AI in the Matter Lifecycle
A paper on artificial intelligence and law firm transformation. April 2026.
This paper is the third in a series. The first — “The Matter Lifecycle as Organising Principle” — establishes the matter lifecycle as the coordinate system within which all law firm operations take place. The second — “Why Data Must Come Before Systems” — addresses the data foundation AI requires. This paper maps AI onto the lifecycle specifically: where it can operate, where it cannot, and what firms need to build now to be ready for what is coming.
AI is present in almost every law firm transformation conversation in 2026. It is present in far fewer transformation programmes in any substantive sense. The gap is not primarily a technology problem. It is a framing problem: AI is being inserted into programmes without a clear account of where in the matter lifecycle it can operate, where it cannot, and what the consequences of getting that wrong are.
This paper addresses AI in the matter lifecycle specifically — not AI strategy in the abstract.
The Adoption Reality
Thomson Reuters Institute data shows enterprise-wide generative AI adoption across professional services moving from 14% in early 2024 to 43% by early 2026. Agentic AI — systems that take autonomous action across multi-step workflows — sits at 15% in widespread adoption, with a further 53% of firms actively planning or considering it.
The trajectory is clear: agentic AI is following generative AI’s adoption curve with an 18 to 24 month lag. Firms designing TOM programmes now that do not account for agentic AI are designing for a state that will be obsolete before the programme completes.
Two Types of Legal Work
The most analytically useful framework currently available for placing AI in the matter lifecycle distinguishes between two types of work:
Repeatable work: Contract review checklists, regulatory monitoring, compliance checks, sanctions screening, term extraction at scale, billing narrative review. The outcome is definable in advance. Best practices can be encoded. Agents execute consistently at scale without requiring human judgment on each instance.
Complex, open-ended work: Full contract suite analysis, end-to-end due diligence, fund formation lifecycle management, matter strategy. The outcome is not fully definable in advance. The agent generates a plan; the lawyer reviews and edits before execution. High-stakes decision points trigger a pause for human judgment before the agent proceeds.
This distinction maps directly onto the matter lifecycle:
| Stage | Work type | AI role |
|---|---|---|
| BD / Origination | Complex | Relationship intelligence, opportunity scoring — augmentation only |
| Conflict Check | Repeatable | Automated search and risk flagging — human decision mandatory |
| Intake / AML | Repeatable + mandatory gate | Risk scoring, data enrichment, CDD automation — human sign-off required by law |
| Matter Opening | Repeatable | Data population, rate configuration, ethical wall triggers — high automation potential |
| Active Delivery | Both | Time capture suggestions, document review, research — varies by task |
| Pre-Bill / Billing | Repeatable | Narrative checking, billing guideline compliance, anomaly detection — high automation potential |
| Matter Closure | Repeatable | Document archival triggers, knowledge extraction, precedent tagging — largely undeployed |
The stages with the highest current AI deployment are also the stages with the clearest repeatable-work characteristics. The stages with the lowest deployment — origination, active delivery in complex matters, and matter closure — are either complex and open-ended or have not yet attracted investment proportionate to their value.
The Gates AI Cannot Cross
Two points in the matter lifecycle are permanent human gates — not because the technology is incapable, but because the law requires a human to make the decision.
Conflicts (Stage 2): Professional conduct rules in every jurisdiction require a qualified lawyer to determine whether a conflict of interest exists and what to do about it. AI can search, flag, and score. It cannot decide. This is not a transitional position that will change as AI matures — it is a professional liability allocation that regulators have no current intention of altering.
AML/KYC (Stage 3): Under WWFT and the incoming AMLR (applying from 10 July 2027), a designated responsible person must make a documented, auditable risk assessment before a matter proceeds. AI-enriched risk scores, automated CDD data gathering, and integrated sanctions screening all assist that decision. None of them replace it. The personal liability attached to the AML sign-off — which sits with a named individual, not a system — is the mechanism that makes the gate permanent.
This has a precise implication for TOM design: STP targets must be set with these gates in mind. A programme that claims full STP across the matter lifecycle is either defining STP incorrectly or has not understood the regulatory architecture. Realistic STP targets acknowledge the gates, measure the time within each stage that can be automated, and treat human decision time as a fixed cost to be supported well rather than eliminated.
