A CEO in financial services laid out his leadership team for us last week. He uses AI every day and can see a dozen places it would help. Below him sit two senior managers who run the day-to-day and employ everyone else. One, he said, has “some understanding and vision around AI, but not a lot.” The other has “almost none.” His conclusion was blunt: “They are not going to see the opportunities for AI themselves. I need somebody who understands how AI can be used to come in and tell me.”

He is describing the most common shape of a stalled mid-market AI effort. The vision lives at the top. The processes live in the middle. Nobody is standing where the two meet.

Why the people who know the work cannot see the opportunity

It is not a talent problem. Managers are closer to the work than anyone, which is exactly why they struggle to see it differently. A process they have run for eight years looks like the only way the process can run.

The data says they are using AI, just not for this. A General Assembly and EZRA survey of business leaders found that 82 percent use AI regularly, mostly for search, summaries and emails, while fewer than a third use it for anything strategic such as scenario planning or organizational design. Salesforce found that 78 percent of managers feel responsible for their team’s AI adoption and 51 percent are anxious about keeping up themselves.

So the manager is responsible, anxious, and using AI as a faster search engine. Asked “where could AI help your department?”, the honest answer is usually a shrug, because the question requires knowing what AI can do and knowing the process at the same time, and almost nobody holds both.

The CEO sees the destination. The manager knows the terrain. AI activation is the work of putting those two people in the same room with a map.

The three gaps between vision and process

1. The vocabulary gap: Executives talk about agents and automation. Managers talk about the Tuesday report, the three-way match and the quote that takes four hours. Until someone translates between them, the opportunities stay invisible to both.
2. The permission gap: Managers assume AI is a leadership project or an IT project. Nobody has told them that redesigning their own workflow is in scope, so they wait.
3. The evidence gap: A manager who has not seen AI do real work in a company like theirs has no reason to believe it will. A communications team we work with put it well during a diagnostic: they “spend a lot of our time hunting for things,” and only when they watched AI do that hunting did the opportunity become obvious.

What closes the gap: a diagnosis, not a demo

The CEO’s instinct to bring in someone who understands how AI can be used is right, but the form matters. A demo impresses the executive and bounces off the manager. That is one reason MIT’s NANDA study found 95 percent of enterprise AI pilots deliver no measurable return: the pilots were chosen far from the work. A diagnosis does the opposite. It starts in the manager’s world.

The way we run it, a two-week Rumble sits with the ops manager, the controller, the service lead and the people who report to them. We walk the work as it is actually done, not as the org chart describes it. Where is energy going that a system should carry? Which steps exist only because the old system could not do them? Where does one person hold knowledge the whole department depends on? Our post on where mid-market companies should start with AI activation describes the questions we ask.

Three things happen when it is done this way. The manager hears their own problem described back to them in their own words. They see a concrete example from a similar company. And they get asked to co-own the first win, which turns a passive skeptic into the person who wants it to work. One quoting team we work with began exactly here. The manager who could not name an AI opportunity in week one now runs a workflow that drafts every request and has kept her team out of backlog status for months.

Build the manager’s capability, not just the system

The gap reopens if the system lands and the manager’s understanding does not move. Writer’s enterprise survey found that 36 percent of companies deploying AI agents have no supervision plan, which is what happens when the people running the work were never brought in. Two investments prevent that.

First, literacy before tooling. Managers need to know what AI does well, what it does badly, and how to spot a candidate task, which is different from learning a particular product. Our piece on the difference between AI literacy and AI training explains why the order matters.

Second, give the manager a seat on the roadmap. Forrester’s 2027 planning guidance says to fund AI only where it can compound value, and the manager is the only person who knows where that is in their department. When the next 90-day review happens, the manager should be in the room proposing the next workflow, not hearing about it afterward. Managers who have seen one win in their own department become the best source of the next five.

That is what the CEO was really asking for: not someone to tell him the answer, but someone to stand between his vision and his managers’ processes long enough that both sides can see the same opportunity. That is the job of an AI Activation Partner, and it is where closing the gap between AI aspiration and AI that actually works begins. If your leadership team sees the opportunities and your managers do not, a two-week Rumble puts them in the same room with a prioritized roadmap and a first win ready to build. Put AI to work for your people, starting with the people who run the work.

Frequently Asked Questions

A CEO told us last week that his board gets the 2027 budget on December 8, and he has about 30 days to decide what AI line to put in it. He is not short of ideas. He is short of a number he can defend, because every option he has seen is either a per-seat tool nobody has a plan for, or a six-figure project with no proof behind it. His exact question was whether his company was “ready to invest at a level with consistency over time,” rather than fund one more experiment.

That question is the whole budget conversation for mid-market companies this quarter. Here is how to answer it.

