Legora and Spellbook both shipped tools that make AI follow a legal team's own playbook, Ivo open-sourced a contract model whose 91% score is its own, and a $60 million seed round for a self-described non-AI legal services company shows where outsourcing money is heading.
Legora launched Skills on October 6, and Spellbook put its Autonomous Contract Management product on general release a day later. Both are bets that the missing ingredient in AI contract work is a team's own rules, written down where the agent can read them. Legora's CEO Max Junestrand framed the problem as "It's not wrong. It's just not how we do it here."
Both launches confirm the diagnosis: the AI was never the scarce part, the playbook was. Neither says who maintains the rules once the first author leaves, or what happens when a deal falls outside them. A tool that reads your playbook is still a tool a lawyer has to drive. The inexpensive work, standard NDAs and low-risk redlines, still arrives at an expensive resource first. Outsourcing it to supervised agents means the routing, the playbook and the escalation rules sit in one place, with a lawyer reviewing finished work.
Teddy AI, a New York compliance and legal services platform incubated by the holding company Tucker's Farm, announced a $60 million seed round on October 7. Founder Kyle Tucker said, "Teddy is not an AI company," and that technology will be used "to the extent it helps." The company says it passed $25 million in business-to-business revenue in 2026 and drew about $115 million in equity interest over three weeks.
This is a company press release, and every figure above comes from it. It matters to in-house teams less for what it proves than for what it shows: investors are putting real money behind outsourced legal services, while the question of how much of that work AI actually does is left open.
When a provider sells you outsourced legal work, can it tell you which parts a model does, which parts a lawyer does, and who reviews the model's part? If the answer is "technology, to the extent it helps," you are buying a headcount, not a process, and the price will reflect it.
Ivo, working with River AI, released Sage, a version of DeepSeek V4 Flash post-trained on long, multi-step contract tasks using public data and lawyer-generated synthetic data. Ivo says the weights are free to download, run and build on, and that the next leap "won't come from bigger models." The reinforcement training lifted the base model's score on the Legal Agent Benchmark contracts set, per Ivo.
The cost claim is only "a fraction" of frontier pricing. The per-task dollar comparisons circulating this week were estimated by a chatbot, not measured, so they are not repeated here. The sources also give the base model's name two ways and do not confirm the licence terms.
A "pass criteria" score of 91% still leaves roughly one criterion in ten unmet. Would you sign an NDA reviewed at that rate with no one checking it? If strong contract models are now free to run, the value moves to the layer around them: your rules, your escalation paths and the review before anything goes out.
Olga Mack, CEO of the contract-rating company TermScout, argued in Above the Law on October 8 that contracts are shifting from records people consult after a problem to inputs that decide what happens next. An agent handling a software renewal could check auto-renewal, permitted price rises, unused licences and notice periods before it asks a person. She is arguing from a vendor's position, and the piece gives no figures.
An agent can test this against a date and act. It needs clean data, a clear version and a link back to the governing clause.
An agent can flag this, but a person still decides what it means. Mack says humans keep approving exceptions and bearing responsibility.
The sorting in this piece is the whole routing question. Direct rules, like payment terms and renewal dates, are inexpensive work that expensive resources still do by default. Interpretive standards are where a lawyer earns the rate. Supervised agents take the first kind end to end and escalate the second with the clause and context attached, so the lawyer spends time only where judgment is needed.
Legora and Spellbook want to read your playbook, Ivo made the model underneath cheap to run, Teddy raised money to sell outsourced work without saying how much is AI, and Mack says contracts will become instructions for agents. The model is no longer the scarce part. The rules and the review are.
Every story this week points at the same gap: inexpensive work is still done by expensive resources, because writing the rules down is not the same as having someone route work by them. A playbook inside a drafting tool leaves the lawyer as the router. A cheaper model leaves the review undone.
Outsource legal work to supervised agents. Flank's agents know your templates, terms and escalation rules, take the NDAs, standard redlines and procurement agreements, and hand finished work to a lawyer for review before anything leaves.