Agentic AI: The Governance Requirement
The distinction between generative AI (which responds to prompts) and agentic AI (which takes autonomous action across multi-step workflows) is architecturally significant for a TOM programme.
Agentic systems do not wait to be asked. They monitor state, identify triggers, and act. In a matter lifecycle context, an agent might monitor a matter’s risk score continuously, trigger a re-screening when a client’s ownership structure changes, and route the result to the responsible partner for review — without a human initiating any of those steps.
This capability is genuinely valuable. It is also genuinely dangerous if the governance architecture is not designed alongside it. The core requirement: “more autonomous does not mean fully autonomous.” The lawyer’s role at each AI-assisted step must be explicitly defined at TOM design stage — not retrofitted after implementation.
Five governance questions must be answered for every agentic AI deployment in the matter lifecycle:
(1) What triggers the agent? System event, time elapsed, data threshold, or human instruction? (2) What actions can it take autonomously? What requires human approval before execution? (3) What is the audit trail? Every agent action must be logged and attributable. (4) Who is accountable for the agent’s output? Partner accountability for AI-assisted work product is non-negotiable — the agent does not absorb the liability. (5) How is the agent’s behaviour constrained? What guardrails prevent it from taking actions outside its defined scope?
A TOM that deploys agentic AI without answering these questions has not deployed AI. It has deployed risk.
The Data Foundation AI Requires
Agentic AI cannot act reliably on data it cannot address. If client identity is distributed across systems with no common record, no agent can answer the question “who is this client, what is the firm’s complete history with them, and what does the relationship network look like?” with any reliability.
Two requirements determine whether AI deployment is additive or disruptive:
Data infrastructure: A matter lifecycle with fragmented data — client identity in one system, time records in another, billing history in a third, with no common matter identifier linking them — cannot be served by an agent that needs a complete picture of the matter to act intelligently. The data architecture decisions made in the current TOM programme determine whether agentic AI can be deployed incrementally or requires a structural rebuild.
Matter state visibility: An agent needs to know where a matter is in its lifecycle at all times — what stage it is at, what the next required action is, what the exception conditions are, and who is responsible for the next decision. Most current practice management systems record transactions but do not maintain explicit matter state. The transition from transaction recording to state tracking is a design decision that must be made in the current programme.
Firms that invest in these two foundations will be able to add agentic capability as it matures without rebuilding their operating model. Firms that do not will face a choice in 2028 between deploying AI on a foundation that cannot support it or restarting the transformation they did not complete now.
The 2028 Horizon
The “world model” concept — a system that observes and operates across an entire organisation, accessing all necessary data and workflows, triggered not by human prompts but by system monitoring — describes where agentic AI is heading. The implication for TOM design now is not to build for that state immediately, but to build the infrastructure that makes it possible: clean data flows, matter state tracking at each lifecycle stage, defined handoff criteria between systems, and an integration architecture that can accommodate new agents without requiring structural rebuilding.
The matter lifecycle is the frame within which AI operates in a law firm. A programme that builds the lifecycle design first, commits to the data foundation second, and deploys AI into a defined and governed structure third will produce compounding capability over time. A programme that deploys AI first and hopes the lifecycle and data will catch up has the sequence backwards.
The Argument
Three propositions for AI in the matter lifecycle:
One. AI belongs in the lifecycle at specific stages, with specific governance, and with permanent human gates at conflicts and AML/KYC. These gates are not transitional. They are structural. STP targets must be designed around them, not over them.
Two. The data foundation is not what you build before AI. It is what makes AI possible. Fragmented data produces unreliable agents. A dedicated data layer, governed by the firm, is the precondition for agentic AI that works at firm level.
Three. Agentic AI is following generative AI’s adoption curve with an 18 to 24 month lag. Firms designing TOM programmes now that do not account for agentic AI are designing for a state the market will have moved past before the programme completes.
Sources
- Thomson Reuters Institute, 2026 AI in Professional Services Report.
- Thomson Reuters Institute, Agentic AI Oversight Challenges, April 2026.
- Harvey, Two Types of Legal Work, One Agentic Platform, April 2026.
- EU AML Regulation (AMLR), applying from 10 July 2027.
- NOvA, Aanbevelingen AI in de advocatuur, 2025.
Part of a series: see also “The Matter Lifecycle as Organising Principle,” “Why Data Must Come Before Systems,” and “Law Firm Transformation: A Practitioner’s Guide.”