Legal AI tools have gotten very good at generating text. They can draft an NDA in seconds and summarize a contract with one click. But here’s the problem: drafting was never the bottleneck.
The real time sink in legal work is validation. Checking compliance, verifying risk alignment, ensuring every clause matches policy, and tracking decisions across redlines and review cycles. That’s where hours disappear, and where most AI tools don’t place their focus. They generate fast, but they can’t always prove their work. An issue that we’ve seen increasingly rearing its head in discussions is that legal teams are left re-checking everything anyway, often spending more time validating AI output than they would have drafting from scratch.
Avokaado has spent nearly a decade building contract intelligence infrastructure, treating contracts as structured data and not static files. Now, with the launch of Avo, they’re applying that foundation to solve what CEO Mariana Hagström calls “the black box problem” of generative AI.
It’s what Avokaado calls an “AI playbook engine”, a system of governed agents that operate within defined rules, escalate when needed, and produce auditable decisions through something called the 5D Engine™. The promise? Legal teams can finally productize their expertise: building autonomous assistants that handle entire contract lifecycles without guessing, hallucinating, or creating more work.
We met with Mariana to understand how Avo actually works, why validation matters more than generation, and whether this marks the shift from contract intelligence to true operational intelligence.
You’ve framed the core problem as “the real bottleneck isn’t generation; it’s validation.” Most vendors are racing to generate contracts faster, but you’re saying that’s solving the wrong problem. Can you walk me through a concrete example?
Mariana: “I think there’s a fundamental misconception about legal work. Because the outcome usually ends up as a document, the profession is often simplified to “drafting legal text.” But drafting is maybe 10% of the actual effort. The real work — and the real bottleneck — is in the remaining 90%: validation, ensuring a contract is correct, compliant, risk-aligned, consistent with policies and regulations, and acceptable to the business. And building technology for validation is significantly harder than generating text with an LLM.
Take something as basic as an NDA. AI can draft one instantly — tools like Legora or Harvey do that exceptionally well, and text generation is a smart first step for the market because there’s no infrastructure barrier for adoption. But once the text exists, the real workflow begins. Legal must check confidentiality carve-outs, governing law, IP ownership, non-solicit language, term and termination, and ensure the entire agreement matches context and internal standards. Then it goes to the business, comes back to legal, then perhaps procurement or security. Redlines start circulating. Five people open the same Word file. Days pass.
So time isn’t lost in writing — it’s lost in verifying, reviewing, negotiating, aligning, tracking decisions, proving compliance. I’m speaking mainly about in-house legal here, where speed and consistency directly impact revenue and risk.
This is why we’re building Avo – not as another drafting tool, but as an end-to-end legal agent engine. It doesn’t require complex prompting or scripting; it generates context and rules directly from the legal team’s existing playbooks and knowledge structures. Instead of just drafting faster, legal teams can finally productise their expertise — create autonomous assistants for the business, iterate rules, set guardrails, and define exactly where AI must stop rather than guess. Avo turns legal knowledge into a governed system, not a probabilistic text generator.
When we say Avo doesn’t hallucinate, we mean it also knows when not to act. It only moves within rules and governance boundaries. And we can do this because we’ve spent nearly a decade automating legal and business workflows with deterministic, rule-based systems — we understand deeply how legal decisions are made and how to translate them into machine logic. Avo brings that discipline into AI. It’s validation-first automation — finally closing the gap between rule-based precision and AI-driven efficiency.”
It’s a compelling reframe: the profession isn’t “drafting legal text,” and the technology shouldn’t be either. If 90% of legal work happens after the first draft, then building AI that only solves the 10% isn’t actually solving much at all. What was particularly interesting was Mariana’s emphasis on “validation-first automation”, the idea that knowing when not to act is just as critical as acting quickly.
The 5D Engine has many parts. How do these layers work together, practically? For example, negotiating liability caps in an MSA?
Mariana: “The way I usually explain it is that 5D mirrors how a lawyer already thinks. We didn’t build a set of micro-agents — one for drafting, one for review, one for negotiation — because in reality legal work isn’t fragmented like that. When you negotiate an MSA, you’re not just writing text; you’re evaluating risk, checking policy, comparing fallback clauses, pulling data, validating approvals, and routing documents operationally. So the 5D Engine bundles all those dimensions into one governed agent for a specific contract type.
If we take your example — negotiating liability caps in an MSA — here’s what happens practically:
D1 (Deal Context) understands industry, contract value, risk level, governing law, vendor vs customer position. If liability cap should be 1x fees for low-risk SaaS but 2x for financial data handling, D1 sets the frame.
D2 (Document Map) tells the agent what the clauses are and where they live, how it interacts with indemnity, limitation of damages, insurance requirements — all the internal cross-references a lawyer holds in their head.
D3 (Data Model) provides the variables: contract value, data sensitivity level, jurisdiction, renewal terms, whether PII is processed. This gives structure instead of free-text guessing.
D4 (Decision Rules) is where the policy logic lives.
