The AI decision nobody can make for you
Executives at service companies are being sold a simple story: pick the right AI tools and models, plug them into your workflows, and watch productivity climb.
That story is now colliding with public data.
IBM’s 2026 CEO Study reports that 69% of CEOs say AI is already changing what they consider core to the business, not just what tools their teams use. 77% say talent and technology leadership are converging, and 76% say they now have a Chief AI Officer, up from 26% a year earlier.
At the same time, Bain’s analysis of IBM Think 2026 makes the blunt translation: the enterprises actually pulling ahead are not the ones with more AI pilots. They are the ones that have redesigned how the business operates.
For a 50–500 person service company, this is easy to dismiss as enterprise theater. It shouldn’t be.
Your biggest AI decision is no longer which model, which vendor, or which use case. It is:
What is the operating model you want AI to run inside?
Everything else is an implementation detail.
From tools to operating model
The newest research converges on a simple pattern:
- IBM says the C‑suite job has shifted from sponsoring AI pilots to rewiring how decisions are made and who has authority to change work.
- Bain says the bar moved "from AI pilots to an operating model" and ties 10–25% EBITDA gains to companies that have actually done that redesign.
- ArticSledge contrasts traditional operating models (slow decisions, sequential handoffs, technology as support) with AI operating models (real‑time signals, parallel execution, technology as a core business capability).
- EverWorker points out that board conversations have shifted from "Should we use AI?" to "Where does AI deliver ROI this quarter?"—and that most “strategies” fail because they never specify how work, governance, and measurement will run in production.
In other words: the value is no longer in having AI. The value is in how you operate once AI is in the loop.
For a service company CEO or COO, that boils down to four questions:
- Where does work live? (Who owns each workflow end‑to‑end?)
- How are decisions made? (What can run on rules, what must stay human?)
- How is risk governed at runtime? (Who can access, change, or approve what, in real time?)
- How does the system learn? (What gets recorded, reviewed, and improved after each job?)
If you can’t answer those questions today, adding agents and copilots won’t fix it. As Florian Schroeder put it in a separate piece:
"If your process is vague, AI will scale the vagueness. If your data is messy, AI will explain messy data confidently. If your approval process is weak, AI will create more things that need approval."
AI multiplies whatever operating model you already have.
Governance has moved from policy to runtime
The second thread in the latest research is about risk.
A Forbes governance playbook reports that 47% of large organizations and 68% of mid‑sized organizations either lack full visibility into employee AI use or have “shadow AI” in place. Check Point’s data suggests that 1 in every 48 enterprise prompts is high risk.
At the same time, Bain highlights IBM’s internal estimate that soon there will be 120 non‑human identities for every human user in large enterprises—agents, automations, background services. Traditional access-control and policy binders do not scale to a 1:120 ratio.
The conclusion is uncomfortable but clear:
Governance can’t live in a PDF anymore. It has to live in the runtime.
For a service company, that doesn’t mean standing up an AI council or hiring a head of risk tomorrow. It means much more concrete moves inside your operation:
- Inventory before policy. Forbes is explicit: you cannot govern what you can’t see. Start by listing every place AI is already touching customer data, employee data, scheduling, pricing, or approvals—whether IT owns it or not.
- Define autonomy tiers. The Forbes model is simple and applicable: decide where AI can act autonomously (e.g., drafts, sorting, low‑impact routing), where it can only recommend, and where human approval is non‑negotiable (e.g., financial commitments, legal communications, safety and compliance actions).
- Bake logging and receipts into the work itself. Jeff Clarke at Dell summarized this as: "In an AI workforce, every action needs a receipt." That is how you build trust in a system that is starting to act on its own.
If you don’t define those boundaries, your people will still use AI—they’ll just do it in shadows and chat windows you can’t see or audit.
The CEO’s job just changed
IBM’s study gives a useful stat stack:
- 69% of CEOs say AI is changing what they consider core to the business.
- 77% say talent and technology leadership are converging.
- 76% report having a CAIO, up from 26% a year earlier.
But the more important shift is qualitative. Axios’s C‑suite brief on agents argues that leaders who understand agents as a management layer, not a productivity tool, will design organizations others can’t compete with on headcount alone.
John Hagel, a long‑time advisor to Fortune 500 CEOs, adds a harder warning:
When the private questions in the boardroom are "How quickly can we automate?" and "How many jobs can we eliminate?", that is a going‑out‑of‑business strategy.
His alternative is what he calls “scalable learning”:
- Automation should free up people to work on novel problems, not simply remove them.
- The most valuable organizations will be the ones where people closest to the work are creating new knowledge about customers, operations, and markets—using AI as leverage.
For a service company, this is not abstract philosophy. It forces concrete design choices:
- Do you use AI to cut coordinators, or to let coordinators handle twice the book of business and spend the extra capacity on proactive customer outreach?
- Do you treat your best technicians’ and project managers’ reasoning as individual heroics, or do you capture their patterns and push them into how the whole operation runs?
- Do you measure AI by headcount reduction this quarter, or by how much new value your existing team can create with the same or slightly higher cost base?
IBM’s data suggests the companies that redesign team interlocks, decision rights, and feedback loops are more than twice as likely to hit their objectives. Hagel’s argument explains why: they use AI to change how the organization learns, not just how it executes.
