Insights
/
Business Transformation
/
Why AI Stalls in Service Businesses: What 105 Real Conversations Reveal
Business Transformation
Most of what gets written about AI in small firms comes from surveys. Respondents tick a barrier from a list, and the list decides what you learn. That works for counting. It's weak at explaining why a 28-person agency that keeps shipping new AI workflows ends up with a team split into two camps.
So we went where agency owners, consultants and MSP operators talk to each other without a vendor in the room. Between October 2025 and October 2026, Orbflo read and coded more than a hundred first-hand Reddit discussions from service businesses, in two connected studies. The first asked why firms adopt, limit or reject AI. The second mapped what goes wrong once they try.
The short version: the hardest problems sit around the tools. Deciding where AI belongs, checking what it produces and finding someone with the time to own it dominate the evidence.
Both studies used the same approach. We searched agency, consulting, marketing, PPC, UK small-business, recruitment, MSP, architecture and accounting communities for first-hand accounts of AI adoption, refusal, ROI, training and day-to-day workflows. Each discussion was coded against a set of themes, and every quote below is an exact excerpt, original spelling retained. We framed each theme as a job to be done: the progress the business is trying to make, rather than the feature it asked for.
A word on what the numbers mean. This is a purposive qualitative sample, so a count tells you how often a theme came up among these conversations, not how common it is across UK SMEs. One thread counts at most once per theme, however many comments it has, and a single thread can support several themes. Positive accounts and successful non-AI alternatives were deliberately included, so the research doesn't assume more AI is always the right outcome.
The voices skew towards operators and employees, with owners making up about a quarter of the usable sources. That matters, because employees describe the effects of decisions and owners describe the decisions themselves. Where an owner subset changes the picture, we show it separately.
The first study coded the decision in each discussion: did the firm adopt, limit, pull back or say no? Very few businesses made a single yes-or-no call. Most adopted one task and refused another.
“I only use it to write job descriptions. I really don't find the need to use AI in any other parts of the process tbh.”
Recruiter, source thread
Only three primary discussions land on rejection or resistance, and one of those is an employee pushing back against a mandate. The larger story is the 15 firms using AI within limits they chose on purpose, and the 6 that pulled back after trying. A recruiter in the sample put the case for a limit plainly: “AI makes everything homogeneous”, so outreach and interviews stay human. A small business owner captured the reluctant middle: “I honestly hate AI but am I really doing the world any good by avoiding it?”
The top reason has little to do with technology. Thirteen discussions turn on skills, ownership and capacity: firms expand when someone has relevant training and protected time, and stall when nobody knows what to do. Close behind is the reason most firms say yes in the first place, freeing capacity to grow.
“The goal isn’t to reduce our existing staff, but to grow without needing to keep adding people for admin, coordination and follow-up”
UK business owner, source thread
Read the conditions side by side and a pattern shows up. Almost every reason to adopt is narrow and specific, and almost every reason to hold back is about the cost of making AI fit.
The second study went deeper on what goes wrong. Of its 105 usable evidence sources, 80 describe a challenge directly, and twenty distinct challenges came out of the coding.
Two things stand out. The top three (tool choice, the correction tax and unproven ROI) account for 63 theme mentions between them. And the owner-only line tells a slightly different story: owners talk most about autonomy breaking at handoffs and not knowing where to start, which are problems of design and priority more than of tools.
Twenty-four discussions wrestle with which platform or partner to commit to. An 8-person digital agency was torn between Gemini and Claude for its core platform. A two-person business didn't want to pay for seats for three employees who don't exist. The research is careful here: the high count doesn't prove that firms want a generic tool-comparison product. The underlying job is fit with their real workflow and budget, which is why our no-code agency playbook starts from the process, not the platform.
“currently trying to nail down which AI platform we should fully adopt into our ecosystem, and we're torn between Gemini and Claude.”
Digital marketing agency, 8 people, source thread
Twenty discussions describe output that needs so much checking that the time saving disappears. It shows up in client emails, analysis, decks and video. One consultant said they were “very sick of spending like half my day every day reviewing barely-disguised Claude output”.
“you end up spending more time 'fixing' the AI’s mistakes than you would have spent just writing the email yourself.”
Agency commenter, source thread
This lines up with the best controlled evidence we have. In the jagged technological frontier experiment with BCG consultants, AI users completed 12.2% more tasks, 25.1% faster and at higher quality on work inside the AI's capability. On a task just outside it, they were 19 percentage points less likely to get the right answer. The service businesses in our sample are discovering that frontier one client email at a time.
Nineteen discussions can't connect AI spend to margin. Usage dashboards show who logs in, not what the business got back, and review time, setup and support rarely get counted. An agency recruiter summed up the other side of the ledger: “Monthly tech spend has gone through the roof.” We wrote about measuring this properly in Cost per Outcome.
“As for impact on margins, we are a small firm and still figuring that out. Like how many hours of consulting time was saved? That is hard to figure out”
Small consulting firm, source thread
Sixteen discussions describe automation that works until something unusual happens. One operator said everything that tried to be a full agent “kinda fell apart the first time something weird happened”. At an insurance agency, AI sorted shared-inbox emails into account records but never told anyone they had arrived, so customers had to call again. The owner of a 37-person European MSP called releasing the technology on customers “scary” because of how little control they had.
“we humans who are supposed to see and reply to the emails, have no idea the emails are there.”
