How are you reading this?
For partners, GCs, ops, and the systems people beside you—pieces that still hold up when someone asks "why is our AI bill structured this way?"
For pilot-scoping detail (same list we use in scoping), switch to Systems Architect above, or open the scoping reference directly.
Maintained by Firm Leverage
Long-form work from builders—useful ammunition for partner meetings, vendor reviews, and ethics conversations.
Governance reading list: we are expanding what we cite here; agent-assisted updates are on the roadmap and will stay labeled when they land. If you maintain standards or bar guidance and want to tighten how we represent something, book a call—we mean it.
A structured read of AI/legal-tech terms—training, retention, subprocessors—for counsel and procurement. Not a verdict; an evidence map you can reconcile with diligence.
Open → Firm LeverageWe host AI for Lawyers—and we attend adjacent legal-AI programming (Boulder cluster, RM Legal AI, sponsors’ lanes) without confusing their format with ours.
Open → Firm LeveragePublic lanes — backend scale, sales and engagement leadership, disciplined city sequencing. Directional naming unlocks verbally with phrases we share outside the crawler story.
Open → Firm LeverageVlogs (YouTube), blogs, podcasts—staging here with honest "coming soon" dates. Long-form digests and email when we ship them.
Open → baur-software.github.ioWhat it is (mechanism, not hype): A mandate-based agent negotiation protocol: cryptographic mandate chains, bounded delegation, ephemeral session DIDs (session keys discarded), co-signed receipts that record what categories/types were disclosed, not the underlying content, and mandate decay where continuation requires active renewal. No new crypto primitives; the published design is built from standard pieces (WebAuthn, DIDs, VCs, SD-JWT, OHTTP, JSON-LD, Schema.org — as laid out in their stack table).
Why we care in law firms: Firm AI stacks are becoming multi-vendor agent workflows under confidentiality, supervision, and vendor ToS reality. Retention, training, and disclosure language isn’t “IT procurement” anymore — it’s “what evidence exists about who touched what, under what authority”. PAP is the rare framing that tries to make scope, minimization, session boundaries, and auditability the protocol’s job—not a PDF policy layer glued on after the fact.
Why Firm Leverage points lawyers here: We’re not trying to be the cleverest prompt writers; we’re trying to ship integrations and operating models that still make sense after the first real matter, the first subpoena-adjacent question, and the first angry ethics conversation. PAP is on this list because it’s the most serious protocol-first attempt we’ve seen to keep human principals in cryptographic control as agents start doing real work across boundaries.
Read → gatherdev.substack.comWhat it is (mechanism, not hype): Peter Bell’s essay (part of his How to Design an Agentic System series) treats a workflow like CI or an ops runbook, but where an LLM does the variable work inside fixed rails: explicit steps, verification gates before the next step runs, constraints that cap blast radius (what files, APIs, or branches are even touchable), concrete exit criteria for “done,” and escalation rules that force a human when limits hit. The spine he keeps returning to is deterministic shell, non-deterministic core—sequencing and gates are code; the model supplies judgment inside those boundaries. He’s also blunt about loops (generate → test → fail → retry): real agent work is cyclic, so the honest design is bounded cycles with retries and budgets—not a tidy DAG that pretends feedback never happens.
Why we care in law firms: A one-off impressive answer is not a supervised practice. Firms need systems where confidentiality boundaries, role separation, and review are encoded in how work moves—not as a PDF next to a chat box. When steps, gates, and human handoff are underspecified, you get the usual failure mode: slick drafts, skipped checks, and no credible afterward story about who authorized what for which client matter.
Why Firm Leverage points lawyers here: Sales demos sell model capability; durable adoption sells governed repetition. Bell is on this list because he writes for the gap everyone hits after the pilot: configuration as infrastructure—durable execution, graph orchestrators, or a script with gates between steps, matched to the risk you actually carry. We send people here when they’re ready to ask what production means under competence and confidentiality—not what the benchmark chart says.
Read → baursoftware.comWhat it is (mechanism, not hype): Todd Baur’s essay treats today’s dominant AI commercial model as a tollbooth: metered inference where the vendor owns compute, pricing, and what you can afford to run at scale (his analogy is long-distance minutes—a monetization shape enabled by infrastructure control, not a law of nature). The counter-shift is not only “cheaper,” it’s who controls inference, what leaves the machine, and what persists when local execution, open weights, and zero-trust-shaped routing are real options. He also names the harder design question: the agent layer operates on judgment formation, not only output—so the stakes are augmented autonomy versus delegated thinking, and what habit that trains.
Why we care in law firms: Firms are being sold multi-year cloud AI packages where cost, retention, and supervisory load scale with every workflow you push through someone else’s meter—and default vendor terms often still reserve retention or using inputs to improve vendor models, so the toll is not only dollars per token but what the contract permits about matter-adjacent data. That is procurement in dollars and confidentiality and competence in practice: the more judgment sits inside a vendor’s default capture path, the more your professional posture inherits someone else’s incentive stack. Whether or not every prediction in the piece lands on his timeline, the tollbooth framing is a sharp way to ask what you are signing before “lock in and tune later.”
Why Firm Leverage points lawyers here: Partners deserve systems language for commercial reality, not only feature matrices. Baur is on this list because the piece connects vendor economics, protocol- and locality-shaped alternatives, and supervision-grade consequences in one thread—and because the same shop has been shipping toward that architecture story (including PAP), not treating it as conference copy. We curate reads that still work when someone asks “why is our AI bill structured this way—and what are we optimizing?”
Read →This list grows. Frameworks and field notes from our work with law firms will appear here as we publish them.
If you've read or written something that belongs on this list, send it our way.
Book a call →