The Agentic Consultant: A Definition and a Standard

Consulting has always run on leverage. One partner’s judgement, multiplied through layers of people who do the labour: the research, the models, the first drafts, the version-seventeen deck at midnight. Every great firm’s economics rested on that multiplication, and so did every consulting career, because the labour was how you learned.

What is changing is where some of the multiplication comes from. AI agents can now perform meaningful parts of that labour: research with sources attached, analysis on real data, drafts that arrive structured, iteration that never tires. AI is becoming one factor in a broader restructuring of how consulting firms recruit, organise and deliver work. McKinsey’s headcount had already fallen from more than 45,000 at the end of 2023 to around 40,000, and in December 2025 the Financial Times reported plans to cut roughly a further ten per cent, concentrated in non-client-facing functions, over about two years. Accenture reduced headcount while reskilling at scale and growing its AI and data workforce to some 77,000 people. PwC has described the firm of the future as a diamond rather than a pyramid, with a smaller junior base and a stronger expert middle. It has also warned that shrinking entry-level recruitment too aggressively could damage the apprenticeship system that produces future experts.

None of this means consulting is ending. It means the traditional bundle of junior tasks is being compressed, faster than the profession’s apprenticeship model can adjust — and that a question lands on every consultant’s desk, employed or independent: what am I, once machines do more of the part I trained hardest for?

I think that question deserves a better answer than either the hype or the doom currently on offer. Here is mine.

The definition

An agentic consultant is a consultant who designs and directs bounded workflows in which AI agents perform parts of the research, analysis, drafting or production, while the consultant retains responsibility for the evidence, the judgement, the client relationship and the final recommendation.

Every word of that is doing work, so let me unpack the two that matter most.

Bounded. Agents are genuinely capable now, and genuinely unreliable in ways that matter to professionals. They can gather material, compare sources, analyse data and iterate against a brief. They can also miss context, invent evidence, and pursue the wrong objective with impressive efficiency. The agentic consultant decides what gets delegated, which sources are acceptable, where review is mandatory, and what is fit to put in front of a client. Delegation without boundaries isn’t leverage; it’s exposure.

Retains responsibility. Four things stay human, and the first three will be familiar — they were always the job.

Judgement. Which question is actually worth answering? What does the evidence support? Which assumptions are carrying too much weight? An agent can generate possibilities; it cannot take professional responsibility for choosing among them.

Relationships. Consulting happens in conversations, conflicts, silences and incentives as much as in documents. Trust is not earned by producing more output. It is earned through consistency and the ability to have difficult conversations well.

Accountability. A client needs to know who stands behind the work. “The system produced it” is not a defence for an invented figure or a recommendation that collapses under scrutiny.

The fourth is new. It is the skill that makes an agentic consultant agentic.

Direction. Agents do not remove the need for management; they create a new form of it. Someone must define the objective, divide the work, set the standards, decide what information each tool may touch, review progress at the right moments, and stop the process when it is heading somewhere wrong. An agentic consultant does not simply use AI. They know how to direct it.

To be clear about what this is not: an agentic consultant is not the same as an agentic-AI implementation consultant — the fast-growing trade of deploying agent systems for enterprise clients. That is a fine business and a different one. This is about how a consultant runs their own work, whatever their subject matter. I’m not claiming to have coined the phrase; versions of it are already in circulation with other meanings. I am proposing the professional standard it should represent.

One footnote on the word itself, because it has a history worth knowing. Psychology has used “agentic” in two opposite senses: Bandura’s, describing people who act on the world deliberately rather than reacting to it, and Milgram’s “agentic state”, describing people who surrender responsibility to an authority. The technology makes both futures available. The definition above, and the standard at the end of this piece, exist to keep this profession firmly in Bandura’s sense.

The evidence, honestly

I’m asking you to rethink how you work, so the sourcing should model the standard.

The best-known field experiment we have comes from Harvard Business School and BCG: 758 consultants, working with 2023-era GPT-4. On tasks within the technology’s frontier, consultants using AI completed 12.2% more tasks, worked 25.1% faster, and produced significantly higher-quality output. On a task outside that frontier, AI users were 19 percentage points less likely to reach the right answer. Both halves matter. AI improved performance where it suited the work and damaged it where people trusted it in the wrong conditions — which is precisely why the definition above says “bounded” and not “everything”. The question is no longer “can AI do this task?” but: under what conditions, with what information, tested how, reviewed where, and accountable to whom.

Since that study, the tools have moved from answering questions to carrying multi-step assignments through checkpoints, with sources attached. The practical advantage of being able to direct that can now be material — but be careful with anyone who tells you it is a settled multiple. It depends on the task, the boundaries, and the quality of the review.

On the supply side: MBO Partners’ 2025 research counts a record 5.6 million Americans earning more than $100,000 independently, up 19% in a year, and finds 74% of independent workers already using generative AI. That describes independents broadly, not consultants specifically — but it tells you the pool this kind of consultant comes from is growing, and already tooled up.

And now the caution, because you should distrust anything in this space that arrives without one. Gartner forecasts that over 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear value or inadequate risk controls. It also estimates that only around 130 of the thousands of vendors using the “agentic” label meet its definition of genuinely agentic. “Agent washing” — relabelling an ordinary chatbot as an agent — is common enough to have earned its own name. Earlier this year the open-source agent platform OpenClaw became a security crisis when, in the platform’s first weeks, researchers reported malicious payloads in roughly 17% of the marketplace skills they analysed.

