Here is the most common thing I hear from leaders who bought AI and are still waiting for results: it is good at drafting, or tools are still in silos, and not much else.
It writes a decent email. It summarizes a document. Ask it to do anything that touches workflows running across the business, and it stalls.
The reason is straightforward: drafting is generic, but execution is organizational. Everything past a first draft relies on internal know-how: how decisions get made, which edge cases matter, and what a finished piece really looks like. That knowledge has always lived naturally in a few key people's heads, unwritten simply because the process never demanded it. Tools do not stall because the technology is limited; they stall because the implicit context around them has not been built into the system yet.
The research says this is close to universal. Grant Thornton asked 950 business leaders to name the factors behind AI underperformance in their organization, offering “AI is not underperforming” as one of the options. Eight percent chose it.
The more useful question is why 2026 produced research saying two opposite things at once. In March, KPMG asked roughly 2,100 senior leaders whether AI was delivering meaningful business value, and nearly two-thirds said yes. In its Q2 wave, fielded April to May and published in June, KPMG asked a separate sample of 2,145 leaders a narrower question: had the organization established ROI? Seven percent had. That rose to 14% where accountability was clearly defined, particularly at CEO level, and sat at 4% where the accountability had not been assigned.
The good news the headlines usually miss
Nearly every AI study in circulation surveyed large-enterprise executives, and the research that looks specifically at companies your size tells a more encouraging story.
Ipsos surveyed 750 owners and operators of businesses with 500 or fewer employees for the US Chamber of Commerce Foundation this spring. Among those already using AI, what they report is not a struggle.
Businesses with 500 or fewer employees
US Chamber Foundation and Ipsos · n=750 owners · 2026
Among owners already reporting a positive impact, 65% say the time saved goes into more or better work, and 56% into learning, planning, or review.
That last point matters more than it looks. Among frontline workers who save a full day a week or more, 66% receive little or no guidance on what to do with the time AI gives back, and more than half never redirect it. Smaller organizations are markedly better at this. You are closer to the work, so the recovered time has somewhere obvious to go.
So the four opportunities below are not evidence that you are behind. They are the four places where the leverage sits.
Opportunity 01 · Direction
Tools arrive faster than the direction to use them.
Asked to name the single biggest driver of AI ROI, 51% of leaders said strategy. That is more than talent, tech stack, and governance combined. Among operations leaders, the function closest to where AI has to work, 22% report a strategy that is both developed and implemented.
In practice, that usually looks like a run of proofs of concept: interesting, successful, and quietly stopped. Some of that is healthy, because a proof of concept that teaches the team something has earned its keep even if it never ships. The pattern worth noticing is when everything stalls in the same place: experiments build competence but never land where leadership has said the business needs to move, so nothing accumulates.
- Pilots that were well received and then simply stopped, with no decision recorded either way
- The person driving AI can describe the goal clearly, and cannot point to where leadership wrote it down
- Expectations that have shifted more than once since the tools were purchased
“As a 65 to 70 person manufacturer, we realized that ‘we need AI’ was too vague to be a strategy; we had to start with specific business problems and bring a small, cross-functional group together to tackle them…”
Thomas Deslongchamps, Pindel Global Precision, in the US Chamber of Commerce Foundation and Google whitepaperEnterprise AI strategy runs on eighteen-month roadmaps, dedicated AI teams, and board reporting cycles staffed by people who do nothing else. None of that is available to an organization where the person leading AI is also running projects, managing vendors, and covering the help desk. The work those structures perform still has to happen. It just has to happen at a size that fits.
It also looks different at this size. In a large company the strategy has not been written. In yours it usually exists, it lives in one person's head, and it has not yet been given a sponsor. What is missing is not planning. It is shared direction, so that asking another department to change how it works reads as a common goal rather than as one person's request. Sales does not take workflow direction from IT, and it should not have to.
Your substitute for a governance council is one sponsor conversation. Ask the CEO or owner to put a single paragraph in writing: the outcome AI is expected to produce this quarter, the workflow it applies to, and who owns it. That paragraph is what turns a personal initiative into an organizational one, and it is what gives a champion the authority to ask another team for change.
Most leadership teams have already said the organization should be using AI. That is a mandate, and a mandate is broad by design. The work of this conversation is to narrow it together into something specific and measurable, because a team cannot act on a direction it cannot picture. You are not asking for a new decision. You are making the existing one concrete enough that people know what to do on Monday.
For more on connecting this to a specific outcome, see Direction, Not a New Assignment.
