Every summer, thousands of people who build, buy, and run the technology behind legal work spend a week comparing notes at ILTACON, the annual conference put on by the International Legal Technology Association (ILTA). If you’ve never been, here’s the shorthand: it isn’t a trade show built around any single vendor’s roadmap. It’s a practitioner-run gathering, with more than 80 education sessions this year chosen from over 400 ideas submitted by the ILTA community itself.

I had the privilege of volunteering with ILTA to help coordinate several of those sessions, from interviewing speakers to guiding them as they created the talk tracks for their sessions.

The thread that connected all sessions this year was this: the legal industry has stopped debating whether to use AI and started grappling with how to be accountable for it – for its cost, its governance, and the people who actually have to use it.

Here’s what stuck with me most.

The Reckoning: From Experimentation to Accountability

For the past few years, most organizations’ AI strategy amounted to enthusiasm plus a credit card: pilot everything, worry about the bill later. NetDocuments CEO Josh Baxter shared a common term for the phase as “tokenmaxxing” – pushing people to use AI, use more of it, experiment with everything, the sky’s the limit. With the advent of consumption-based pricing instead of an all-you-can-eat token buffet, that phase is ending. Now, plenty of organizations are rationing instead – usage caps, budgets, decisions about who gets access to which tools.

The numbers explain why. Revenue at large firms is up more than 12% in the first six months of the year, with demand for legal services (measured in hours worked) up 4.8%.  Yet, clients are simultaneously pushing hard for AI to lower what they pay. Thursday’s closing keynote, a recap of conversations from ILTA’s G100 and G200 leadership gatherings, moderated by NetDocuments Chief Product Officer Dan Hauck, put real numbers behind that nervousness.

Reported AI spending disclosed by firms spans an enormous range, from roughly $100,000 to as much as $35 million, and a rough industry rule of thumb is emerging that total AI spend should land somewhere near 1% of revenue. Layer in consumption-based pricing – where the bill depends on how heavily a tool gets used – and a lot of finance and innovation leaders are quietly nervous about costs that are hard to predict in advance.

A recent Gartner survey of more than 780 IT leaders found that only 28% of AI use cases fully succeed and meet ROI expectations. One CIO on the panel offered the most honest framing of the week for measuring return in a fast-moving market: “the tools we have today aren’t what will be the tools we have tomorrow … so do I want to spend 500 hours to calculate the ROI today? It’s a meaningless notion. As long as it’s positive, move forward.” Another panelist reframed the whole exercise as “fail fast, return on effort” rather than return on investment – a distinction that matters when the tools themselves have a shelf life measured in months.

That same panel surfaced something worth sitting with regardless of your industry. James McKenna, CIO of Foley & Lardner, described watching clients evolve, in his words, from “you can’t use AI,” to “which AI are you using,” to now, simply, “tell us how you’re using AI.” The demand for AI itself has quietly become a demand for transparency and pricing simplicity. “You need to understand how much it costs to get an answer,” as he put it. “Don’t overcomplicate billing. Keep it simple.”

It’s showing up in how firms think about their own business model, too. A session called “The Billable Hour Paradox: Making Non-Billable Work Visible (And Valuable)” tackled the same tension from a different angle, and the Thursday panel spent real time on what one panelist called “the rise of the legal engineer” – new roles sitting between lawyer, technologist, and knowledge manager, because delivering legal work profitably in a consumption-priced world takes a different team shape than delivering it by the billable hour.

Gartner has started naming this shift directly, describing the discipline organizations now need as “context engineering” – using it to “improve AI model output accuracy, enhance relevance and reduce token usage, thereby lowering operational costs.” Whatever you call it, the message is consistent: the conversation has moved from “are we using AI” to “can we prove what it’s worth.”

Context Is the Asset That Compounds

The most useful mental model I heard all week was also the simplest, and NetDocuments used it to open our own Company Update: three things determine whether an AI agent is actually good at its job:

• The model underneath it,
• The harness (the application or assistant) built around it, and
• The context it can draw on while it works.

Models are rented; they improve and get replaced on a lab’s schedule, not yours.

Harnesses get swapped constantly, and most organizations already run more than one.

Context – the organized, governed record of an organization’s own documents, precedent, and institutional knowledge – is the one layer an organization actually owns, and the only one of the three that compounds instead of resetting every time a tool changes.

That distinction is testable, not just a slogan. NetDocuments published the Legal Context Engineering Benchmark to measure what better context is worth – in both answer quality and dollars – by asking the same 300 questions to a fixed AI model and harness/application, changing only the context the agent could reach. In one test case using a real-world merger, with poor context, the answer scored 0.10/1.0; with richer context, 0.67 — nearly 7x better.

