From Prompts to Agents: How Consultants Should Actually Work With AI Now

Most people formed their opinion of what AI can do sometime in 2023 or 2024, typing questions into a chat box and reading the answers. If that is still your mental model, it is now about two product generations out of date — and the gap is quietly costing firms a great deal.

I want to give you an honest map of where things actually stand in 2026, why the shift is bigger than it looks from inside a chat window, and what it means for how a consultant should work. I will use the six frameworks from my book as the backbone, because the underlying discipline has not changed. But almost everything around that discipline has.

What actually changed: from answering to acting

The single most important development is easy to state and easy to underestimate. AI stopped being something that answers and became something that acts.

Two years ago, the state of the art was a model that wrote you a paragraph. You prompted, it replied, you copied the useful bits into your own work. It was a very well-read assistant trapped behind a piece of glass — it could tell you what to do, but it couldn’t reach out and do anything.

That wall has come down, through three developments that arrived in quick succession.

The first is connection. In late 2024, Anthropic introduced the Model Context Protocol — a common standard for letting AI systems plug into real tools and data. It sounds like plumbing, and it is, but plumbing is what turns a demo into infrastructure. Within a year it went from one company’s idea to the industry’s default: OpenAI, Google and Microsoft all adopted it, and in December 2025 it was handed to a Linux Foundation body to steward as a neutral standard. By early 2026 there were more than ten thousand public connectors and tens of millions of monthly downloads. In practical terms, your AI can now see your calendar, your documents, your CRM, your analytics and your codebase — not by pasting things in, but through a live connection.

The second is agency. Once a model can connect to tools, the obvious next step is to let it use them in sequence, on its own, toward a goal. That is what an “agent” is — not a science-fiction robot, but a model that can plan a multi-step task, take the steps, check its own work and keep going. Two years ago this was a research demo for engineers; by early 2026 it had shipped to ordinary knowledge workers under names like Claude’s Cowork and ChatGPT’s Workspace Agents — systems that plan and run long tasks, produce the documents and decks and analyses at the end, and keep working while you’re asleep or in a meeting. The names will keep changing and the launches will keep coming; the capability is the thing to hold onto.

The third is specialisation. Alongside the agents came a quieter idea — the industry has settled on the word “skills”: reusable packets of expertise — a defined workflow, the context it needs, the standard it should hit — that turn a general-purpose model into a specialist for a particular job, and can be shared across a team. Instead of re-explaining how your firm builds a market assessment every single time, you capture it once and hand it to the agent as a repeatable capability.

Put the three together and the picture is genuinely different. The thing on the other side of the glass is no longer a witty intern who can only talk. It is a junior team that can connect to your systems, carry out a plan, and specialise in your way of working. That is a categorical change, not an incremental one.

The paradox nobody is talking about enough

Here is where it gets interesting, and where the real opportunity for consultants sits.

By every adoption measure, this technology has crossed the chasm. In professional services specifically, Thomson Reuters’ 2026 research found that around 40% of professionals say their organisation now uses generative AI — nearly double the year before. Most weekly users expect it to be central to how they work within five years. On paper, the battle for adoption is over.

And yet almost nobody can tell you what they are getting for it. In the same body of research, only about 18% of firms track the return on their AI investment at all; a further 40% don’t even know whether anyone is measuring it. On the agent front, McKinsey’s 2026 work puts the share of enterprise functions actually using AI agents at roughly one in ten, despite the enormous noise. Gartner has gone further and forecast that more than 40% of agentic-AI projects will be scrapped before the end of 2027 — not because the technology fails, but because of unclear value and weak controls. Security and risk, not capability, is the barrier firms name most often.

Sit with that combination for a moment, because it is unusual. Nearly everyone has adopted the tools. Almost no one can prove they are working. That is not a technology gap — the technology plainly works. It is an operating gap.

This is the most important thing I can tell a consultant in 2026: the tools have been commoditised, and the discipline has not. Your client has the same models you do. Your competitor down the corridor has the same connectors and the same agents. Access is no longer the edge. The edge is entirely in how well you brief the work and how well you govern it — and that is a set of skills, not a subscription.

Why structure matters more now, not less

When AI could only write a paragraph, a lazy prompt was cheap. You typed a vague question, got a bland answer, shrugged, and rewrote it yourself. The cost of a poor brief was thirty wasted seconds.

That arithmetic has completely changed. When you hand a task to an agent, a vague brief no longer wastes thirty seconds — it sets an autonomous process running for minutes or hours, across your real tools, in the wrong direction. The model will confidently pursue whatever you actually asked for, not what you meant. The brief is no longer a question; it is a specification for work that will be carried out while you are not watching.

This is why I am more convinced than ever that the humble skill of briefing well is the highest-leverage thing a knowledge worker can develop. It was valuable when the downside was a weak paragraph. It is decisive now that the downside is a confidently wrong deliverable, built on your live data, that you then have to catch.

And catching it is the other half. The research is blunt about this: the thing holding agentic AI back is not intelligence, it is trust — security, risk, and the absence of a way to know the output is sound. Which is precisely the muscle a good consultant already has. We are trained to pressure-test, to ask what would have to be true, to find the assumption that breaks the recommendation. That instinct, which felt like table stakes in human work, turns out to be the scarce and valuable layer in AI work.

So the consultant’s job in the agent era has two clear parts, and the frameworks map onto both: brief the work with enough structure that an autonomous system can execute it well, and govern the output with enough judgment that you would put your name on it. Everything else is detail.

