Inside A&O Shearman's $1.1 Billion AI Bet: 3% of Expert Time

A&O Shearman announced a five-year AI commitment valued at $1.1 billion on November 5, 2026. The consequential choice is not a software purchase. The firm says the majority of the stated investment will come from every practice allocating at least 3% of its time to developing specialist AI tools and services. This is valued professional capacity, not a disclosed $1.1 billion cash expenditure. The firm also reported strong participation in September: 96% of partners and 93% of lawyers used its AI tools, with more than 558,000 activities across ContractMatrix and Harvey workflow tools. It reported ninefold growth in ContractMatrix client uptake over the prior 12 months. These figures are self-reported and unaudited. They show participation and demand signals, not proven productivity, profit, quality, or customer value. The operating idea is still important. Protected domain time can turn expert judgment into a reusable, versioned asset instead of leaving AI adoption at the level of licenses and logins. Enterprises should test that model through a tightly governed 90-day 3% Domain Capacity Pilot. They should select one revenue-critical workflow, reserve expert capacity, run the artifact in shadow mode, and use predetermined production and stop gates. Expert time should be treated as an explicit investment. Its return is reusable asset yield and verified outcomes, not activity counts.

What the $1.1 billion figure represents

A&O Shearman's November 5 announcement describes a $1.1 billion commitment over five years. It says the majority will come from each practice investing at least 3% of its time in specialist AI tools and services for its area of law. The announcement does not disclose a $1.1 billion cash budget or explain how the firm valued that professional time.

That distinction matters. Time from scarce experts carries an opportunity cost. It can reduce near-term billable or productive capacity. It also creates an incentive to label ordinary work as investment activity. A governed allocation makes both costs visible. It asks leaders to specify what asset the time must create, who can release it, and what evidence will justify continued funding.

The firm says its approach combines lawyers, data scientists, and developers, with qualified lawyers retaining responsibility for advice and governance committees overseeing deployment. It also says ContractMatrix client uptake grew ninefold in the preceding 12 months. These are company-reported descriptions and operating signals. They do not establish the return on the time allocation.

Why protected capacity can produce an asset

A license gives an expert access to a general tool. Protected capacity gives the expert a mandate to encode practice-specific judgment into something the organization can use again. The intended asset can be a versioned workflow, evaluation set, decision rule, structured knowledge module, or service component. It should have an owner, change history, defined users, and a release standard.

Without protection, immediate client work usually wins. Experts may use AI, but they rarely have enough uninterrupted time to define exceptions, test edge cases, document controls, and make a workflow reusable. The 3% allocation is therefore best understood as a portfolio rule. It forces domain knowledge into the investment process. It is not evidence that 3% is the correct level for every organization.

Bloomberg Law reported that A&O Shearman expects hourly billing to become less important, although it will not disappear, and that the firm's share of income from fixed fees rose about 15% over two years. Bloomberg reported the firm's statements; it did not independently audit the valuation or operating metrics. The shift still explains the economic logic. Reusable assets matter more when revenue can become less tightly coupled to hours worked.

Separate participation from results

The firm's reported September usage is extensive: 96% of partners, 93% of lawyers, and more than 558,000 activities across ContractMatrix and Harvey workflow tools. Usage and activities are input measures. They can show reach, fluency, or system load. They cannot show whether work became faster, safer, more profitable, or more valuable to a client.

Client uptake is closer to an outcome, but it still needs a denominator, retention evidence, commercial terms, and customer results. A useful operating model separates three levels:

• Inputs: allocated expert hours, active users, activities, training, and build capacity.
• Outputs: one tested, versioned workflow artifact; completed evaluations; documented exceptions; and releases.
• Outcomes: cycle time, rework, error rate, contribution margin, and a defined customer result.

The 90-day 3% Domain Capacity Pilot

This is a Stratex-authored implementation framework. It is not an A&O Shearman program, and 3% is a testable allocation rather than a universal optimum.

Days 1 to 15: select and baseline. Choose one revenue-critical workflow with repeated demand, stable enough inputs, a named customer or internal user, and errors that can be detected. Exclude safety-critical decisions and work where shadow evaluation would expose protected data without approval. Reserve 3% of participating experts' capacity in calendars and workload plans. Do not count normal tool use toward the allocation.

Assign four accountable roles. The domain experts define judgment, exceptions, and acceptance examples. The automation builder creates and versions the artifact. The risk owner defines data, quality, security, and stop controls. The product lead owns the user problem, evidence, economics, and day-90 decision. One person may hold two roles in a small team, but no role may be unowned.

The required deliverable is one reusable, versioned workflow artifact with an owner, scope, change log, test set, operating instructions, and rollback path. Establish comparable baselines for end-to-end cycle time, rework, detected error rate, contribution margin, and one customer outcome. Record volume, work mix, and exceptions so the comparison is not distorted.

Days 16 to 45: build and verify. Build the smallest artifact that can complete the selected workflow. Version every material change. Test normal cases, known exceptions, and deliberately difficult cases against the baseline process. The risk owner can halt the work for a control breach or evidence gap.

Days 46 to 73: run four weeks in shadow mode. Run the artifact beside the existing process without allowing it to control the customer deliverable. Compare like-for-like cases. Record disagreements, human corrections, cycle time, rework, errors, customer outcome, and the hours consumed by building and oversight. Do not convert a shadow result into a production claim.

Days 74 to 90: apply the gates. Move to limited production only if the artifact passes the predetermined data and security controls, meets the error tolerance, improves or preserves the customer outcome, and has a credible contribution-margin path after build and oversight cost. Stop immediately for a material confidentiality or security breach, an error beyond the agreed severity threshold, repeated untraceable output, or loss of accountable human review. At day 90, choose continue, revise, or stop. Continue only when verified outcomes and reuse potential justify the expert time.

A compact decision scorecard

In prose, the scorecard asks six questions. Did cycle time improve? Did rework stay controlled? Did errors remain inside tolerance? Is the contribution-margin path credible after all costs? Did the customer outcome improve or hold? Did the team create a versioned artifact that can be reused? Each answer needs a baseline, target, observed result, evidence owner, and continue, revise, or stop status.

The trade-off must remain visible

Protected expert time is not free. It reduces capacity available for current customers or other productive work. A percentage target can also be gamed through weak time classification, low-value prototypes, or activity inflation. Leaders should audit the artifact and evidence, not reward hours booked to AI.

The investment is justified only when one governed artifact can be used again and when customer, quality, and economic outcomes clear explicit gates. If the work produces logins, activities, or demonstrations without an accepted asset and verified result, stop or redesign it.

A&O Shearman's announcement makes domain expertise visible as investment capital. That is the idea worth testing. The practical standard is not adoption at scale. It is reusable asset yield and verified value after the opportunity cost of expert time is counted.

For a related framework on distinguishing activity from system outcomes, see Stratex's software delivery load test.