From pilot to billable work: how London professional-services firms should adopt AI

A London professional-services firm should adopt AI around one defined client workflow, not around a general instruction to use the technology. The first production use should have a valuable task, permitted data, an accountable owner, a documented human review and a baseline against which quality, time, rework and margin can be compared. A fast draft that creates extra checking or a confidentiality risk is not an improvement.

The commercial test is whether the redesigned workflow produces work that the client can rely on and the firm can defend. Time saved matters only when it becomes better capacity, faster delivery, a stronger fixed-fee margin or a more valuable service. The firm remains responsible for the advice, analysis and documents it supplies, whether a model helped to produce them or not.

Start with a client workflow, not a licence count

The most useful starting point is a recurring task with a clear input, output and reviewer. Examples might include classifying a controlled document set, extracting clauses for review, creating a first draft from approved source material, preparing a meeting record or checking a deliverable against a defined standard. The example must be adapted to the profession and the firm's obligations.

Map the present workflow before introducing a tool. Record who supplies the input, what evidence the practitioner checks, how exceptions are handled, how long each stage takes and what makes the final output acceptable. That baseline prevents an attractive demonstration from being mistaken for an operating improvement.

The government's AI adoption plan for professional and business services describes increasing adoption alongside persistent barriers involving skills, safety, transparency and implementation cost. Its sector figures are useful context, not evidence that a particular tool will work inside a particular firm. A managing partner still needs a workflow-level case.

Rank candidate tasks on four dimensions:

• value if the task improves;
• frequency and volume;
• cost of an error;
• sensitivity of the information involved.

The best first use is often important enough to measure but bounded enough to stop. High-consequence advice with uncertain source material is a poor unattended pilot.

Give each use case an accountable owner and risk class

Ownership cannot be delegated to a model vendor. One senior person should accept the intended use, data category, controls, reviewers and stop conditions. The business owner understands the client outcome; information security, data protection and the relevant professional function test the risk.

A simple classification can separate low-risk internal assistance from client-facing or high-consequence work. Ask whether the system:

• receives personal, confidential, privileged or market-sensitive information;
• influences a decision about a person;
• generates analysis that a client may act on;
• retrieves sources that can be inspected;
• sends data to another supplier or jurisdiction;
• can take an action in another system;
• is subject to a regulator, court, audit or contractual record requirement.

The ICO's artificial-intelligence guidance should inform uses involving personal data. The precise assessment depends on purpose, lawful basis, data flows, automated decision-making and the rights of affected people. A general corporate approval for one product does not settle every later use case.

Regulated firms must also apply their own rules. For example, the Solicitors Regulation Authority's warning notice on misuse of AI addresses supervision, accuracy and client confidentiality in legal practice. An accountant, architect, surveyor or financial adviser needs guidance from the regulator and professional body relevant to its own work rather than borrowing a law firm's control list.

Decide whether to buy, configure or build

Most firms should compare three options. A general enterprise tool can support common tasks but may offer limited workflow control. A specialist product can encode sector processes and evidence, but creates supplier dependence. A custom system can fit the firm closely, while placing more security, testing, maintenance and documentation obligations on the firm.

Due diligence should cover more than a product demonstration. Examine the contractual data terms, retention, model training settings, sub-processors, hosting locations, access controls, audit logs, deletion, incident handling, service continuity, intellectual-property terms and the firm's ability to export its records. Test what happens when a model or underlying service changes.

The National Cyber Security Centre's secure AI system development guidance organises controls around secure design, development, deployment and operation. Its lifecycle approach is useful even when the firm buys rather than builds, because configuration, integration and user behaviour can create risks that a vendor's general assurance does not remove.

A build decision needs a total cost, not only developer time. Include evaluation data, integration, security review, model or API usage, monitoring, support, retraining, vendor changes and the cost of keeping evidence available for later review.

Control confidential information before people start experimenting

The firm should state which tools and accounts are approved, which data categories may enter them and which uses are prohibited. Consumer accounts, personal logins and copied client files create a shadow workflow that can bypass contractual and technical controls.

Create a data-flow record for each production use. It should identify the source system, prompt or instruction, attachments, model provider, storage, retention, output destination, user roles and deletion process. Where personal data is involved, decide whether a data-protection impact assessment is required and document the conclusion.

Minimisation is a practical control. Remove names and irrelevant fields, retrieve only necessary document passages and keep the model away from an entire matter or customer database when a smaller approved set will do. Permission boundaries should follow existing client and matter access rather than creating a new shared pool.

Contracts with clients and technology suppliers may be stricter than data-protection law alone. Confidentiality, privilege, information residency, audit rights and subcontracting all need a specific check. A tool being described as private or enterprise-ready does not prove that a proposed use complies with those promises.

