What happens when AI joins the project team?

DO
Start on the nearly black screen. Click once for the question and let it sit for two seconds. Click again for your name before you move into the setup.

That is the question. Not whether AI is coming. It is already in the building.

What I want to work out with you is what happens to an engagement when there are four of them working alongside you. Who does what. Who decides. Who signs off.

I am going to show you my actual working setup, not a product tour. Some of it you can use this afternoon. Some of it needs your technology organisation. I will be clear about which is which.

Thirty three minutes. Then ten for the argument.

DO
Say the first line and stop. Then click for the last block and slow right down.

Let us put the real tension on the table right away, because everything after this depends on it.

Your advice is not what is under pressure. Your judgement is fine, and your client relationships are fine.

What is under pressure is how the advice gets produced. The reading, the synthesis, building the pack. Those three weeks between the kick-off and the first real conversation.

And somebody out there is already quoting your client a shorter version of the same pre-study. Same rigour. Different price.

(Click.)

Take this back to your leadership. The firms that learn to deliver this way first are the ones who decide what a pre-study is worth. Everybody else ends up negotiating against a number that somebody else picked.

That is the whole reason I am standing here.

DO
Do not explain tokens. Land the middle two numbers, they are about your audience's job.

So why now, and not three years ago?

Two things moved. The cost of running machine reasoning fell by a factor of two hundred and eighty in two years. You do not need the technical detail behind that. It means work that was once too expensive to touch became work nobody thinks twice about.

The other two numbers matter more to you, because they were measured on your job.

Harvard and BCG took seven hundred and fifty-eight consultants and randomly assigned them to work with or without a frontier model on eighteen realistic consulting tasks. Over twenty-five percent faster. Over forty percent better, scored by human graders.

Be honest about the timeline here. Both numbers are older than they look. The price figure stops at the end of 2024 and has kept falling since. The study ran in 2023 on models two generations behind today, so treat the exact percentages loosely. The direction has not reversed.

DO
Title lands. Pause. Click for one. Let it sit. Click for two. Click for three, then wait properly.

These are not hypothetical. Your clients sit in the same conferences you do, and some of them are further along than the firm advising them.

(Click.)

What did you automate on this engagement, and what did that save me? Procurement will ask that one in writing.

(Click.)

Who checked the output before it reached my board? That comes from a risk committee, and it is a very good question.

(Click.)

Then this. Your firm got faster this year. Why is my fee the same?

(Pause. Look at the room.)

Three years ago none of these came up. Somebody in this room has already been asked one of them across a table. I would rather you had an answer ready than improvise it.

DO
Click through the five phases. Do not read every cell. Click once more for the bottom line, land it and stop.

People hear AI and picture one task, a summary or a draft email. That is too narrow. It runs across the whole assignment.

Prepare: it reads the context and shows what matters for you to decide.

Interview: it transcribes, highlights contradictions and keeps you focused while you are still in the room.

Analyse: it preserves the evidence and sources so you can judge what is good enough.

Recommend: it drafts options and trade-offs. You make the call.

Present: it prepares questions and follow-up. You take the pushback.

The split stays simple. The volume goes. The judgement stays, and the recommendation keeps your name on it.

DO
Click through the four rows. Read the task first, then the right-hand result. Keep the pace up.

So what changes in a pre-study?

Fourteen interview transcripts. Today, four get read properly. With agents, all fourteen get themes and contradictions flagged.

Market benchmarks. Today, whatever you can dig up in two days. With agents, you get a sourced set, the method checked, and weak references dropped.

The steering pack is usually assembled by hand the night before. With agents, it starts from the evidence. You still edit the argument.

Between engagements, the evidence base keeps running instead of going quiet until the next tender.

Same six weeks. Different amount of work inside them.

DO
Read the left panel first. Then the commercial panel. Pause on capabilities compounding.

The delivery model is already changing the business model.

If a pre-study takes half the calendar time, you cannot keep pricing it the same way and expect the model to hold.

The delivery team changes first. A smaller senior team directs a fleet of agents instead of a pyramid doing the assembly work.

Then the economics change. Compressed timelines force you to price the outcome rather than the hours. Prompts, methods and agents improve between deliveries. That is IP. It compounds.

The relationship does not have to stop when the work is delivered. A retainer keeps the evidence, benchmarks and advisory work moving.

This is not a blueprint every firm should copy. But the old model cannot remain the default once the delivery engine changes. Decide how you will capture that compounding IP: retainer, outcome pricing or licensing.

