AI for Professional Services
The sector where an AI mistake is signed by a professional and filed under their name. Nine verticals, written for the partner who carries the liability rather than the analyst writing the disruption thesis.
Two things make this sector different, and both are uncomfortable.
The first is liability. In most industries a model error is an operational problem. In professional services the output is delivered under a professional's name, into a court filing, an audit opinion, a tax position or a client recommendation, and the professional carries it. The evidence for how badly that goes is unusually precise: as at 5 September 2026 the AI Hallucination Cases database recorded 2,022 decisions worldwide in which a court found a party had relied on fabricated AI content — 1,379 in the United States, 217 in Canada, 110 in Australia, 69 in the United Kingdom. No other sector has a public tally of its own AI failures this exact.
The second is that AI attacks the unit of sale. A business that bills hours and deploys technology that removes hours has a commercial problem its own efficiency creates, and pretending otherwise makes for bad advice. We have told firms that the productivity case they were shown implies a revenue reduction nobody in the room had modelled. That conversation is more useful early than late, and it usually changes what gets built rather than whether anything does.
Where AI actually earns its place in professional services
Ranked by evidence rather than by how often it appears in a firm's innovation update. The maturity column is our own read; the catch column is what the vendor demonstration leaves out.
| Where | What it does | Maturity | The catch |
|---|---|---|---|
| Document review and extraction | Pulls facts, clauses, dates and figures from large document sets. | Proven | The strongest case in the sector, and the least discussed. Errors are checkable against the source, which is the whole point. |
| Retrieval across the firm's own work | Finds precedent, prior advice and internal knowledge with provenance. | Proven | High value. The hard parts are confidentiality boundaries and telling current material from superseded. |
| Drafting from firm precedent | Produces first drafts constrained to the firm's own templates and language. | Strong | Real time saving. Constrained to your own material it works; open-ended generation is where citations get invented. |
| Contact centre assistance | Suggests answers and drafts responses for agents in real time. | Proven | Well-evidenced, particularly for less experienced agents. Deflection claims are usually softer than assistance claims. |
| Transcription, notes and summarisation | Turns meetings, interviews and calls into structured records. | Proven | Mature and immediately useful. Consent and recording law is the constraint, not accuracy. |
| Candidate and applicant screening | Ranks or filters applicants against a role. | Regulated | High-risk under the EU AI Act from December 2027, and already subject to bias audit and notice rules in several US jurisdictions. |
| Machine translation with post-editing | Translates at volume with human revision. | Proven | Genuinely transformed this work. The unresolved question is commercial, not technical. |
| Research and analysis synthesis | Summarises sources into an analytical position. | Good | Useful with provenance and a competent reviewer. Sold as an analyst replacement, it fails on the judgement it cannot show. |
| Synthetic respondents | Simulates survey participants instead of fielding to people. | Contested | Reproduces what a model has read about a population, not what that population thinks. Directional at best, and not evidence. |
| Autonomous professional advice | Produces client-facing advice without professional review. | Not credible | The liability does not transfer to the model, and the 2,022 court cases are what happens when firms act as if it does. |
The pattern in the failures is narrower than the headlines suggest
Almost every AI failure in professional services has the same shape: open-ended generation, no source constraint, no verification step, delivered under a professional's name. The same technology constrained to the firm's own documents, with provenance on every claim and a competent person reviewing before it leaves the building, has an unremarkable safety record. It is worth noting that in the hallucination database self-represented litigants account for more entries than lawyers do — 1,163 against 805 — which tells you the failure is about process rather than about the tools being inherently unusable. Firms that put the process in place are getting the productivity without the sanctions.
Professional obligations first, then the AI-specific rules
This sector answers to professional conduct rules, client confidentiality and privilege, and a growing body of AI regulation concentrated on employment decisions. This is our reading as at September 2026 and we work alongside your compliance, risk and general counsel functions rather than in place of them.
The oldest rule is the one that bites
Duties of competence, candour to the court or regulator, and supervision apply to work produced with AI exactly as they apply to work produced by a junior. Courts have been consistent that the obligation to verify what you file does not change because a machine drafted it. Most sanctions in the hallucination cases turn on failure to check rather than on use of AI as such.
The question is where the data goes
Client confidential and privileged material entering a third-party system raises confidentiality, privilege and often engagement-letter questions before it raises a technical one. Deployment model, data retention, training use and jurisdiction are the terms that matter, and they belong in procurement rather than in a pilot.
Recruitment is where this sector meets Annex III
AI used for recruitment, selection, promotion, termination, task allocation and monitoring of workers is high-risk under Annex III. Following the Omnibus agreement reached in May 2026, those obligations apply from 2 December 2027. Article 50 transparency obligations were not deferred and applied from 2 August 2026 — which reaches client-facing chat and generated content across this sector, not only HR systems.