Across 300 test questions spanning 10 real legal matters, the pattern held: better context made correct answers roughly 48% cheaper to produce, without sacrificing accuracy, and the benefit grew stronger on more capable models rather than weaker – the basis for a projected $942,000 in annual savings for a 2,000-user firm.

Gartner’s research points the same direction: firms report roughly a 30% accuracy gain in agentic AI when strong context is in place. NetDocuments published its full methodology specifically so any organization can run the same test on its own documents rather than take the results on faith. The bottom line: context dramatically reduces token use and AI costs.

The week also produced a genuinely useful, vendor-neutral tool for cutting through the noise: five questions to ask the next time anyone – including a vendor you already use – claims their product has “context.”

• What does the system actually understand or extract from your documents?
• How does it find the relevant information it needs?
• Whose permissions determine what the AI can retrieve?
• Can that governed context reach the AI tools your organization actually chooses?
• And what evidence shows the effect on answer quality and cost?

Architecture matters, but so does evidence.

It’s worth being precise about what “context” bundles together, because most vendors, NetDocuments included, are describing several distinct technical layers under one word: automatic classification and metadata extraction that requires no setup, custom profiling an organization configures for its own taxonomy, semantic search that finds meaning rather than just keywords, and a connective graph tying documents, people, and matters together – still in private preview as of this year’s conference, not a shipped feature every customer has today. Treating those as one thing instead of four is exactly how “context” claims get overblown.

Another key topic of discussion around context was being able to connect to an organization’s AI tools of choice via Model Context Protocol (MCP) versus having to move files outside of the system of record, usually the firm’s document management system. The proof, for what it’s worth, is starting to come from customers rather than vendors.

Thomas Kline, Director of Information Technology at Stark & Stark, described connecting the Perplexity AI-powered search and answer engine to his firm’s document repository through MCP (more on that below): “No uploading documents. No selecting files. Our lawyers simply point AI at a matter, and it works from everything they already have permission to access.” And Jeff Westcott of Akin, on a customer panel, offered the plainest advice of the week for anyone drowning in point solutions: “the more you can do in-platform, the better.”

Autonomy Without Governance Isn’t Progress

As AI tools move from answering questions to taking actions – drafting, filing, cross-referencing, flagging – the phrase “agents are users” came up repeatedly this week. NetDocuments Chief Technology Officer John Motz noted during the Company Update session: “People are starting to realize that agents are users, and they’re really smart users, so we have to think about their access and controls just like a normal user, if not more critically.”

Amazon Web Services’ Pallavi Nargund followed with the security use case in more concrete terms: customers have access to a layered, isolation-based model with quantum-resistant encryption already applied to stored documents, so that even if one layer of security is breached, the data, the memory, and the underlying hardware stay independently sealed off, protecting client confidentiality, data privacy, and the organization’s intellectual property.

That same tension – more autonomy requiring more governance – showed up as a load-bearing theme across sessions with no connection to each other: “Beyond the Charter: AI Governance Frameworks That Actually Work,” “Secret Agentic Man: When AI Starts Acting on Its Own,” and “Matter Mobility: Managing Risk, Access, and Accountability” each worked a version of the same question. A Q&A during the Company Update panel pressed on exactly this point: can you trace what an agent actually did, question by question, and can you revoke or roll back its actions if something goes wrong?

Those aren’t hypothetical questions once an AI tool has standing access to a client’s files rather than reviewing output an attorney signs off on afterward.

Gartner’s research backs the stakes here: without a semantic foundation underneath it, the firm projects 60% of agentic AI projects relying solely on point-to-point connections will fail by 2028. The connection itself isn’t the hard part. What has to travel with it – permissions, ethical walls, audit trails – is.

This is already showing up in how AI vendors wire products together. During the week, Google and NetDocuments announced a new connection between Gemini Enterprise for Legal and NetDocuments through Model Context Protocol (MCP), joining existing connections to Claude, Microsoft Copilot, and Perplexity and other legal tools like Harvey and Legora.

The design principle behind it is the one worth borrowing regardless of which vendors you use: governance has to travel with the connection automatically, not as a setting someone has to remember to turn on.

Adoption Is a People Problem, Not a Software Problem

Some of the most practical conversations of the week weren’t about frontier models at all – they were about people. NetDocuments masterclass, “Adapt, Adopt, Transform,” opened with a line from futurist Alvin Toffler, written more than 50 years ago but newly urgent: “The illiterate of the 21st century will not be those who cannot read and write, but those who cannot learn, unlearn, and relearn.”

Legal innovation partner Kathleen Hogan built on that with Dr. Larry Richards’ research on the “Lawyer Brain”: lawyers are trained to be skeptics (finding what could go wrong is the job), to value precedent over novelty, to prize autonomy over being told what to do, and to write off an imperfect tool entirely rather than debug it. These professional strengths, she was careful to note, are not character flaws – but ones that make standard change management backfire in a law firm.