Opportunity 02 · Workflow
What changed about the work itself?
Deloitte found that 48% of organizations introduced AI without redesigning the workflows or roles it sits within. Meanwhile the Chamber research found that 73% of small business owners using AI say it has already changed employee roles and responsibilities.
Those two findings sit together well, and the space between them is the point. The work is already changing. The question is whether anyone is designing that change.
BCG put a number on the difference. Surveying 11,749 workers across fourteen markets, published June 3, 2026, it found:
BCG, AI at Work, fourth edition, June 2026, n=11,749
The most common measurement mistake follows directly. Organizations build a scoreboard from what is easy to count, which is licenses issued, logins, token consumption, and training completed. Every one of those numbers can rise while nothing about the work changes. Probe the same organization on productivity or on what has been built that could not be built before, and the answers thin out.
- Your AI reporting is mostly usage: seats, logins, consumption, completions
- Two teams solve the same problem in parallel, in different ways, without knowing it
- The knowledge a workflow depends on lives in one or two people's heads and has never been documented
- Nobody can produce a before-and-after for a single process
Choosing which workflow to rebuild is where instinct tends to fail. The visible ones nominate themselves: well documented, widely disliked, easy to point at. Those make credible quick wins and they are worth doing. The transformational ones are usually buried, running quietly across two or three departments that never compare notes, or narrowing to a single person everything has to pass through.
Two axes sort them:
- Business impact, meaning what happens if this runs slowly or fails
- Execution speed, meaning how the process runs today, manually or automatically
High impact and still manual is your risk zone, and that is where AI belongs first. Running that exercise properly takes the people who own those processes in the room together, which makes it a question of shared direction rather than scheduling.
A tool can be licensed by one person and used occasionally by another without anyone owning the workflow end to end. That is access, not adoption. And the reason so much AI stalls at drafting sits right here: nobody decided which workflow mattered enough to map, so the knowledge that runs it stayed where it has always been, and the tool was left to work without it.
At Epsilon I watched this play out on client data onboarding, which ran from analysis and requirements through documentation, configuration, testing, and deployment before a single file reached the platform. Rebuilding it meant treating the whole chain as one workflow rather than speeding up the step at the front. Across the workflows my team rebuilt that way, keeping a human in the loop at the judgment points, the workflows returned 30 to 40% of the hours they used to take. That figure is my own observation from those engagements.
Pick one workflow from the risk zone, not from the complaints list. Give it a named owner, a documented before state, a defined target, and a date ninety days out when you will measure. One workflow rebuilt completely is worth more than five touched lightly.
One test before you commit to it. The people who would rebuild it are already busy, so the workflow has to be one they personally lose time to now. If the redesign does not give those specific people hours back inside the ninety days, the work competes with their day job and loses to the first real deadline. That is the most common reason a well-chosen pilot quietly stops.
For more on choosing the right workflow, see Which Workflow Is Worth Rebuilding First.
Opportunity 03 · Governance
What happens when the AI is wrong?
Most people pause on that question. Not because they have not thought about it, but because there is no written answer.
Seventy-eight percent of organizations lack strong confidence that they could pass an independent AI governance audit within 90 days. The reason shows up as soon as you line up what was approved against what is being watched.
Share of organizations, same survey base
Grant Thornton 2026 AI Impact Survey · n=950
Approval moved faster than the structure needed to evaluate it. If you do not have a board, read that as your owner, your CFO, or whoever signed the purchase order, because the sequencing is identical and the oversight layer is usually thinner.
Deloitte maps the consequence. Sixty-nine percent of organizations sit at the most conservative end of AI autonomy, either running none at all or limiting AI to low-risk, reversible actions. That caution is not principled, it is the default position for any organization that has not yet defined what AI can be trusted to handle. Grant Thornton found 74% are already giving agentic AI — software that takes multi-step actions without a person approving each one — access to their data and processes, against the same 20% with a tested incident playbook.
Here is the version of this that shows up in practice. A client-facing assistant is answering questions consistently and correctly. The platform underneath it ships a new model version. The answers change. Nothing broke, nothing was misconfigured, and no alert fired, because the system was working exactly as designed. The only thing that would have caught it is somebody comparing today's answers against what a good answer looked like last month.
- You could not produce a list of every AI tool running in the business this week
- Software you already own has switched AI features on by default, and nobody approved it
- There is no written description of what a correct output looks like for any workflow
- There is no AI incident plan, or one exists and has never been run against a real scenario
Large companies are working through ungoverned deployment. You are working through ungoverned proliferation: a personal chatbot account in marketing, someone in finance pasting into a generative tool, an AI notetaker expensed on a credit card. Most of that is not misconduct. It is motivated people routing around tools that lack the access or context to be useful, which makes it a signal about configuration rather than about discipline.