DO
Make this personal. First three fast. Slow right down on the last one.

This is the personal part. What happens to your career if the work changes and you do not?

Your clients will expect it. They are running their own pilots, and soon they will ask what you changed in the delivery.

Someone is learning faster. Maybe not a competitor you watch. A smaller firm with less to unlearn.

Your senior time gets more valuable. Less of your week goes on work that never needed your judgement.

And then the one that reaches your own career.

Reading, drafting and being corrected built judgement. If the machine takes that work, you need a new way to learn it.

I do not have a finished answer. Nobody I work with has a proven one. But the choice is becoming clear. Start now, while you can learn the new way of working with support, or wait until it is simply expected of you.

Your title may stay the same for a while. The work underneath it will not.

With this many AI products changing every month, where should a consultant even begin?

DO
Say the first line, then let the tool names sit for a beat. Explain the jobs, not every product. Click once per job lane. Then click for the proof overlay, Chat and Cowork. Keep this bridge short.

There is a ladder here, from chat to agents that keep working. But it is not a product ladder. The tools overlap. Start with the job.

Microsoft 365 Copilot is the familiar door. Ask for help in Word, Teams or Outlook, and keep moving.

Cowork is for a bounded outcome. Hand over the goal, review what comes back, and approve anything sensitive.

Scout is for work that should continue. It can keep the engagement context, watch for change and delegate to specialist agents.

Copilot CLI is where I build my own toolbox. Claude Code and Codex are well-known examples of the same build-oriented end of the spectrum. You can create skills, connect tools and shape a repeatable workflow without starting from a blank chat.

The question is not which product wins. How much do you want to configure, control and leave running?

(Click.)

I need help now. Microsoft 365 Copilot.

(Click.)

Hand off the outcome. Cowork.

(Click.)

Keep the work moving. Scout. Persistence is the difference, not intelligence.

(Click.)

Build my own toolbox. Copilot CLI, Claude Code or Codex. The job is to make the toolbox itself.

(Click.)

This is what happened when I kept going.

These are examples from my own role: Partner Cockpit, Opportunity Hunter, Matchmaking, Signal and AI Chief of Staff. Different jobs, same principle. Start with a real problem, not a software roadmap.

(Click.)

Before the case, look at the two tabs.

Chat is familiar. I stay in the conversation and decide the next move.

(Click.)

Cowork changes the hand-off. I give it a bounded task, then review what comes back. The plan, files, skills and tools stay visible.

Cowork could carry much of the case I am about to show. Scout is for a project that should persist, with a heartbeat, specialist agents and a stop when the evidence needs my judgement.

Now let us put the choice into a real engagement.

DO
Read the client question out loud. Then say the honesty line about the demo before you switch screens. None of that is on the slide.

Those four modes were about your own week. Now put one inside a real engagement, the kind every one of you has actually run.

Nordic industrial group. Four service organisations. The board wants to know whether they should be one. Fourteen interviews, forty one documents, six weeks, and a steering group on the eighteenth of September that will not move.

It is invented. Nothing about it is unfamiliar.

One thing before I switch screens. I want to be straight with you, and this is not on the slide.

What matters here is why it can start in the right place. The project folder carries the context: the brief, the source material, the method, and the instruction files. Custom agents add the roles and rules for the work. Copilot is not starting from a blank chat. It can see the purpose of the engagement and the boundaries you gave it.

DO
Space steps the demo. Fifteen states. Backspace steps back. R resets.
The panel below shows the line for each state. Both decisions are spoken while the waveform moves, then appear in the composer on the next step.

(Per-state script is in the panel. Advance with Space.)

DO
Make the product boundary explicit. The demo was personal Scout, on my permissions, for one task. The right-hand side is the operating-model direction, not a claim about today.

Everything you just watched was personal. Scout worked for me, on my machine and under my permissions.

The next jump is from personal Autopilot to project agent.

A governed agent in the project channel can serve the partner, engagement manager and workstream leads from the same place. It might own the morning evidence check, keep the decision log current or watch the client actions everybody forgot after Tuesday's meeting.

This pattern is already emerging in Teams and Slack. People use the agent, then correct it inside the flow of work. They improve its instructions, add examples and tighten the checks. Over time, it becomes part of how the team operates.

One useful middle step was released last week. GitHub Copilot can now start a collaborative cloud-agent session from a Teams thread, where the team can see and steer the work while the issue or pull request stays connected to the discussion. I have already picked that pattern up in one of my own Teams workflows, using GitHub for the collaboration and review trail.

