If you are a CEO or CXO, here is the short version: AI tools plus AI training will not transform your company. Transformation means redesigning the organization so people and AI agents share the work, and that redesign starts with you, not with IT.
Almost every leader I talk to in Dubai, Abu Dhabi or Riyadh has the same story. The licences are bought. A pilot ran. People were trained. Then the work carried on much as before, and the board asked where the return went. In my experience that is rarely a model problem. It is an organization problem, and it is the one thing a CEO can fix and nobody else can.
MIT Sloan's Paul Cheek, who advises executive teams on exactly this, puts it in one line: tools plus training is not transformation.1 What follows is what that means for you, mixed with what we see at Skcript Applied Labs when companies try to move past the pilot.
Why your AI rollout is not transforming anything
You bought the tools. You ran the training. Productivity per person went up a bit. Cheek sees that as roughly the 10% lift chatbots give, a stepping stone and nothing more.
The trap is below the individual. Willy Walker, CEO of Walker & Dunlop, described it from his own firm.1 Each banker has a small dedicated team of one to four people. Make that team faster with AI, and nothing changes if the banker still originates the same number of deals a year. The saved time stays trapped inside the team structure. To capture it, the firm has to move to pooled teams and put the freed capacity behind wider origination.
Check your own company for the same pattern. If efficiency gains stay local, you will not see them in revenue or margin, however good the tools are.
Redraw the org chart with agents on it
Cheek reframes "organization" as a group of nodes: some human, some AI agents. For every agent you deploy, someone on your team must be able to answer four things:
- Who does it report to?
- Who is accountable for its output?
- Whose budget pays for the inference it uses?
- What data can it touch?
If your org chart cannot answer those, you have agents running without a manager. Redesign here does not mean removing people. It means rethinking roles, workflows and skills so human and AI nodes work toward the same goal. His book title, No One Works Here, points at the shift from doing the work to setting intent and goals for systems that do it.
Your personal time is the budget that matters
Cheek will not start a transformation program until he has sat down with the CEO and heard a clear yes. He asks the CEO to commit personal time to the same upskilling journey as the team.
The logic is practical. An employee cannot easily justify an hour of AI learning when a deadline is due. They can when you are doing it, the leadership team is doing it, and their manager is doing it.
Your board matters too. Cheek says AI adds a new kind of actor to the company, one that does not behave like anything the board has managed, and that adds risk. The interview cites figures1 from his writing that about 78% of boards do not really understand AI and roughly 92% of C-suite executives overstate their own understanding or use of it. I could not verify those numbers, so treat them as his claims. The remedy he describes is simple: entire executive teams and boards are blocking two or three days for focused AI literacy.
Two cheap habits help:
- Tinker time. Take one day and use AI on something personal and low stakes. You learn where it fails before you try it on a decision that matters.
- Weekly show and tell. Teams present what they built with AI, working or not, and the best go to the all-hands. It makes learning visible and changes habits, which course completions never do.
Four questions to pick where AI goes
The question Cheek hears most from executives is "what do we do with AI?" There are a million options, so he asks four questions of each:
- Feasibility: can AI do this today?
- Business impact: will it move the needle?
- Risk: what happens if it is wrong?
- Human or AI: should AI do this, or should a person?
Most teams skip the fourth. Walker's mortgage example shows why. A single-family loan is a small bet pooled with thousands, so it suits algorithms. A $200 million commercial loan is bespoke, and the human judgment is part of what the client pays for. Same industry, opposite answers.
Cheek draws the line the same way personally. His agents draft follow-ups after every meeting, but they are not allowed to send anything. AI could send it. He decides it should not.
Run the four questions across your functions and you get a ranked list. That list is your roadmap. For a worked version on a single task, see our pilot to production playbook.
Do not pay frontier prices for every task
Match the model to the job. R&D may need a frontier model; finance or HR may be fine on a cheaper open one. Walker said his firm uses frontier models today but is weighing open models on its own servers because of data security. Cheek expects token costs to keep falling and a mix of models to be the norm inside one firm.
