Building Better Professional Services with AI

The knowledge in a professional firm is not reducible to the documents on its server. Much of it resides in distinctions that experienced people have learned to make.

A credible professional knows that two sources of equal apparent authority may not deserve equal weight. A client's account can be accurate as testimony and misleading as evidence. A precedent can be technically relevant and practically inapplicable because the jurisdiction, institution, incentives or risk tolerance differ. A rigorous method does not remove judgment; it makes the use of judgment more visible.

This is why professional concern about AI is ultimately a concern about epistemology: how the firm knows what it claims to know. A system can retrieve a precedent without understanding why it was persuasive. It can summarize a methodology without recognizing where that methodology ceased to be appropriate. It can reproduce the language of confidence while being unable to accept responsibility for the consequences.

Over years of practice, a firm accumulates more than finished work. It accumulates methods, interview practices, source standards, client histories, failed approaches, proposal evidence and tacit knowledge about how decisions are actually made. If AI is introduced as a generic writing layer, detached from that body of knowledge, it will make the firm sound more fluent while making its work less particular.

The ethical boundary is equally important. Confidentiality, consent, intellectual property and client expectations cannot be delegated to a model. Nor can authorship. A person must remain able to explain why a consequential claim was included, what evidence supports it, what uncertainty remains and who approved its release. The presence of a human in the process is not sufficient; there must be meaningful human ownership of the judgment.

Research assembly, document comparison, retrieval, calculation and routine monitoring are means. Judgment and accountability are the obligations those means must serve.

Respect for professional judgment cannot become an assumption that a human being is the better instrument for every intellectual task simply because the human being is human.

People forget. We search unevenly, privilege recent experience and stop looking when a plausible answer appears. Our recall of earlier projects depends on who happens to be present. Under pressure, we reuse familiar sources, carry forward old language and make inconsistent choices about work that appears routine. Expertise reduces these failures; it does not abolish them.

A well-designed system may be able to search a firm's approved archive more broadly than any individual can remember it, compare more documents without fatigue and apply a specified check with greater consistency. It may notice that a current proposal resembles work completed twelve years ago, or identify a contradiction across evidence that no one person had reason to hold in mind. None of this gives the system wisdom. It does mean that human authorship is not, by itself, proof of superior care.

There is an ethical dimension to this counterargument. If scarce senior attention is consumed by retrieval, formatting and repeated assembly, less of it is available for interpretation, challenge and conversation with the client. If institutional memory remains dependent on individual recollection, the firm exposes clients to avoidable inconsistency. If recurring business-development or knowledge work disappears whenever delivery becomes busy, that too is a design choice with consequences.

The relevant comparison is not between an ideal professional and a reckless machine. It is between actual ways of working, each with its own failure modes. AI can hallucinate, overstate and execute a poor instruction at speed. Humans can overlook, tire, rationalize and allow important work to stop.

Many experienced professionals tried generative AI in 2024 or 2025 and found it unreliable, cumbersome or intellectually mediocre. That judgment was often sound. It is no longer sufficient evidence about what is possible now.

The important change is not simply that models write better sentences. Current systems can be given a bounded objective, approved context, access to defined tools and a sequence of work to complete. They can search across files, compare evidence, calculate, edit documents, monitor a recurring process and return an artifact with sources and exceptions for review. Persistent context also makes it possible to connect the work to a firm's own methods and history rather than begin each exchange from a blank chat.

These capabilities do not make the systems reliable by default. They change the unit that must be evaluated. The old test was often whether a prompt produced an impressive answer. The relevant test now is whether a specified, controlled workflow produces better work after the cost of supervision, verification and correction is included.

The familiar objections—confidentiality, hallucination, intellectual property, client expectations, quality control and over-reliance—belong inside the inquiry rather than outside it. Each should determine where information may go, what the system may do, what evidence it must return and where a person must intervene.

The evidence does not establish a return for any particular firm or show that automated advice is superior. It does show that AI is already inside reputation-sensitive professional work. Refusal is not a control: informal use can leave a firm carrying the exposure without gaining the operating benefit.

The counter-risk is also real. Competitors may combine equivalent judgment with broader research and faster execution. A firm's accumulated knowledge may remain difficult to retrieve. Opportunity monitoring, client follow-up and knowledge capture may continue to stop whenever delivery becomes busy.

To be AI-first does not mean asking a machine to do everything first. The phrase names a discipline of work design. Before scarce human time is assigned, the firm asks what a system could do independently, what it could prepare alongside a person and what must remain a wholly human act of interpretation, relationship or accountability.

The best first workflow is rarely the most dramatic. It repeats often enough to learn, contains a visible quality signal and fails safely before anything leaves the firm. Unreviewed client advice, autonomous outreach, payments and contracts are poor starting points. Opportunity monitoring, proposal assembly and project-close knowledge capture are better candidates because their outputs can be compared with approved evidence.

The first project record may save little time. The advantage appears when the next pursuit can retrieve relevant assignments, approved evidence and delivery lessons without asking who remembers them. That memory must remain permissioned, source-linked and exportable. A firm should not bury its intellectual property so deeply inside one vendor that it cannot reconstruct or move the corpus.

Staff need to know what information may be used, where it may be used, how far a system may act and who approves the result. A practical policy begins with operating permissions.

At minimum, the firm needs a named owner, approved enterprise terms, a simple data classification, least-privilege access, source lineage, known-result tests for important workflows and a way to export or re-route firm memory.

A maturity score creates false precision. The useful output is a practical answer to where AI adds value, what controls it requires and what the firm should test first. That requires inspecting actual work: recurring tasks, hand-offs, wait time, rework, client promises, systems, archives, permissions and the points at which scarce judgment is being spent on assembly.

By quarter end, the test is whether staff trust a few governed workflows because they have survived real client and firm work. Continue only if the pilots show net value after review, correction and risk controls.