever-evolving knowledge of the market and of the strategies to navigate the complexity of modern business." His sharper observation is that expertise accumulates across deals, not documents. A single purchase agreement can't tell you why one party conceded the indemnity escrow or how a defined term shifted across a dozen amendments. That pattern lives in the relationships between documents, across a portfolio, over years, and generic systems that read text sequentially can't reliably see it. The knowledge is real. The structured, queryable version of it usually isn't.

    That's the gap. A firm can buy Litera's Lito, or Harvey, or stand up something on Amazon's stack, and still get thin results, because the agent inherits whatever state the firm's data is in. Point a capable agent at an organized, annotated, connected knowledge base and it produces the firm's work. Point the same agent at a shared drive full of unstructured documents nobody has ever mapped, and it produces a plausible average of the public internet with the firm's letterhead on top. The interface is identical in both cases. The output is not, and the difference is entirely the dataset.

    This is why the CAIO hires and the Centari argument belong in the same section. The work of the next two years is not choosing an agent, because the agents are converging and will keep converging until they're roughly interchangeable at the interface. The work is building the data layer the agent runs on, which is unglamorous, slow, and closer to knowledge management and information architecture than to anything that demos well. It is also the only part of the stack a competitor can't buy off the shelf, because Amazon can replicate the front door in a quarter and no one can replicate thirty years of a firm's judgment.

    What This Means For Your Firm

    Three implications for any firm running an AI program through the back half of 2026.

    First, stop evaluating vendors on the agent and start evaluating them on what they do with your data. The interface is commoditizing in real time, and this week showed it. The questions that matter at the point of purchase are where your data lives, whether the platform structures and enriches it or just reads it, who owns the encoded output, and what happens to that asset if you leave. A firm that buys the best-demoing agent and gives no thought to the dataset behind it has bought the one part of the stack that's about to be free.

    Second, treat the data layer as the strategic project it is, not as cleanup work to be done later. In ALPS terms, this is the connective tissue getting something worth connecting. Most firms own the raw material — the matters, the redlines, the deal histories, the institutional judgment — and have never turned it into anything an agent can query. The Circle Method™, the seven-dimension methodology inside Circle 360®, exists to do exactly that work of building and maintaining the firm's encoded knowledge so the tools produce the firm's practice rather than the market's average. Architecture defines what the ecosystem is; the methodology is how the dataset underneath it actually gets built. More on the full framework at cmgconsultants.com.

    Third, read the CAIO wave as an org-chart signal about where the work lives. Arnold & Porter and Morgan & Morgan didn't hire agent operators. They hired people to own the firm's data and knowledge strategy, one from inside the applications function and one from big tech. The role is being defined around the layer that survives every interface change, and the firms building it now are the ones that will have something to point an agent at when the agent stops being a differentiator, which this week suggested is already happening.

    The interface is collapsing to a single agent, and that agent is quietly becoming something anyone can build. What stays scarce is the accumulated knowledge underneath it, structured well enough to be useful. As Laura's book, Escaping the CTO Gap Zone, argues throughout, the firms that treat that knowledge as a strategic asset and invest in it with discipline are the ones that pull ahead. The rest will keep shopping for agents and wondering why the answers come back generic.

    Laura Kennedy and Chelsey Lambert are Co-CEOs and Co-Founders of Legal Tech Decoded. Rising Above the Noise lands every Monday at legaltechdecoded.com.

    The Daily Brief

    Every morning, the most important legal tech developments — curated and context-rich — delivered before 8 AM.

    Free. No spam. Unsubscribe any time. Read by 28,000+ legal professionals.