How to make a consulting firm AI-native: a playbook

In October 2025, an A$440,000 report Deloitte wrote for Australia's Department of Employment and Workplace Relations was found to cite academic sources that do not exist and to carry a made-up quote from a federal court judgment. Deloitte issued a revised report and agreed a partial refund[37]. A month later, a CA$1.6 million, 526-page health workforce plan Deloitte wrote for Newfoundland and Labrador turned out to cite at least four research papers that cannot be found. Two of the researchers it named said the papers attributed to them do not exist[36].

That is the first problem: what the firm signs. When an agent drafts the research, every claim and citation still goes out under the firm's name.

The second problem is in Accenture's annual report. It warns that unless new work makes up for it, AI-driven automation "will lead to reduced demand for our services and/or adversely affect the utilization rate of our professionals", and that results suffer "if our clients do not accept new pricing or commercial models that reflect the value of these AI-enabled solutions"[4]. That is how the firm charges. In most sectors AI cuts cost. In a firm that bills by the hour, it also cuts revenue on the same work.

An AI-native firm handles both. Agents draft; partners sign every price, claim and deliverable, behind a reference check no deliverable can skip; and repeatable work moves off hourly pricing.

An AI-enabled firm gives its people AI tools. An AI-native firm redesigns the engagement so agents do defined steps, under the firm's rules and the client's terms, and people decide what carries the firm's name.

Most firms are on the left. No survey covers consultants directly. In a vendor survey of the neighboring professions (legal, tax, accounting and risk), 34% said they use AI tools their employer has not approved, and while 78% of corporate clients said AI-enabled quality is essential from their providers, 6% said providers deliver it[25]. Assume your Stage 0 inventory will find tools nobody approved.

Firms will not replace their PSA, CRM or document store to become AI-native. The agents are arriving inside them. Certinia lists agents for RFP responses, estimation, SOW drafting, resourcing, shortlisting, work reallocation and project management[27]. Kantata's agent does skills matching and forecasting[28]. Copilot searches every file a user can view[26].

Each tool can govern its own agent. None answers the questions across all of them: which engagement is this agent on, whose authority is it using, did the client allow this model to see this data, and who approved what went out.

OrchKernel is built to be that layer. Agents ask it before they act. It checks the request against the firm's rules, holds it for a named person when needed, runs it with credentials the agent never sees, and writes it to a tamper-evident log. What it is not: it is not a PSA, a conflicts system or a document store. The PSA stays the system of record for projects, time and billing, and the conflicts system still decides who is conflicted. The detail is in the OrchKernel blueprint.

CRA International puts the model in one sentence: "Our revenues and profitability are largely based on the bill rates charged to our clients, compensation costs and the utilization of our consultants"[6]. Even fixed fees are usually priced as hours by role times the rate card, so margin depends on whether actual hours stay under the estimate.

The public numbers are thin, because the large strategy firms are private. What filings show: FTI Consulting billed an average of $529 an hour in Corporate Finance, $442 in Forensic and Litigation Consulting and $583 in Economic Consulting in 2025[5]. Gross margin, revenue less direct cost, was 32.1% at FTI in 2025 and 32.0% at Accenture in the fourth quarter of fiscal 2026. Operating margin was 10.3% at FTI, 10.5% at Huron and 11.1% at CRA, against 15.4% at Accenture[3,8].

Utilization, billed hours over available hours, is where AI first changes the economics. The firms that report it billed between 57% and 78% of available consultant hours in 2025:

The unbilled 22% to 43% is business development, proposals, bench time, training and admin. It is the firm's own cost, so an agent that shortens it adds margin without touching revenue. That is why the staged path starts there.

The same agent that saves an unbilled hour on a proposal removes a billed hour on a deliverable. Under time-and-materials billing that hour is revenue the firm no longer earns. The margin case works only if the firm moves pricing off hours, fills the freed hours with more billable work, or both. Accenture names new "pricing or commercial models" as the condition for AI to help rather than hurt[4]. Pricing when the hours shrink covers what that means for a mid-sized firm.

Advisory is still growing. US employment in management, scientific and technical consulting was 1,888,100 in September 2026, up 1.5% on a year earlier, while computer systems design fell 1.3%[1]. The BLS projects 10% growth in management analyst jobs from 2025 to 2035, and 14% of the 1,077,100 analysts were self-employed[2].

At one public firm, revenue and headcount have already moved apart. FTI's billable headcount fell 3.2% in 2025, from 6,633 to 6,421, while revenue rose 2.4% to $3.79 billion[5,8]. Federal-heavy consultancies shrank for a different reason, government buying: Booz Allen's revenue fell 6.4% in its year to March 2026 and ICF's 7.3% in 2025[8].

We list categories, with vendors named in their own sources as examples, not recommendations. Their performance figures are their own.