Governance at this scale is smaller than the enterprise version, not lighter. It splits into two halves that need different kinds of attention, and both need a named owner.
- What the AI is doing, and in which workflow
- Which outputs it can act on alone, and which need a named reviewer
- What good output looks like, so quality can be judged
- What happens, step by step, when it is wrong
- A written AI policy people have read and can point to
- A standing group that reviews AI decisions, however small
- A path to approve a new AI workflow before it goes live
- Data handling, vendor review, and where liability sits
The left column is inseparable from workflow redesign, because a process cannot be rebuilt around AI without deciding what the AI is allowed to decide inside it. The right column is a different discipline. Whether you build it in-house or with outside help depends on the HR and IT leadership already in place, and there is now a growing set of platforms that carry much of the policy, review, and evidence work for organizations without a compliance function. Knowing which position you are in is itself a finding, and it belongs in an AI strategy rather than in a binder discovered after an incident.
Answer three questions in writing before AI touches customer data, financial records, or anything customer-facing. What is the AI doing? What is it deciding without human review? And what happens when it is wrong?
The third is where most organizations stall, and it has a practical answer in four steps:
- Write down what good looks like. A wrong output cannot be defined until a correct one has been.
- Review against that baseline on a cadence. Someone checks output on a schedule rather than waiting for a complaint.
- Name the escalation path. When a deviation appears, can the person who found it resolve it, or does it go to someone else?
- Classify what you found. A defect in the system, or an operator who needs clearer instruction about the expected output. In my experience it is more often the second, and the fix is context rather than technology.
Those four steps describe one workflow. The organization also needs the structure that holds them: a written policy, a small group reviewing AI decisions on a regular cadence, and a defined path for approving a new AI workflow before it goes live. At your size that group may be three people meeting monthly, and the review may look closer to peer review than to a committee. Small is fine. Absent is not.
For more on what this looks like in practice, see What Your AI Does Not Know About Your Business and Deciding What Success Looks Like Before You Start.
Opportunity 04 · Workforce
Your people are further along than the org chart suggests.
This one is the most often misread, because the leaders most confident about readiness tend to have the least direct visibility into it. Grant Thornton found CIOs and CTOs five times more likely than COOs to say their workforce is ready, at 39% against 7%. Twelve percent of leaders say their people are fully ready, and 81% say “fairly” or “mostly.” As Grant Thornton puts it, ‘Mostly ready’ is not close enough.
But the framing of a readiness problem misplaces the cause. Ipsos and Google found that only 37% of employees have received any organizational guidance on AI use, and 14% have been offered training. Small businesses are nearly three times less likely to offer AI training than large ones.
People are not unready. They have mostly been handed a login and left to work it out, and the ones who have worked it out did so on their own time.
You can identify them without a survey. They are the ones asking for more access, or quietly using something outside the approved set because the sanctioned tool has not been given what it needs to be useful to them. In most businesses that is a signal rather than a violation. If you handle customer financial data, health records, or anything under a contractual confidentiality obligation, it is both, and the exposure starts the day it begins. Either way the message underneath is the same: your most motivated employees are telling you exactly which piece of the work has never been written down.
The people not experimenting are not the problem either. What surfaces when you ask is fear about what this means for their job, uncertainty about what to do with time that gets freed up, and a skills gap nobody has offered to close. The first deserves a direct answer: the Census Bureau found AI-related employment reductions in 2% of firms.
- Someone has asked for access or a system connection more than once and has not received an answer
- Tools are being used that IT did not approve, by people who are good at their jobs
- No written guidance exists on what employees are expected to do differently
- Nobody has said out loud what happens to the time AI gives back
The skills gap tends to close under one condition, which the Chamber and Google research names as the practice separating the organizations making real progress: a small cross-functional group, drawn from operations, finance, marketing, and leadership rather than from IT alone, working one specific business problem together.
Formal training for everyone is usually out of reach, and there is a harder constraint underneath it. The two or three people already experimenting are typically your two or three most capable employees, and they are the ones you cannot spare.
That is the real objection, and it has an answer. It is why you run one workflow rather than five, and why the workflow should be one those same people already find painful. You are not adding to their week, you are trading against it. If the pilot does not give them time back, the workflow was the wrong choice.