Already in force, and varying by state
New York City requires an annual independent bias audit, published results and candidate notice for automated employment decision tools. Illinois requires notice where AI is used in employment decisions and prohibits using zip codes as proxies for protected classes, effective 1 January 2026. California's employment regulations took effect 1 October 2025, making evidence of anti-bias testing relevant to claims and defences and extending record retention for automated decision system data to four years, including training data. Colorado's AI Act was postponed to 30 June 2026.
Frequently the binding constraint
Many client agreements now restrict AI use, require disclosure, or prohibit client data leaving named environments. In our experience this constrains firms more immediately than any statute, and it is often discovered after a tool has been deployed. Reviewing the engagement terms of your largest clients before selecting a tool is unglamorous and it prevents an expensive reversal.
The constraint on the easiest win
Transcription and meeting capture are among the most reliable applications here and sit under recording consent law that varies by jurisdiction, plus data protection obligations covering retention and access. This is a solvable problem and it is not a footnote — it determines whether a tool can be switched on for a multi-jurisdiction client base.
What this means for a build
Three design consequences. Generation should be constrained to the firm's own material with provenance on every claim, because open-ended generation is the mechanism behind almost every reported failure. A verification step belongs in the workflow rather than in the training, since the sanctions record turns on failure to check. And anything touching hiring needs a documented position before deployment, because in several jurisdictions it is already regulated and in the EU it becomes high-risk. See AI policy development and AI risk assessment.
Who we write for
Each page starts from that organisation's own problems, names the regulatory exposure it carries, and routes into the engineering. Depth varies and is stated on each page.
Legal & law firms
Document review, precedent retrieval and constrained drafting — built around the verification step the sanctions cases turn on.
Read more →Accounting & tax
Transaction extraction, explainable anomaly detection and research over verified sources — documented to a working paper standard.
Read more →Management consulting
Engagement knowledge retrieval, research synthesis and proposal production — plus the pricing conversation efficiency forces.
Read more →HR & recruiting
Screening, matching and HR operations — built around rules that are already in force, not arriving.
Read more →Staffing agencies
Database reactivation, matching and fill-rate analytics — with the obligations that attach to the provider, not just the employer.
Read more →Marketing & advertising agencies
Content production, media analytics and reporting automation — with rights and disclosure settled before the work ships.
Read more →BPO & call centres
Agent assistance, quality analytics and process automation — with deflection claims separated from assistance evidence.
Read more →Market research
Open-end coding, qualitative analysis and panel quality — with a clear position on synthetic respondents.
Read more →Translation services
Quality estimation, domain adaptation and effort routing — plus the pricing model this technology already changed.
Read more →FAQ
Marked up with FAQPage schema so these answers can surface in search results and inside AI assistant responses.
Do you have professional services experience?
Yes in document extraction and review, retrieval over firm knowledge bases, contact centre assistance, and analytics for agency and consulting operations. Less in audit-specific tooling, and none in practising law, accountancy or actuarial work — we build systems for professionals rather than performing their work, and we would not claim otherwise. Each of the nine vertical pages states our depth in that area rather than implying uniform expertise across a sector spanning a magic circle firm and a translation agency.
How do we avoid ending up in the hallucination database?
By changing the process, not by choosing a better model. The 2,022 recorded cases share a shape: open-ended generation, no source constraint, no verification step, and delivery under a professional's name. Constrain generation to your own documents, require provenance on every factual claim, and put a verification step in the workflow rather than in a training session. Firms doing that are getting the productivity without the sanctions, and it is worth noting that self-represented litigants account for more of those cases than lawyers do — the failure is procedural.
If AI makes our people faster, does that not reduce our revenue?
In an hourly model, yes, and it should be modelled rather than avoided. We have shown firms a productivity case that implied a revenue reduction nobody present had calculated. The useful responses are real ones: take on work that was previously uneconomic, shift the pricing basis on the work AI most affects, or accept a lower-margin service line and grow volume. What does not work is deploying the efficiency and hoping the commercial model absorbs it quietly, because the realisation shows up in the numbers about two quarters later.
Is our client data safe in these tools?
That depends on the deployment terms and it is a procurement question rather than a technical one. What matters is where the data is processed, whether it is retained, whether it is used for training, which jurisdiction it sits in, and what your own engagement letters and your clients' contracts permit. In our experience client contract terms constrain firms sooner than any regulation does, and they are usually reviewed after a tool has been selected rather than before. Reviewing the terms of your largest clients first costs a week and prevents a reversal.
What should a firm build first?
Retrieval over its own work, in almost every case. Professional firms hold decades of resolved problems — prior advice, precedent documents, working papers, prior engagements — in systems only the author can navigate, and the recurring cost is rediscovery. Retrieval with provenance addresses that directly, uses material you already own, and carries far less risk than generation. The design requirements are real: confidentiality boundaries between client teams, and the ability to tell current material from superseded, which is the difference between an asset and a liability.
Start with the problem, not the technology.
Thirty minutes, no charge, no deck. Tell us what is going wrong in your organisation and we will tell you whether AI is the right instrument — including when it plainly is not.