For example:
– If contract value > €500k → escalate cap beyond 1x fees
– If personal data category = special category → require cyber insurance + breach notification window ≤ 72h
– If counterparty proposes unlimited liability → auto-generate fallback clause & escalate to legal
The agent never guesses — no rule, no action.
D5 (Dialogue Strategy) manages the negotiation style.
If the counterparty pushes back, the agent explains policy, offers fallback language, and if thresholds break — hands it to legal with context.
So instead of an LLM improvising, the 5D Engine orchestrates AI with structure — context → map → data → rules → interaction. Lawyers can make it as light or as deep as they need; an NDA might have 10 rules while an MSA could have 200. The engine doesn’t limit creativity — it gives legal teams a scalable framework to build as far as they want. No vendor tickets, no waiting — in an AI-native system the only limitation is your own knowledge.”
What’s striking here is the deliberate rejection of micro-agents. Rather than fragmenting legal work into discrete AI tasks, Avokaado mirrors how lawyers actually think, which is holistically, with context and constraints always in view. The liability cap example makes the architecture tangible: it’s not just rules firing in sequence, but layers of intelligence operating simultaneously to produce outcomes that legal can defend.
You’re solving AI’s black-box problem. You’re still using probabilistic models. How did you architect around non-determinism?
Mariana: “We don’t try to make the LLM deterministic — instead, we contain it and govern it. Avo is architected so that the AI never acts alone. The model generates options, not final decisions. The rules layer determines the context and what’s acceptable, the Verifier checks compliance before anything is approved, and the Scorecard records every step and rationale.
Think of the LLM as the creative engine, and the 5D framework as the governance cage around it. Avo doesn’t hope the model chooses the right clause — it guides the AI through structured rules, ensuring outcomes are explainable, auditable, and repeatable. We use AI for what it’s brilliant at, while 5D enforces the discipline legal requires.”
The “governance cage” metaphor is apt. Avo doesn’t try to fix the LLM, it contains it. That’s a fundamentally different approach than hoping prompt engineering or fine-tuning will make models more reliable. By positioning the LLM as the creative engine rather than the decision-maker, Avokaado sidesteps the determinism problem entirely.
The Scorecard — how are early access users using it? Efficiency tool or governance infrastructure?
Mariana: “The Scorecard is designed to start as a review accelerator — one view that shows what changed, who made the change, and why, without digging through redlines or emails. But structurally, it is built to evolve into something much more powerful: a governance and audit layer for legal AI operations. Once live, we expect teams to use it not just for faster review, but for:
• compliance and audit evidence
• policy-alignment monitoring
• vendor and contract risk comparison
• systematic improvement of playbooks
• onboarding and training support for legal teams
We see Scorecard becoming the layer that transforms legal AI from text generation into traceable, explainable, defensible decision-making. The review-speed benefit is immediate — but the long-term value is in governance and accountability at scale.”
To us, this means that the Scorecard is designed to grow with organizational maturity. It starts as a practical review tool – no one wants to dig through email chains – but its real potential lies in what comes next: audit trails, policy monitoring, systematic improvement. If legal AI is going to scale beyond individual tasks, it needs this kind of accountability layer built in from the start.
You mentioned agents know when to escalate. How are escalation triggers defined and customizable?
Mariana: “Escalation lives inside D4 – Decision Rules, and it’s built to be fully configurable per organisation, contract type or even per agent. Instead of hard-coded logic, teams define when automation can act alone, when it must pause, and when a human must step in. Think of it less like “AI making decisions” and more like a control panel for risk.
In practice, rules may look like:
• Liability cap exceeds policy threshold → escalate to legal review
• Counterparty in high-risk jurisdiction → require compliance approval before continuing
• Missing DPA or KYC document → block signature
• Redline to a protected clause → escalate to legal
Every parameter — thresholds, approvers, conditions, exceptions — can be adapted to how legal operates in that organisation.”
Escalation as a “control panel for risk” is the right framing. Legal teams don’t want black-and-white automation; they want configurable boundaries that reflect how their organization actually manages risk. The flexibility here, with defining thresholds per contract type, per jurisdiction, per clause, suggests Avo is built for legal reality, not idealized workflows.
Previously Avokaado was described as the bridge to operational intelligence. Is Avo the fulfilment of that vision?
Mariana: “Yes — Avo is the realisation of that vision. For almost a decade, Avokaado has been building the infrastructure to treat contracts as data, not just documents. That foundation is what makes governed AI possible. Instead of generating text in a vacuum, Avo operates inside structured context — parties, clauses, rules, history, risk parameters.”
Core laid the groundwork.
The 5D Engine turns that into operational intelligence.
With Avo, contract data becomes active, not archived. Agents draft, review, negotiate, route, escalate, and prove decisions with traceable logic. It’s the shift from document workflows to policy-driven automation that learns and adapts.
This is how legal moves from support function to strategic operator — not by replacing lawyers, but by scaling their judgment across the organisation.”