The money is loosening, and the overruns are real

The appetite is back. Forrester’s 2027 planning guides report that more than 80 percent of leaders expect larger budgets over the next 12 months, with up to a quarter expecting growth of 10 percent or more. The same guides warn that where you invest matters more than how much, and that AI funded into a broken process scales the dysfunction.

The overruns are the other half of the picture. Futurum’s survey of technology buyers found that 46.9 percent of enterprises spent over their AI budget this year, and nearly half of those went back for supplemental funding. Finance teams noticed. The report expects AI budgeting to become a formal approval gate in the next planning cycle, which means the days of “we will figure out the cost later” are over.

The companies that overran did not budget too little. They budgeted for the wrong thing.

What a 2027 AI budget should actually contain

Most mid-market AI budgets fail because they contain one line: “AI.” A defensible budget has four, each with a different owner and a different confidence level.

1. Diagnosis: A fixed-price, two-week look at where your people spend energy that a system should carry, ending in a prioritized roadmap and a first win ready to build. This is the cheapest line and the one that de-risks every other line. We price a single-function version at $10,000 and a business-wide version at $25,000, because every dollar is spent deciding what to build before anything is built.
2. The first proven win, and the cadence behind it: One production system, scoped to a measurable outcome, typically starting around $25,000. A steel manufacturer we work with began with the reading of certificates that seven people handled as part of their jobs. Removing that work freed the equivalent of a full-time employee each year and paid for the next project. That is the point: wins compound, and the budget should fund the cadence that produces them. Depending on the scope of work and the size of the organization, a partnership that delivers a consistent cadence of wins runs from $12,000 to $90,000 per month. Budget it as a cadence with a 90-day gate, not as a one-off project.
3. Run costs: Hosting is modest. Model usage is the variable, and it is falling fast. In October, OpenAI shipped GPT-6.1 Terra at half the price of its predecessor and Anthropic cut Haiku pricing by roughly 90 percent. Watch the expiry dates too: Gemini 4 Argon’s launch price doubles on December 1. One distribution client runs about a thousand AI-assisted quotes a day for under $20. Budget the run, but do not inflate it. Our post on why tokens got cheaper but your bill did not explains where the real costs hide.
4. Tools and training: Per-seat licenses for the people who will use them, plus the enablement that makes the seats worth paying for. Keep this separate from the build budget so it cannot quietly absorb it.

The budgeting mistakes we see every December

The first mistake is budgeting a tool rollout as if it were a strategy. MIT’s NANDA study found that 95 percent of enterprise AI pilots deliver no measurable return, and the cause was integration and learning, not the tools. A hundred ChatGPT seats is a cost, not a plan, and without a roadmap it turns into shadow AI by March.

The second is budgeting the build and forgetting the run. Several companies in the Futurum data carried their overrun into the next year because nobody owned the usage line. Put a name on it.

The third is asking the board for a large, open-ended number. Boards fund sequences, not leaps. A diagnosis in Q1, a first win by Q2, a second by Q3, each gated by results, is a budget a CFO can approve and a board can follow. Our guide on how to pitch an AI initiative to the board walks through the framing.

The fourth is budgeting hourly. AI work carries unusual scope uncertainty, and hourly billing hands all of it to you. Fixed prices for the diagnosis and the first win keep the risk where it belongs.

How to answer the consistency question

The CEO’s real question was about commitment, not cost. Here is the honest answer we give.

Yes, AI that works takes consistent investment over time, because the gap between AI aspiration and AI that actually works is closed one workflow at a time, with the people who do the work. No, that does not mean a large upfront bet. It means a small diagnosis, a proven win, and then a partnership that scales as the wins land, with every 90 days a checkpoint where results, ROI and risk are reviewed and anything that is not working is fixed at no additional cost.

The balance the Rumble is built to strike is between the two halves of that answer. Too little investment and the wins arrive one at a time, months apart, each one re-justified from scratch. Too much and you are back to the open-ended number the board will not approve. A steady monthly commitment, sized to your scope and your organization, is what lets each win fund and de-risk the next, so the returns compound instead of resetting. For most mid-market companies that commitment lands between $12,000 and $90,000 per month, and the roadmap says which end of that range you need and why. That is how we structure every engagement, and it is why more than half of our revenue comes from clients who came back.

The December board meeting does not need to approve “AI.” It needs to approve the first two lines and the rule for unlocking the rest. Our post on measuring AI ROI before you invest gives you the numbers to attach.

If you are building the 2027 budget right now and want a number you can defend, a two-week Rumble produces the roadmap, the first win and the cost of each. Put AI to work for your people, and give the board a sequence instead of a leap.

Frequently Asked Questions