Then ask the cheaper question first: do you need AI at all? Many problems being handed to the newest generative or agentic AI can be solved by an earlier generation or plain automation, at lower cost and risk. Build once with the new tools, then run cheaply. More on this in when not to use AI and how we test AI models on your work.
Set honest expectations on ROI
Walker admitted his team was using AI to become more knowledgeable, yet when he asked how it helped win a pitch, there was one example.
Cheek's answer: a student starting a degree is not judging it by the cafe job's hourly wage. They are buying a long-term salary multiplier. Expecting needle-moving ROI in the first months is short-sighted. Most companies are in a learning phase, and the bigger risk is strategic: firms that embed AI in every function will replace those that do not.
Measure from day one, but tell your board which phase you are in. When the numbers do arrive, our CFO guide to the ROI of AI shows how to count them.
What we see when companies try this
Three patterns repeat in the work we do, and they line up with Cheek's advice.
The pilot has no starting number. Teams cannot say how long the task takes today, what it costs or how often it goes wrong. Without that, nobody can tell a leader whether the pilot worked, and the board hears "it felt faster". We insist on measuring the task before anything is built, and on writing down one money number that would convince you to go live.
Nobody owns going live. A sponsor attends the demo, the builder moves on, and IT or security were never in the room. The pilot stalls in the gap. A named owner on your side, with authority over access and process questions, closes it. This is the org-chart point in practice.
Success means a good meeting. I have seen pilots called a success because the board liked the demo, and weeks later nobody opened the tool. We judge a pilot on whether the team uses it every day on real work, with numbers you can read.
If you want the full version of this, it is in why AI pilots stall and the pilot to production playbook.
Speed is the small win
One more idea, from a separate MIT Sloan conversation between Bill Aulet and Paul Cheek.2 Aulet's warning is blunt: AI lets you do the wrong thing faster. If the task should not exist, or nobody wants the output, speed only gets you to a bad result sooner.
The bigger prize, in his framing, is what you do with the time saved. Spend it on talking to customers, running more small experiments and testing assumptions, the way a spreadsheet changed finance because of what-if analysis, not faster addition. He also treats AI output as roughly 80% right and useful for starting a team discussion, not for ending it.
Cheek adds a practical fix for companies where everyone uses AI their own way. Keep one source of truth, your strategy, as shared context, and generate downstream work from it so tactics do not drift from strategy. For you, that means asking a simple question in every review: what did we do with the time AI gave back?
If you are the incumbent
Cheek's advice on startups chasing your market is useful to you:
- Creation is cheap, distribution is hard. Anyone can build a product now. Getting the first customers is the bottleneck, and you already have them.
- Use that edge. With distribution in hand, building new things costs close to nothing. Move.
- Separate protective friction from plain slowness. Regulation in banking or aviation is a real moat. A committee chain that takes weeks to approve a decision is not.
If you sell services, expect buyers to pay for outcomes rather than hours, because agentic AI delivers finished work. Cheek still sees a big opening for services firms, but the pricing logic shifts.
A seven-step start for this quarter
- Commit your own time to the learning, not just budget.
- Start AI literacy with the board and leadership, then everyone, with protected hours.
- Run tinker days and weekly show and tell.
- Run an opportunity assessment with the four questions and get a ranked list.
- Redesign roles, workflows and the org chart so every agent has an owner, a budget and data limits.
- Pick the cheapest model that works per task and revisit as costs fall.
- Measure learning first, P&L second, and brief the board on timing.
Cheek calls this a people problem more than a technology problem. I agree. The model is the easy part.
Working through steps four and five and want a second pair of eyes? That is what we do at Skcript Applied Labs. If your pilots have already stalled, start with why AI pilots stall, and check where you stand with is your business ready for AI adoption.
Footnotes
-
Walker Webcast, "Building the AI-Driven Enterprise with Paul Cheek", Willy Walker in conversation with Paul Cheek, MIT Sloan. Watch on YouTube. Paraphrased in our own words; statistics are as cited in the interview. ↩ ↩2 ↩3
-
"MIT Sloan's Bill Aulet and Paul Cheek: AI + Entrepreneurship". Watch on YouTube. Paraphrased; the talk is about entrepreneurship, so the enterprise reading is ours. ↩
