The Legal AI 'Step-In Rights' Blind Spot: Why Your Contract Review Tool Is Telling You Half the Story
The clause is there. The AI found it. Your summary report shows a green tick next to "step-in rights identified." You move on.
The clause is there. The AI found it. Your summary report shows a green tick next to "step-in rights identified." You move on.
That's the problem.
AI contract review tools have become genuinely useful for high-volume clause detection. Products like Thomson Reuters CoCounsel Legal, Luminance, and Spellbook have demonstrated real value in surfacing provisions lawyers might otherwise miss across large document sets. But there is a specific, repeatable failure mode that practitioners in outsourcing and infrastructure work are beginning to encounter, and it is serious enough to warrant a direct conversation about where AI review ends and legal judgment must begin.
The failure mode is this: AI tools reliably identify that a step-in rights clause exists. They consistently fail to assess whether the performance thresholds that trigger it are defined with enough precision to actually be exercised.
What Step-In Rights Actually Need to Work
Step-in rights give a customer—or in some financing structures, a lender—the ability to take over delivery of a service when a supplier is failing. In outsourcing agreements covering IT infrastructure, facilities management, or critical business processes, they represent a last-resort operational lever. In project finance and PFI-style arrangements, lenders use them to protect debt service by stepping in before a project company defaults.
For that lever to work under pressure, two things need to be true. First, the right needs to exist. Second, the conditions triggering it need to be measurable at the moment you want to exercise it.
AI tools are very good at confirming the first. They are structurally blind to the second.
The distinction that matters in practice is between subjective and objective triggers. A clause that permits step-in upon "material breach of service levels" sounds robust. It is not. "Material" is a legal characterisation, not a measurement. When a supplier is failing and you want to exercise step-in rights at pace—potentially under time pressure, with operational continuity at stake—you will first have to establish that a breach exists, then argue that it is "material," and almost certainly deal with a supplier who disputes both conclusions. The clause exists. The right is functionally unusable under any time horizon that matters.
Contrast that with a clause triggered by "three consecutive calendar months in which uptime falls below 95% as measured by the monitoring tools identified in Schedule 7." That trigger is binary. Either the condition is met or it is not. Exercise of the right becomes a notification exercise, not a legal argument.
Why AI Summaries Treat These as Equivalent
The architecture of AI contract review tools explains the failure. These tools are built to classify and extract: does the document contain a step-in rights provision, and if so, what are its general parameters? That classification task they perform well. What they cannot do is evaluate the operational sufficiency of a trigger condition—whether it will yield a clear, unambiguous answer under real-world conditions when a supplier's lawyers are looking for exits.
Illustratively, consider a GC reviewing a summary of a managed infrastructure agreement. The AI output reads: "Step-in rights are present. Triggered upon service level breach. Customer must give 30 days' notice." That summary is technically accurate. It is also useless as a risk assessment. The question that determines whether the clause has any value—is the service level breach defined in a way that can be established without litigation?—is entirely absent.
This is not a criticism of the tools' stated capabilities. It is a criticism of how practitioners are being trained to interpret their outputs. A clause detection summary is not an enforceability assessment. When lawyers treat it as one, they are substituting a presence check for an adequacy analysis.
What Goes Wrong in Outsourcing and Infrastructure Contracts
The consequences show up at the worst possible moment. Suppliers in distress are not passive. When a customer attempts to exercise step-in rights against a failing service provider, the first response from the supplier's legal team will be to attack the trigger. If the trigger relies on a subjective characterisation—"material breach," "persistent underperformance," "failure to meet service expectations"—that attack has real traction. Disputes about whether the condition is met can, and do, run longer than the operational crisis the step-in right was meant to address.
In complex outsourcing agreements, this problem compounds because the service level schedules and the step-in rights clause are often drafted at different times, by different teams, with limited cross-referencing. The SLA schedule may contain precise metrics. The operational services agreement may refer to those metrics obliquely, or may have been updated through amendments that the metrics section never caught up with. An AI tool reviewing the agreement as a whole will identify the step-in clause and the SLA schedule separately. It will not flag that the trigger in the former no longer maps cleanly onto the metrics in the latter.
What Practitioners Need to Do Instead
The answer is not to stop using AI review tools. It is to be precise about what they have done when you receive their output.
AI detection of a step-in clause should prompt a structured human analysis of three things. First: is the triggering condition objective or subjective? If it relies on a characterisation (material, persistent, significant), treat it as requiring litigation to exercise and advise accordingly. Second: does the trigger cross-reference measurable metrics that are defined elsewhere in the agreement, and do those definitions still hold after amendments and schedules are read together? Third: is there a measurement mechanism—a designated monitoring tool, an agreed reporting cadence, a specific data source—that produces the evidence needed to invoke the right?
ABA Formal Opinion 512, issued in July 2024, addresses AI competence requirements for lawyers, including the duty to understand the limitations of AI tools being used. The practitioner who relies on an AI summary to sign off on the adequacy of a step-in rights clause without conducting this analysis is not satisfying a competence standard. They are outsourcing a legal judgment to a classification tool.
The Right Frame for GCs and Transactional Lawyers
Step-in rights are insurance. Like all insurance, their value depends entirely on whether the conditions for paying out are met. An AI tool that confirms the policy exists but cannot tell you whether it will pay out is not doing the job that the stakes require.
For GCs reviewing complex services agreements at execution stage, and for transactional lawyers advising on outsourcing or infrastructure mandates, the step-in rights clause is not done when the AI finds it. It is done when a human lawyer has confirmed that every trigger condition in it produces a yes/no answer that can be established from the contract record, without argument, on the day you need to exercise it.
That analysis is not something a current AI review tool performs. Build it into your process or inherit the consequences.