Hogan also included business leadership consultant Andy Stanley’s Trust Gap – the distance between what people expect from leadership and what they actually experience. The session tied the Trust Gap to usage of AI tools as well and pointed out that it breaks asymmetrically: one bad AI experience outweighs 10 good ones, because “it didn’t work that one time” becomes the whole story people tell.

She illustrated it with a real example: an early AI pilot at a firm quietly transposed a single digit in a client’s phone number, an error nobody caught for a long stretch, precisely because nobody was watching for it. The lesson wasn’t “don’t use AI.” It was that verification has to be designed into a workflow deliberately, because reviewing AI output is itself a demanding task that degrades exactly when people are busiest.

Kyle Kissell, director of NetDocuments’ Legal Innovation Partner team, made the point that the gap between understanding a tool and just pointing it at a problem is where most disappointment lives. He shared an analogy of asking his son to mow the lawn and getting exactly that, minus emptying the grass catcher and bagging the clippings, because he didn’t give his son those specific instructions. Vague instructions get literal, technically-correct, sometimes useless results – from a person or an AI agent.

This sentiment was mirrored in one of my own sessions entitled “Human-Centered AI Adoption: A Practical Training Playbook for Legal Organizations and Corporate Law Departments.” Speakers included a technology trainer, a Director of Knowledge Management, a lawyer on his firm’s innovation team, and co-founder of Hotshot, a company that specializes in AI adoption. The consensus was that there is no substitution for hands-on, real-world work within a platform to drive adoption – short, guided, experiential learning that is persona-based and practice-specific.

This isn’t specific to law firms. It’s a reminder that in any industry, the bottleneck on AI adoption is rarely the technology. It’s whether the humans around it actually understand how to use it effectively in their specific work tasks and trust what it’s doing.

Know What You Have Before You Buy What’s New

The most grounded advice of the week was also the least flashy: before buying anything new, ask what you’re actually trying to accomplish, and let that workflow – not a vendor’s roadmap – drive the investment.

With more than 1,200 AI tools now competing for a legal buyer’s attention, a genuinely useful diagnostic question came up more than once: can your existing technology stack already cover 80% of the workflow in question?

And rather than guessing where the gaps are, look at your own prompt logs to see what people are actually asking for. Real usage data beats intuition every time.

That same discipline applies one layer down, to the data underneath any AI initiative.

In a session on building what NetDocuments calls an MCP-enabled data estate, NetDocuments’ Nate Ruiz opened with a simple argument echoed by his co-panelists: if you don’t know what data you have, you don’t know what you can do with it – and you won’t know what’s being exposed to a generative AI tool, either.

Firms are increasingly blending structured data (the kind that lives in a database with metadata and a taxonomy for context) with unstructured data (the kind with details buried in documents and email), and the shared advice was less about picking a single “right” architecture and more about building toward a clean, well-understood data foundation before layering more AI on top of it.

Whatever comes next in this space, and something new arrives every week, a firm that already knows its own data is in a better position to use it.

What This Looked Like on the Ground

For those curious about the specifics: NetDocuments used the week to share a run of updates that fit squarely into these themes. Alongside the benchmark report referenced above, the company launched new AI capabilities for tabular review and finding legal authorities inside documents, announced the Gemini Enterprise connection, and was recognized on the 2026 Inc. 5000 list for the fifth consecutive year.

Our reimagined user experience and Legal Context Graph, first introduced in May, are rolling out deliberately rather than all at once: a redesigned core experience available to every customer now, with the AI-powered Legal Context Graph following in private preview and no forced migration timeline – a pace one presenter summarized as “not one big rollout, but a cascading set of changes.”

The experience of being an ILTACON volunteer is unique. Having the opportunity to work with other business partners and ILTA members on building this conference is truly rewarding, from the conversations with my team to working with speakers. It is the people who make ILTA special, and this year’s conference may have been the best one yet. Thanks to everyone who submitted a session idea, spoke on a panel, or simply showed up ready to compare notes. Conferences like this only work because practitioners are willing to be this generous with what they’ve learned. If this year’s conversations are any indication, the next chapter of legal AI won’t be won by whoever moves fastest. It’ll be won by whoever builds the most trustworthy foundation.

Keep the Conversation Going

If this piece has you thinking about what your own organization’s context actually looks like right now – and whether it would hold up to a test like the one described above – NetDocuments is hosting a webinar on September 17: Legal Context: The Key to Trusted Content and Better AI. It’s a practical next step, not a sales pitch: time with the people behind this benchmark, walking through the methodology and what it takes to apply the same thinking to your own documents. Register here to save your spot.