Executives need an AI plan, not another demo

On September 29, OpenAI used its DevDay stage to introduce dots, always-on agents that pursue goals in the background and can be messaged through Slack and Teams, as TechCrunch reported that day. They went live for ChatGPT Pro and Business Premium users the same day. The examples were good: a developer’s agent that watches customer feedback and ships fixes, a scientist’s agent that reruns analysis when new data lands. TechCrunch also noted that much of this was already possible through Codex and similar tools. Dots package it around an agent that acts on its own.

The next morning, BCG published its annual study of 1,300 executives. The headline was encouraging: about half of companies now get value from AI. The detail underneath was the part every CEO should read. Among the companies BCG calls future-built, 61% have moved from isolated pilots to a multiyear, enterprise-wide AI program. Among the laggards, 17% have.

Put those two days side by side and you have the problem in front of most executive teams this fall. Demos are arriving every week. Plans are not.

DataCompanies running a multiyear AI program, not pilots

Future-built companies61%
Laggards17%

The leaders stopped running disconnected pilots and committed to a program. Source: bcg.com

A demo can’t tell you which work to hand over first

I like demos. I’ve spent twenty years building companies where media meets technology, and watching a new capability for the first time is still one of the best parts of the job. But a demo answers one question: can this be done? Your company needs answers to different ones. Which work do we hand over first? Who owns it? What will we stop paying for? How will we know it worked?

A demo can’t answer any of those, because they depend on your workflows, your people and your numbers. An executive team that watches five demos a quarter and writes nothing down ends the year with a lot of opinions and the same operating model it started with.

The BCG data says this plainly. Future-built companies put the C-suite in charge of the AI program, and 95% of them use clear KPIs or track the profit-and-loss impact directly. Companies that track AI on the P&L realize 3.6% AI value, against 1.2% for companies that don’t formally measure it, by BCG’s count.

The plan matters more as agents get more autonomy

Dots and the agents like them are built to act with minimal oversight. That raises the stakes on planning.

BCG found that 42% of companies expect AI agents to have real decision-making authority by 2030, but only 5% have the full set of controls in place today, per its September 30 release. Companies with all six controls BCG tracks generate three times the agentic value of those with just one.

That gap is a planning gap. Nobody buys “controls” from a vendor during a demo. Controls come from decisions an executive team makes in advance: what an agent may touch, who reviews its output, what it may never do without a person, what gets logged. If those decisions aren’t written down before the agents arrive, they get made one incident at a time.

By the numbers

of companies have the full set of controls for AI agents today, while 42% expect agents to have real authority by 2030

Source: bcg.com

What a real AI plan contains

A plan doesn’t need to be long. For a company doing $10M to $100M a year, I’d expect it to fit on a few pages and cover these things.

The first three to five workflows, ranked. Each one is named specifically, with the department, the weekly volume and the hours it takes today. Rank them by value and by how easy each is to measure.

An owner for each. This is a person who runs that part of the business and has the authority to change how the work is done.

The measure for each. It might be hours saved, cost per unit, response time or revenue. Pick one per workflow and take the baseline now.

The rules. Write down what data agents may read, what they may change, what always needs a person’s approval, and who sees the logs.

The people plan. Say who gets trained, on what, by when, and how you’ll know they can do it.

The budget and the tools. List what you already pay for, what you’ll add and what you’ll cancel.

A review date. Put it ninety days out, on the calendar, with the CEO in the room. At that review, each owner reports the measure against the baseline, and the team decides which workflows expand, which get fixed and which get dropped.

That is my working version of what BCG means by an enterprise-wide program. It is also the document most executive teams don’t have.

As for who writes it, the CEO has to own it. Delegating the AI plan to IT turns it into a software purchasing exercise, and BCG’s finding that more than 80% of AI spending now sits outside the enterprise IT budget tells you the work has already moved into the business units. The plan should follow the work.

That doesn’t mean the CEO writes it alone. The best version comes from the CEO and the people who run each function, in the same room, working from their own workflows and numbers. It’s the reason I co-host AIXE with Daniel Shemtob: two days where founder-CEOs build their company’s AI plan instead of watching more presentations about what’s possible.

Whatever format you use, the output should be the same: a written plan with names and dates on it.

Run every new demo through the plan

Once the plan exists, the demos stop being a distraction and start being useful. Every new launch gets one question: does this help with one of our ranked workflows? If it does, the owner of that workflow tests it against their baseline. If it doesn’t, it goes on a list for the 90-day review.

This is how you keep up with a market that ships something new every week without rebuilding your strategy every week. The plan sets the order of work. New tools compete for a place inside it.

It also changes the conversations you have with vendors. A sales team that walks into a company with a written AI plan gets asked specific questions: here is our workflow, here is our volume, here is our measure, show me. That is a far better meeting for both sides than a tour of features, and it is much harder for a weak product to survive.

What to do this week

ChecklistYour first draft of the plan

Three to five ranked workflows
One owner per workflow
One measure and a baseline each
Written rules for agent access
A review date 90 days out

Block two hours with your leadership team. No vendors, no demos.
Ask each leader to bring one workflow from their area with this week’s volume and the hours it takes.
Rank the workflows together and pick the first three.
Assign an owner and a measure to each before you leave the room.
Write the rules for what agents may touch, even if you haven’t deployed one yet.
Put the 90-day review on the calendar today.