But be precise about what you saw. That was not Scout serving a team. Scout is personal today. A shared project role needs a separately published and governed agent.

DO
Point back to the conflict Scout could not resolve on its own. Keep this practical and personal.

Scout stopped a minute ago because leadership, operations and market evidence pointed one way and challenge pointed another. That pause is not a limitation to apologise for. It is the point.

If the machine carries the volume, what exactly are you being paid for?

Three things. These are the expensive ones.

Context. What this client can actually absorb, and what the board will refuse to hear this quarter. No model in the room knows that.

Quality. Which evidence is strong enough to carry the firm's name. That judgement is about risk. A statement can be technically accurate and still be unsafe to use.

And the call. You own the recommendation itself, together with the risk of being wrong in front of a steering group with the CFO in it.

That is why the stop in the demo matters. A confident answer is cheap. Knowing when the evidence needs a human call is the valuable part.

DO
Move clockwise. Keep this about consulting IP, not model training. The correction you just chose on the conflict is the raw material for this loop.

Consulting firms have always had IP: method decks, interview guides, benchmark databases, quality checklists. Useful, but passive. Somebody still has to remember where those assets are and apply them properly.

Agents change the form of that IP.

Project work produces evidence, decisions and outcomes. The choice you just watched, one operating model with two exceptions, becomes a correction. Those corrections become better instructions, stronger examples and tighter checks. The next engagement starts from a reviewed capability instead of an empty chat window.

The model is not secretly learning from client data. The firm is deliberately improving an asset it controls.

That makes IP more important. When everybody can generate another app or another analysis, the advantage is the delivery system that improves every time your people review it.

That is how AI-driven delivery becomes a competitive advantage. The next team does not start from scratch. It starts from what the firm already learned.

DO
Let the scattered labels sit for a beat before the click. On the click, watch the room, not the slide. Land the line underneath and stop.

Zoom out from one engagement to the whole firm.

Models here. Agents there. Data somewhere else. A workflow tool nobody fully owns. Human decisions scattered across email threads. Governance added afterwards. Most firms are already living here, whether they planned it or not.

(Click.)

The same six things now need to work together.

More agents create an estate. A pile of capable tools is not an operating model. Orchestration is what makes the estate work as one system, not six.

DO
Read layer by layer, top to bottom on screen. Do not read every product name in the connect row. Land the closing line and stop, this slide has to stay under 75 seconds.

Somebody in this room will ask me after the session how this fits together at KPMG scale. Here is the honest map.

At the top, the surfaces people actually work in. Microsoft 365 Copilot, Cowork, Scout, Teams, your own business applications.

Underneath, the places you build and orchestrate multi-step work. Copilot Studio, Power Automate, Foundry workflows, and the Microsoft Agent Framework for code-first teams.

Then a thin connecting layer, mostly invisible to you. MCP connects tools and data. Agent-to-agent messaging handles coordination. The Activity protocol reaches Microsoft 365 channels.

And at the base, one thing, not four. Agent 365. It does not orchestrate your recommendation. It watches the whole estate: identity, access, visibility, protection, lifecycle, and who is accountable when an agent gets something wrong.

Several ways to build. Standard ways to connect. One control plane, not one engine pretending to be everything.

One honest caveat, because somebody will check. Foundry workflows are still in preview, and Microsoft's own documentation already points new work toward the Agent Framework.

DO
Say public twice. Do not speculate about the Workbench roadmap. Point at each tag as you say the word, quietly.

I am a guest here, so every line on this slide is public.

Workbench gives KPMG a multi-agent platform, the piece you build on. Copilot is deployed across more than two hundred and seventy-six thousand people, the surface people work in. Agent 365 is being adopted to govern the agents themselves and watch the estate.

Work surface. Platform. Governance. The same parts from the map I just showed you. This time they are already funded, live and named in public.

The question for this room is what you build on top of them. Which roles will agents take? Which methods will they run? Which services become possible?

And which parts of KPMG's expertise improve every time the system runs?

DO
Ask the question. Pause. Then click for the Teams notification and leave it there.

COPY
https://www.linkedin.com/in/johanwallquist

One question before I take yours.

Pick one role or workflow to test next week. Not a strategy, one recurring task with a clear owner and a clear boundary.

Then ask what it should learn each time the team corrects it.

Your IP used to sit in slides. Now it can run.

(Click.)

This is a published project agent, not personal Scout, reporting inside the team's own channel: what changed, what it did, and what still needs a human.

That notification is the direction, not a claim that personal Scout already serves a whole project team.

Thank you.