The cohorts I have run show that when people work a real problem alongside peers, support comes from the group rather than from a curriculum.
Name the two or three people on one team who are already experimenting. Give them a single shared workflow problem and a few weeks. Make the result visible to everyone else. When the person at the next desk can see a colleague getting real time back on work they both dislike, and can walk over and ask how, the adoption calculus changes. That is not a training program. It is what training programs were trying to reproduce.
For more on finding them, see Who Are Your AI Champions, and What Did They Learn?
How This Work Sequences
None of these four are technology failures. They are questions of direction, structure, accountability, and ownership, and most organizations have simply not had the framework to work through them in order.
There is also a distinction worth naming directly: choosing a tool and changing how work gets done are two different efforts. The first is a purchasing decision, and organizations are practiced at those. The second is organizational work: deciding what should change, agreeing on what better looks like, and carrying people through it. Without budgeting time and resources for the change management work that follows access to the tools, progress will stall.
Resourced deliberately, the work has a sequence, and following it keeps progress from stalling.
Understand
Leaders need a clear picture of what AI can realistically do inside their own operation, not what it does at a company ten times their size. This article is part of that step, and so is the assessment below. The purpose is a shared, honest read of where things stand, so the next decision rests on evidence instead of pressure.
Identify
A diagnostic across five dimensions, built from direct observation rather than self-assessment, finds where AI is most likely to produce a measurable result. It scores candidate workflows on business impact and execution speed, surfaces the ones running quietly across departments, and names the one worth rebuilding along with the person accountable for it. Undocumented knowledge, missing policies, and ungoverned tools surface here as findings with owners attached.
Adopt
The chosen workflow runs as a pilot with a goal, an internal champion who owns it, and a measure defined before the work starts. Ninety days, with checkpoints. Each deployment carries a written statement of what the AI does, what it decides, where the human review threshold sits, and what happens on an error. The exit condition is a documented result and a team that could run the next pilot without outside help.
What the good version looks like from the inside is fairly ordinary. Work that used to take hours arrives in minutes, and the person receiving it spends their time applying judgment and context to serve the client better. An owner finds the team relies on them less, and people spend more of the week inside the work they were hired for.
So the question worth sitting with is not whether AI can work in your organization. It is which of these four opportunities is the one still open in yours.
Common Questions
What is the difference between AI access and AI adoption?
Access means people can log into the tool. Adoption means the work itself changed: a specific workflow runs differently, someone owns that outcome, and the difference was measured against a baseline. Login counts, license utilization, and active-user numbers measure access. They are the most common proxy for adoption and the least reliable one.
Why does our AI only seem good at drafting?
Drafting is the one task that requires no knowledge of your organization. Everything past it depends on how a specific piece of work gets done, who decides what, and what a good result looks like, and that knowledge usually lives in a few people's heads rather than anywhere a tool can use it. Choosing one workflow and capturing how it really runs is what moves AI from writing about the work to doing part of it. The system connections follow that decision; they do not lead it.
Why is our AI not producing measurable results?
In the 2026 research the recurring answer is not the technology. No workflow was redesigned around the tool, no single person owns the outcome, and no baseline was captured before deployment, so there is nothing to measure against. KPMG found organizations with a named accountable executive were more than three times as likely to have established ROI.
Does a company with 50 to 500 employees need an AI policy?
Yes, and it can be short. A workable policy names which tools are approved, what data may never be entered into them, which decisions require human review, who approves a new AI workflow before it goes live, and who to tell when something goes wrong. It needs a named owner and a review cadence.
You do not have to draft it from a blank page. The NIST AI Risk Management Framework and ISO/IEC 42001 are the recognized structures, and an adviser can map either one down to your size in an afternoon rather than handing you the standards to read. The same conversation should cover which state requirements apply to you, since several states now have AI disclosure or employment rules in effect and they differ.
What should a mid-market company do first with AI?
Diagnose before deploying anything further. Identify which high-impact workflow is still running on manual effort, confirm someone with authority will own the outcome, capture a baseline, and set a measurement date ninety days out. Selecting the workflow correctly matters more than selecting the tool, and it is the step most often skipped.
Sources
Every figure above is drawn from the primary source listed here. Where a source hedges its own finding, that hedge has been preserved.
Most AI research in circulation is published by firms that sell AI services or advisory, including several cited here. Where a government probability sample exists, it is used in preference to an executive panel, and the distinction is noted in the text.
Last reviewed 26 August 2026. Regulatory statements are current as of that date and are reviewed quarterly.