No independent survey of AI use inside consulting firms exists that we could find. Across the US workforce, 23% of employed adults had used generative AI for work in the previous week and 9% every workday, and time savings came to about 1.4% of all work hours[24]. The large firms report far more, with AI teams of hundreds behind them. BCG expected work on AI to supply 20% of its revenue in 2024 and 40% by 2026[42]. PwC became the largest customer of ChatGPT Enterprise in May 2024[40]. Accenture bought the UK AI firm Faculty in January 2026, at a reported $1 billion valuation[41], and IBM bought the data consultancy Hakkoda in April 2025[35]. None of that is a benchmark for a 200-person firm with no AI team.

Capital is going to two models. The first buys traditional firms and runs them with AI. Thrive Holdings raised $2 billion at a $12 billion valuation in August 2026 to do this; it owns more than 50 accounting firms with over 2,000 professionals and about 20 IT firms, and says a tax agent handled "over 7,000 returns at 98% accuracy" (company claim, reported by the press)[33]. OpenAI took a stake in December 2025 and embeds its staff in the portfolio companies[34].

The second builds engineering-heavy firms that deliver working systems rather than reports. Distyl describes itself as an AI-native operations consulting and deployment firm with "50+ Fortune 500 enterprise deployments" (vendor claim)[32]. Both sell outcomes more than hours. Neither is proven at the scale of a large consultancy yet.

Six stages, in the order of the engagement and of the risk. The firm's own unbilled work comes first, because nothing reaches a client without a person and the savings are margin. Client deliverables follow, behind a reference check. Pricing changes alongside, before delivery savings leak. Agents that act across the firm's systems come last. Practices can sit at different stages: a litigation practice may keep Stage 2 tight long after the operations practice moves on.

Stage 0, two Stage 1 workflows on real work, then the reference check in one practice. We suggest proposal drafts and weekly status reports: both go after unbilled hours, neither reaches a client without a person, and both give numbers you can compare within the quarter.

At day 90, ask two questions. Have people stopped rewriting the drafts? Could you show a client who checked the sources of any deliverable, and who released it?

Pricing decides whether AI adds margin or only cuts revenue, and it is a partner decision. The arithmetic shows why the order of the stages matters.

Fixed fee per deliverable. Works for repeatable work with a clear output: a market scan, a diligence module, a benchmarking pack. It needs your own data on actual hours, which Stage 2 produces, and a scope agent watching for work outside the SOW, because the overrun is now yours, as FTI's 10-K warns[5].

Subscription or managed service. Recurring work, such as monthly reporting, ongoing pricing analytics or a run-the-process service, sold as a running service. Managed services overtook consulting at Accenture in fiscal 2026, $37.30 billion against $36.88 billion[3].

Outcome-based fees. Paid on a measured result, for firms with the data to price the risk. We found no public figure for how common they are.

Hourly, honestly. Some work stays hourly: testimony, open-ended advice, early-stage strategy. Bill the time actually worked, and say how AI was used if the client asks. Writing eight hours of narrative for twenty minutes of agent output is Scenario E in when it goes wrong, and it is what a client's invoice audit looks for.

The path at Stage 4. Agents read, draft and check; people decide anything that commits the firm, reaches the client or affects a colleague's career.

An agent can prepare each of these: gather the facts, draft the document, propose the answer. A named person makes the decision, and the log shows whose it was.

The failures that have cost consulting firms money and reputation so far were invented citations and broken walls. Each case names the control points that would have stopped it.

These are scenarios, not reported events. Each is something an agent with too much reach could do in a consulting firm, and each maps to one of the control points.

Consulting is lightly regulated as a profession. Most obligations arrive through client contracts, through public buyers and the client's own regulators, and through general law on data, employment and copyright. Public buyers moved first. Law firms and audit firms share most of this workflow but carry extra professional rules, noted once below.

Across these rules and contracts, four duties keep coming back. The exact form differs by jurisdiction and use, but a firm that builds these four into how agents work has the base the others extend:

Hiring your own consultants. If you screen applicants with AI, the hiring rules in New York City, Illinois, Colorado, California and the EU apply to you as an employer. The staffing playbook covers them in detail.

Our reading of where the work moves, not a survey finding. In a small firm several will be parts of existing jobs.

OrchKernel sits between AI agents and the systems the firm connects, as described in the missing layer. Agents ask it before they act; it checks the rules, holds what needs a person, runs what is allowed with credentials the agent never holds, and records all of it, engagement by engagement.

Where a consulting firm needs a control whatever tools it runs, who owns it, and what enforces it.

OrchKernel is source-available under the Business Source License 1.1 and runs on your own servers, so you can read the code that enforces these controls. Running it for your own firm is one deployment for one organization, which the license allows, and so is a separate deployment for each client. One deployment serving several organizations as a hosted service needs a commercial license (see the license page). Whether one firm deployment whose agents work inside many client environments fits the standard license is to be confirmed with us.

Record the baseline before Stage 1 and track the same numbers at each stage. For most consulting operating metrics, such as realization, write-offs, win rate and proposal cost, there is no public benchmark: the standard professional services benchmark is paywalled, and we have not quoted any figure we could not source.

Sources were read in October 2026; dates are publication or data dates. Last reviewed October 2026.