What this means for a 50–500 person service company
Most mid‑market service leaders we speak with are stuck between two bad options:
- Tool sprawl. Each department experiments with its own AI apps. There are local wins, but no shared operating model, no shared memory, and no shared controls.
- One big program. A transformation initiative or external pilot that takes months, burns cash, and comes back as a slick demo that never survives contact with actual work.
The research points to a third path: start small, but start at the operating‑model layer.
A practical sequence looks like this:
1. Pick one workflow that actually matters
EverWorker’s executive guide is blunt: if a use case doesn’t move a C‑suite metric, it’s not an AI priority.
For a service company, that usually means one of:
- First‑response to inbound requests
- Scheduling/dispatch and rescheduling
- Job or project close‑out and billing
- Compliance checks or quality reviews
Pick one where:
- Volume is high and repeatable
- Outcomes are visible (you can tell good from bad)
- Risk is understandable (you can set autonomy tiers)
2. Design the workflow like an operating model, not a feature
Use the operating‑model components that recur across IBM, Bain, and ArticSledge:
- Objective and outcome. What is this workflow supposed to achieve, in language a CFO recognizes?
- Process redesign. What are the steps today? Which can be standardized or eliminated before any AI touches them?
- Data in the flow. What information does each step actually use? Is that data available where the work happens, in a structured way, or locked in documents and inboxes?
- Decision rights. For each decision, what can be rule‑driven, what needs judgment, and who owns that judgment?
- Governance and risk. Where can AI act alone? Where does it recommend only? How will you log and review actions?
- Measurement and learning. What will you track to know if it’s working? Who will review the data and make changes?
Only once those are clear does it make sense to ask which AI components to add.
3. Move governance into the work, not onto a shelf
Borrow directly from the Forbes playbook and Bain’s runtime‑governance framing:
- Build role‑based access around the workflow: who can see what, who can change what, who can approve what.
- Ensure every AI action carries a timestamped, human‑readable log attached to the job, ticket, or order it affected.
- Define simple incident‑response rules: if an AI action is wrong or risky, how does someone flag it, roll it back, and improve the rule or prompt so it doesn’t repeat?
You don’t need a policy binder to start. You need a single workflow where these controls are real, visible, and used.
4. Treat your best people as “intelligence managers”
Arjun Prakash, writing for the World Economic Forum, describes the core shift this way: as AI absorbs more execution, the marginal cost of work collapses, and the scarce resource becomes experts who can direct, govern, and improve that work.
In a service company, that means:
- Your best dispatchers, team leads, or project managers design the logic: when to escalate, what “good” looks like, what exceptions matter.
- They review AI output in higher‑risk steps and tune prompts, rules, or templates.
- They own the learning loop: after each week or month, they look at exceptions and adjust the system.
That is a promotion in responsibility, not a demotion to “checking the machine’s work.” It’s how you convert individual judgment into shared infrastructure.
Avoiding the two traps
If you take this operating‑model view seriously, two common traps become easier to avoid.
Trap 1: Shadow AI as the only real AI
If you don’t provide a visible way to use AI in the work—with clear rules and logging—your people will still use it, just not where you can see it. That’s how you wake up one day with customer data in unvetted tools and no audit trail.
The fix is not a crackdown. It’s to offer a safer, sanctioned path that is genuinely useful in the workflow that matters, and move governance into that path.
Trap 2: Automation used to defend an obsolete model
Hagel’s warning is aimed at large enterprises, but it applies directly to regional service firms.
If every AI conversation in your leadership team is about how many coordinators, schedulers, or back‑office staff you can cut, you are using new technology to preserve an old way of working: centralized decisions, rigid roles, and people measured only on volume.
The alternative is harder and more valuable:
- Use AI to take routine work off the plate of good people.
- Use the freed capacity to improve customer experience, shorten cycle times, and create new services your competitors can’t match.
- Measure AI by improvements in outcomes (margin, response time, NPS, revenue per employee), not by how many line items in the payroll budget disappeared.
The services and consulting giants you buy from are already being forced into this shift. Analyst firm TBR’s 2026 outlook notes that they are moving from time‑and‑materials to fixed‑price, outcome‑based services, and from human‑intensive delivery to AI‑driven operations. Mid‑market service companies will be held to the same expectations.
The operating-model test for your next AI decision
The next time a vendor pitches you an AI solution or your team proposes a pilot, ask three questions before you sign or approve anything:
-
What part of our operating model does this change?
- If the answer is "none, it just makes people faster," you’re probably buying a 20–40% gain at best—and adding complexity without changing how the business runs.
-
Where will governance live when this is in production?
- If the answer is "we’ll write some policies" or "IT will handle it," you’re likely heading for shadow AI and audit surprises. Look for concrete answers about visibility, logging, and autonomy tiers inside real workflows.
-
Who owns the learning loop?
- If nobody can name a specific person who will review metrics, exceptions, and outputs and then adjust the system, the gains you see in month one will decay by month six.
You do not have to solve “AI for the whole company” this quarter.
But you do have to decide—as CEO, COO, or head of operations—what kind of operating model you are building for the next five years, and whether your first AI steps move you closer to it or just decorate the old one.
That decision can’t be delegated to a model vendor, a line manager, or a pilot team. It is now the core of your job.