I include the ugly numbers deliberately. The coming disillusionment will punish overpromising, and it will reward practitioners who can verify their work and stand behind it. That is not a threat to the agentic consultant. It is the opening.

The trust layer

The deepest change is not speed. It is that when drafts are nearly free, the scarce skill is knowing which draft is right. In my own testing, an agent will state an invented number with exactly the same confidence as a sourced one.

So the working discipline of an agentic consultant includes what I call a trust layer. Every material claim should be traceable to a source or explicitly labelled as calculated, estimated or uncertain. Review should be designed into the workflow and placed at the moments when a wrong assumption could distort everything downstream, not bolted on at the end. Every important recommendation should be subjected to a deliberate opposing pass. What contradicts it? What has been omitted? Under what conditions would it fail? Client information should enter only tools approved for that class of data: confidentiality is an architecture decision, and if you work inside a firm, your firm’s AI policy is a hard boundary rather than an obstacle. And nothing should reach a client merely because a workflow marked it complete.

The trust layer is what lets you sign the work. Being able to sign the work is the entire game.

How to move

The shift is a sequence, not a leap, and it does not begin by assembling a collection of agents. It begins with how you already work.

Make your method explicit. Choose one recurring deliverable — a research synthesis, a business case, a proposal. Write down how you actually produce it: the information you need, the questions you ask, the standards you apply, where your judgement enters, what goes wrong most often. You cannot delegate a method that exists only in your head. This is why structured frameworks matter more in the agent era, not less; a framework is how you brief something that cannot read your mind. (The six I developed are free here; they were written for prompting and they transfer directly to briefing agents — a shift I traced in From Prompts to Agents.)

Separate tasks from decisions. Agents can be useful for collecting, structuring, comparing, calculating, drafting and checking. Decisions about relevance, risk, organisational context and recommendation stay visibly human. The boundary moves by engagement; drawing it is part of the job.

Build one bounded workflow. Not an automated engagement — one repeatable stage with clear inputs, a verifiable output, source requirements and failure conditions, tested on non-sensitive work until you know where it is dependable and where it is not. A market-entry research stage, say: the agent returns a sourced landscape and stops; you review — some sources will be thin, one market-size estimate won’t survive its own assumptions, a line of enquiry needs killing; your steer goes back before conclusions are allowed to form. There is friction at every checkpoint. The checkpoints are the job. The goal is not maximum autonomy; it is controlled leverage — because a weak method, automated, produces weak work faster.

Add the trust layer before you scale. A workflow is not finished until you know how its output gets verified. Without verification, each additional automated step can compound an error rather than create leverage.

Then change how you describe the value — and, if it’s yours to set, the price. Clients are not buying an afternoon of document production; they are buying a clearer decision, a defensible recommendation, a risk retired. If you’re independent, compressed delivery time creates a commercial question. Hourly billing can quietly hand much of the value of your improved operating model to the client. Not every engagement suits outcome pricing; a fixed deliverable, retainer or staged fee may fit better. The principle is simpler: time saved should not automatically become value surrendered. An independent consultant can often reconsider the commercial model on the next suitable proposal. Large firms may move more slowly because their utilisation, staffing and pricing structures are harder to change. If you’re employed, you can’t reprice, but you can become the person who improves the method: better briefs, defined review standards, and the role I’d bet on most — the trusted verifier of AI-assisted work. The tools belong to your firm. The discipline is portable, and it leaves the building with you.

For the independent, the destination has a name worth knowing: the firm of one — one accountable human directing research, analysis and production capacity that used to require a team. Don’t take it literally; software is not a workforce, and one person’s knowledge still has limits that collaborators and challengers must fill. What changes is the point at which growing your capacity requires growing your headcount. Solo consultancies are not new. This degree of leverage, at this price, is.

The standard

A definition means nothing without obligations attached. The agentic consultant holds to five:

1. I delegate work, not accountability. I remain responsible for what I use and what I recommend.
2. Every material claim carries a status. Sourced, calculated, estimated or explicitly uncertain — an unlabelled figure is a defect.
3. Human judgement is designed into the workflow. Review happens at defined checkpoints, not only at the end.
4. Client trust governs the technology. Confidentiality, permissions and appropriate tool use take priority over convenience.
5. Method comes before automation. Platforms change monthly; a sound brief, a disciplined process and a defensible standard of evidence do not.

Adding AI to an unchanged workflow can produce a real but limited improvement. Rebuilding selected parts of a practice around method, boundaries, direction and trust is a more fundamental change — one clients can feel. Consultants who make that change, and hold themselves to a standard the hype merchants cannot meet, need a name. Whether you claim it is up to you.

The six frameworks — the briefing discipline everything above rests on — are free here. They are developed further in The Consulting Prompt Playbook, extended into 100 ready-to-run prompts in the companion Prompt Vault for readers, and inform the Consulting AI Operating System now in development.

Three questions people ask

Is an agentic consultant the same as an AI consultant? No. An AI consultant advises clients about AI. An agentic consultant may advise on anything — strategy, operations, change — and uses directed AI agents in how the work gets produced, under human review and accountability.

Do I need to be independent to be one? No. The method, the boundaries and the trust layer work inside a firm on approved tools. Independence changes what you can price, not what you can practise.

Doesn’t this just mean replacing junior consultants? It means the traditional bundle of junior tasks is being compressed, which is a fact about the technology, not the point of the term. The unresolved question — how the next generation develops judgement when machines do more of the apprenticeship work — is one the whole profession, including this site, has to answer honestly.