Insurance agency employee, source thread
Fourteen discussions can't pick a first use case and eleven have no one with the capacity to implement. These two travel together. The people who understand AI are usually the busiest experts, and an MSP leader said exactly that: “Our staff that know it are at the limit”. In smaller firms the default owner is the founder, which is the dependency we unpacked in The Founder Tax.
“Considering hiring someone to come in to focus on AI implementation but without an understanding of exactly what they should and should not be doing, even that seems tricky!”
UK services business owner, just under 30 staff, source thread
Fourteen discussions trace bad output back to scattered or stale context. One agency had agents pulling from outdated process docs because a linked database view hadn't refreshed. An implementer described a marketing agency where “the problem wasn't lack of AI. It was that none of their tools talked to each other.” If that sounds familiar, Your Business Context Is Your Biggest AI Asset covers why this matters more than model choice.
Grouping the 20 challenges makes the shape clearer. Most of the friction sits in deciding what to do and in delivering it reliably. People and capacity come next, then trust and control, with upkeep as a smaller but stubborn cost.
The urgency and impact judgements add a second lens. Ten of the 20 challenges were rated high on both, including several with modest counts, such as confidentiality and pilot scope. A challenge doesn't need to be frequent to be expensive when it hits. Pilot scope is a good example, and it fits the argument in our analysis of why AI pilots fail to scale: the trouble usually starts before the pilot, in data, workflow and ownership.
Look past the tools and a cluster of people problems sits underneath: skills that don't match the work, nobody with time to own the rollout, adoption that depends on a few enthusiasts, and a workload that grows as output speeds up. Together they account for 42 theme mentions across the 80 challenge discussions.
At a 28-person agency, a leader wrote: “Team gets "AI fatigue" every time we ship something new.” A senior MSP technician watched new hires “just ask ChatGPT about any issues and blindly trust it”, which is a skills problem and a review problem at once. And one agency employee described the side of speed nobody budgets for:
“It takes him 20 seconds to assign us weeks of work.”
Agency employee, on their CEO's AI-generated ideas, source thread
None of these get fixed by buying a better tool. They need owners, practice time on real deliverables and someone deciding what the extra capacity is for.
Frustration is the dominant tone, recorded in about a third of the usable sources. Uncertainty and pressure follow. The positive accounts are real but outnumbered.
“No internal training, no sandbox, no AI center of excellence. I feel like I'm lagging behind and I'm honestly overwhelmed by the sheer amount of material out there and I don't know where to start.”
Consultant, 5-person boutique firm, source thread
Good research includes the accounts that complicate the story, and this sample has plenty. Some firms solved the problem without AI. Some disclosed openly and found clients didn't mind. Some decided, for sound reasons, that AI isn't for them.
Sources for the table above: PR agency disclosure, in-house agency applications, reporting without AI, agree first, draft second, UK employer on paying for implementation, ethical rejection.
The practical lesson is that "where does AI belong?" is a better question than "how do we adopt more AI?" Sometimes the answer is a narrow draft-and-review workflow. Sometimes it's a plain integration. Occasionally it's nowhere, and a firm that knows why it said no is in a stronger position than one that bought six subscriptions and hopes.
A ranked list of problems is only useful if someone works through the fixes. So starting with this report, I'm opening a new series on the Orbflo blog. Each article takes one of the 20 challenges, unpacks what's really going on using the evidence from this research, and sets out a practical solution a service business can apply. The aim is a working answer for each challenge, tested against the conditions owners told us matter: net time after review, clear ownership and a business case that holds up.
We'll start at the top of the ranking with tool selection and the correction tax, and the order may shift as readers tell us which problems are costing them most. If one of these is stalling your business right now, tell us, and it moves up the list.
Find out which challenges are holding you back
The AI Operating System Scorecard shows where your business sits on workflow, data, ownership and measurement, so you can see which of these 20 challenges to tackle first instead of guessing.
Take the AI Operating System Scorecard →
No, and it isn't designed to be. It's a purposive sample of public Reddit discussions, which carries selection, self-reporting and survivorship bias, and most posters don't disclose company size or location. Its strength is explanation: it shows why firms make the choices they do, in their own words. For prevalence, national data such as the ONS AI in UK businesses analysis is the better source. Our next step is interviews with owners of verified 20 to 250-person UK agencies and consultancies.
It depends on where they are. Before buying, the common blockers are not knowing which tool fits and where to start. After trying, the blockers become the correction tax and ROI that can't be shown. Across both studies, the thread connecting them is capacity: someone with the skills and time to own the workflow. The Harvard and BCG experiment helps explain the after-trying problem, since AI performs well on some tasks and worse than a person on others.
Start with one repeatable workflow and measure it before and after, counting review time. If checking the output takes longer than doing the work, narrow the task or leave it alone. Compare AI against ordinary automation too, since several firms in our sample solved their problem with integrations alone. Our cost per outcome method gives a simple way to run that comparison.
The AI Operating System Scorecard is a diagnostic tool that measures whether your business is structurally built to make AI compound, across nine dimensions including how decisions get made, how clearly your processes are defined and how your team is using and integrating AI.
The output is a clear view of where your biggest leverage gaps are and where to focus first.
One practical AI operating-system insight bi-weekly.
No fluff, no spam.
© 2026 AV Ventures Limited trading as Orbflo | All Rights Reserved