Design human review around the type of failure

"Human in the loop" is not a complete control. The reviewer must know what to check, have enough source evidence and possess the authority to reject or redo the work. If a junior employee is asked to approve an output outside their competence, the review step is largely decorative.

Build a review checklist from likely failure modes. For a research memo that might include factual support, missing contrary evidence, dates, source authority and citation accuracy. For a contract task, check document completeness, jurisdiction, defined terms, clause relationships and deviations from the agreed playbook. For numerical work, recalculate critical figures outside the model.

Keep an evaluation set of representative and difficult cases. Run the old and proposed processes on the same material, then compare accuracy, completeness, consistency, turnaround and reviewer effort. Record failures and update the instruction, source set or workflow. The tool should not enter live client delivery merely because a small set of favourable examples looked convincing.

The companion analysis of AI in London finance examines a more regulated setting. The shared principle is that a technical capability becomes commercially useful only after responsibility and evidence are designed into the process.

Decide what the client should know

Disclosure should follow the engagement, the materiality of the AI contribution, professional duties and the client's procurement or security requirements. A firm does not need to turn every grammar aid into a dramatic announcement. It should not hide a material change in how sensitive work is produced or reviewed.

Before starting an engagement, clarify permitted tools, restricted data, subcontractors, record keeping, approval responsibility and whether the client expects notice of material AI use. For an existing engagement, check whether the change falls within the agreed delivery and security model.

Provenance also matters. The file should show which client and approved sources informed the output, which version of a tool was used where relevant, who reviewed it and what was changed. A practitioner must be able to explain the basis for the final work without presenting a generated answer as its own authority.

Copyright and licensing questions are fact-specific. The government's copyright and AI consultation records policy work, but a consultation is not a permission to use protected material. Firms should assess the material placed into a system, the provider's terms, database rights, licences and the intended use of the output.

Convert time saved into a credible commercial model

AI can unsettle an hourly model because less recorded time may reduce the invoice even when the outcome improves. It can also weaken a fixed-fee model if the firm lowers price before understanding review, technology and exception costs.

Measure the full workflow first. If a drafting stage falls from four hours to one but review rises from one hour to three, only one net hour has been saved. If rework or claim exposure rises, there may be no commercial gain at all.

The firm can then choose how to use genuine capacity. Options include faster turnaround, a fixed fee with stronger margin, a lower-cost entry service, more senior attention on judgement, or a new monitoring or analysis product. Pricing should reflect client value and risk, not the number of model calls.

For work sold overseas, the operating and contractual questions in our guide to exporting professional services remain relevant. Cross-border client data, local professional rules and tax do not disappear when part of delivery is automated.

Train roles and change incentives

Training should be tied to approved workflows. Users need to know the permitted data, expected instruction, source standard, review checklist, escalation route and incident process. Partners and managers need enough understanding to supervise, not just enough access to request a summary.

Incentives also need attention. A utilisation target can discourage people from reporting time saved. A public adoption target can encourage use where it adds no value. Reward documented improvements in quality, turnaround, client outcome and reusable knowledge instead of prompt volume.

Define a route for mistakes and near misses. Staff should be able to report an exposed document, a fabricated authority or an unexpected system action without concealing it to protect an adoption metric. Incident learning belongs in the governance record and, where required, in client or regulator notification.

Measure quality, margin and rework together

A production scorecard should include at least:

• output acceptance against a defined quality standard;
• reviewer time and number of corrections;
• end-to-end turnaround, not model response time;
• rework after client review;
• technology and implementation cost;
• gross margin or capacity created;
• incidents, exceptions and prohibited-use attempts;
• client satisfaction or renewal evidence where available.

Compare a sufficient period before and after introduction and note other changes, such as a new team, template or client mix. Vendor claims and government adoption figures are not substitutes for the firm's own controlled evidence.

The planned guide to sustainability reporting requests for London suppliers shows why provenance matters beyond AI. Enterprise clients increasingly ask suppliers to substantiate operational claims. A firm should expect the same discipline around an AI-enabled service.

Scale only when the workflow meets agreed thresholds and the control owner signs it off. Keep a rollback option. The strongest proof of adoption is not a firm-wide licence but repeatable client work whose quality, responsibility and economics are understood.

Limitations

Professional duties differ across legal services, accountancy, audit, architecture, engineering, consulting and regulated financial work. This article cannot decide whether a particular use satisfies a firm's contracts, insurance, regulator or data-protection obligations. The law and regulatory guidance around AI, copyright and automated decisions continue to change.

Product behaviour, hosting and terms can also change after procurement. Every live workflow needs current technical and contractual verification. Productivity findings from one team or task should not be generalised to the whole firm without comparable evaluation.