The frameworks, in the agent era

The six frameworks in The Consulting Prompt Playbook were written for prompting. Their value has, if anything, gone up — because a framework is just a repeatable shape for a brief, and briefs now matter more. What has changed is that you are increasingly briefing an agent that will act, not a chatbot that will reply. The shape holds; the stakes are higher.

Two of them do most of the daily lifting.

GOAL — Goal, Output, Audience & Context, Leverage — is the fifteen-second structure for ordinary, fast work. What do you want to achieve; what format do you need; who is it for; what constraints or style apply. It was always the daily driver. In an agent world it is the minimum viable brief: even a quick delegated task needs those four answers, or the agent fills the gaps for you, usually blandly.

RCAS — Role, Context, Ask, Style — is for work that has to leave the building and survive scrutiny. Give the model a role to reason from, the context it is missing, a precise ask, and a clear standard for what good looks like. This is the shape I reach for when an agent is going to produce something client-facing, because it front-loads exactly the judgment an autonomous process cannot supply for itself.

The other four each cover a mode of work that has only become more relevant: RISE for genuine strategy, when you want the system to explore real options rather than the obvious one; CRISP-C for communication, the board narrative or the difficult message; CAPE for research and analysis, so an investigation stays focused instead of sprawling — and sprawl is exactly what an unsupervised agent does worst; and IDEA for innovation, pushing past best practice into next practice.

You do not need to memorise them to take the point. The point is that “just wing the prompt” was a forgivable habit when the machine only talked. It is an expensive habit now that the machine acts. A framework is simply how you make a good brief the path of least resistance instead of an act of willpower at 11pm.

If you want the ground-level mechanics, I’ve written separately on why structure beats one-liners and on using GOAL to brief in fifteen seconds. This piece is the map; those are the technique.

The judgment layer is the job now

There is a temptation, reading all this, to conclude that the goal is to automate yourself — to build the agents, connect the tools, and step back. That is exactly the mistake the cancelled-project statistics are made of.

The firms getting real value are not the ones delegating the most; they are the ones delegating the right things and keeping a firm hand on the rest. And knowing which is which is a consulting skill, not a technical one.

A useful way to think about it: sort the work by two questions — how reversible is a mistake, and how much does it depend on judgment the model cannot have? Low-stakes, well-defined, easily-checked work — first-draft research, formatting, reconciliation, summarising a data room — is exactly what you should be handing to agents, and you are leaving value on the table if you don’t. High-stakes, judgment-heavy, hard-to-reverse work — the recommendation itself, the framing of the problem, the number you will defend in a board room — is where you stay firmly in the loop, using the AI to widen your options and stress-test your thinking, never to make the call.

Most of the disappointment with AI comes from getting this sorting wrong in both directions at once: firms automate the judgment (and get confidently wrong answers they can’t defend) while doing the drudgery by hand (and wonder why there’s no productivity gain). Get the sorting right and the same tools that are failing elsewhere start compounding.

This is also, incidentally, the honest answer to the anxiety about consultants being automated away. The parts of the work that an agent can do end-to-end were never where your value lived. The parts that need someone to decide what matters, to carry the risk, to look a client in the eye and say “this is what I would do” — those are becoming more valuable precisely because the surrounding work is getting cheaper. Augmented judgment beats both unaided human effort and unsupervised automation. That is the whole game.

What this means for your clients

There’s a market signal worth ending on, because it changes the commercial calculus. Clients have noticed all of this too. The Thomson Reuters research found that roughly two-thirds of corporate clients now actively want their outside firms to use AI — while fewer than one in five formally require it. Read those two numbers together and you get a clear message: the expectation has arrived, but the standard has not been set.

That is an opening. In a market where every firm has the same tools, “we use AI” is not a differentiator — it is table stakes that clients increasingly assume. What clients cannot yet get, and will pay for, is disciplined AI: work that is faster and cheaper because it is intelligently delegated, and trustworthy because it is properly governed. The firm that can show its working — here is what we automated, here is where a human made the call, here is how we checked it — wins the trust that the cowboys, racing to automate everything, are busy destroying.

That is a discipline you can build deliberately. It starts, unglamorously, with the brief.

Where to start

If you take one thing from this, let it be the reframe: the job is no longer to use AI — nearly everyone does that now — it is to direct it well and govern it honestly. Tools are commoditised; operating discipline is not.

The practical on-ramp is small. Before your next real delegation to an AI, spend fifteen seconds on the four questions underneath every good brief: who do I want it to be, what does it not yet know that I do, what exactly am I asking for, and what does good look like. Then, before the output goes anywhere, spend two minutes as the sceptic: what would have to be true for this to be right, and where would I look for the crack. Brief with structure; govern with judgment. That is 80% of the value, available today, with the tools you already have.

When you want the full system, I’ve made the six frameworks free — you can get all six here. And the hundred ready-to-run prompts built on them, each tagged with its framework and adaptation notes, come with the book as the companion Prompt Vault for readers.

The technology will keep moving; by the time you read this, there will be a newer model and a louder launch. The discipline underneath won’t move at all. The consultants who thrive in this decade won’t be the ones who adopted AI earliest or automated the most. They’ll be the ones who learned to brief it like a professional and govern it like one — and who remembered that in a world where everyone can generate an answer, the rare and valuable thing is judgment about which answer is right.