A rival answers enquiries before you have finished the day's jobs. A platform offers to run your marketing, bookings and customer service, provided everything stays inside its system. Your administrator wants training, but there is nobody to cover their work.
These are more useful starting points for the AI inequality debate than predictions about everyone becoming unemployable. Small businesses have something to gain from AI, and something to lose if the gains depend on resources they cannot afford or platforms they cannot leave.
The sensible response is to take the risk seriously without treating anyone's future as settled.
What does ‘permanent underclass’ mean?
In this debate, the phrase describes a feared situation in which some people become persistently excluded from economic opportunity: decent work, income growth, ownership and a realistic route to improving their circumstances.
It is a contested, harsh phrase. It does not mean that those people are inferior, less deserving or incapable of learning. Nor is it an established prediction that AI will create such a group. “Permanent” is the claim we should question most carefully.
The concern is that those who own valuable technology and businesses could capture much of the benefit, while people whose work changes struggle to find another route to security. For small firms, there is a related possibility: remaining open but becoming increasingly dependent on bigger companies for software, customers and permission to compete.
Buying an AI subscription does not automatically solve either problem.
What the small-business evidence says
The OECD's report, published on 5 November 2025, draws on a late-2024 survey of more than 5,000 small and medium-sized enterprises in Austria, Canada, Germany, Ireland, Japan, Korea and the United Kingdom. Across that seven-country survey, 31% of SMEs used generative AI, meaning tools that create material such as text or images.1
Among the SMEs using it, 65% reported improved employee performance. That means 65% of users reported a benefit; it does not mean their employees became 65% more productive.1
There is an encouraging competitive signal too: in the same survey, 29% of AI-using SMEs said it helped them compete with larger companies, while 26% reported increased revenue.1 Those findings support testing useful applications. They do not establish that adopting AI caused business growth, or that firms without it will fail.
The staffing evidence also deserves space. Among AI-using SMEs in those seven countries, 83.0% reported no effect on overall staff needs, 9.1% a decrease and 5.5% an increase. The remaining responses were “not applicable” or “don't know”. The OECD did not ask the size of the changes, so these figures cannot establish a net number of jobs gained or lost.1
These are owners' or managers' reports about their businesses at a particular time, not an experiment or a forecast of employment in 2030. They cover seven countries, not every small business worldwide. They also concern generative AI, rather than every form of automation.
A separate 2025 International Labour Organization assessment estimates that one in four workers worldwide is in an occupation with some degree of generative-AI exposure. Exposure means tasks could be affected, not that one in four workers will lose their jobs; the ILO expects transformation to be more common than redundancy.2
Why worry if the tools are becoming cheaper?
A low subscription price removes one obstacle. It does not create spare cash, organised records, paid learning time or access to customers.
Larger businesses can spread the cost of specialists, testing and failed experiments across more sales. A small employer may need the same person to answer the phone, chase invoices and learn the new system. Asking them to “get better at AI” in their own time shifts the cost onto someone with little room to absorb it.
The OECD survey illustrates the skills barrier: 50% of non-adopting SMEs across the seven surveyed countries cited employees lacking the necessary skills. But 57% cited AI being unsuitable for their work.1 Some businesses need help getting started; others have sensible reasons to wait. Non-adoption is not evidence of incompetence.
Ownership and market access matter too. A company that controls customer discovery, advertising or bookings can set terms that affect a small firm's margins. If it also supplies the AI running those activities, switching away may become difficult. Customer records, templates and working knowledge can accumulate in a system the business rents rather than controls.
That creates a practical test: if the supplier raises prices, changes its rules or closes your account, can you still serve your customers?
Workers and customers can also lose out even when the firm benefits. An administrator may lose the tasks through which they learned the business, without gaining training for more demanding work. A customer who struggles with digital services may face a chatbot where a helpful phone call used to be. Faster administration should not quietly remove either person's route to participation.
A realistic pilot: quote administration
Consider a hypothetical local maintenance firm. Its owner wants help turning enquiry notes into draft quotes. This is an illustration, not a customer case study or a promised return.
Run a four-week pilot on one familiar service. First record how the existing process performs. Then let an AI tool prepare a draft from an approved service description and price list, with missing details clearly flagged. Keep unusual jobs outside the trial.
A named employee checks the scope, price, exclusions and proposed appointment before anything reaches the customer. The tool cannot send quotes, promise availability, take payment or alter the price list. Calculations should come from the firm's checked pricing process rather than a model's guess.
Start with fictional or properly anonymised enquiries. Before using real customer information, check the provider's retention, training-use, access and deletion terms, along with the firm's data-protection obligations. Do not paste addresses, access codes or sensitive circumstances into an unapproved account. Keep the approved quote and customer record in the business's own system.
Agree the measures before the trial:
- Time: record total staff minutes per completed quote, including preparation, checking, corrections and rework. Track customer waiting time separately.
- Quality: count incorrect prices, omitted exclusions and other errors across all trial quotes. Record serious errors individually, even if the average looks good.
- Cost: include subscriptions, setup, training, integrations and staff time at a stated hourly cost. Compare cost per approved quote with the baseline; do not count the same saved time twice.
- Customer experience: record complaints, requests for clarification and whether people can still reach a person. Quote acceptance is useful context, but changing demand or job mix can affect it.
Choose a spending cap and stopping rules with the team. Pause for a privacy incident or a serious pricing error. Continue only if measured results meet the agreed quality standard and justify the full cost. Time saved is useful capacity, not automatically cash saved or extra revenue.
Share the gains and keep an exit
Involve the administrator in choosing the workflow and judging the results. Give them paid learning time and a say in where any released hours go: following up customers, resolving awkward enquiries or improving records. Discuss redeployment before quietly allowing automation to hollow out the role.
Preserve work through which junior colleagues learn judgement. Someone must understand why a quote is wrong, and that knowledge needs practice. Keep a phone or email route for customers who cannot, or do not want to, use an automated service.
Before committing to a supplier, test an export of customer records and templates in a usable format. Check what cannot be exported, what deletion involves and how the process would run during an outage. Retain direct customer relationships and a documented manual fallback.
There is a public-policy job here as well. Affordable training needs paid time and practical support to be usable by small employers. Business groups can organise shared learning and independent advice. Governments should address barriers to finance and training, support people changing work, and scrutinise market arrangements that make customers or suppliers difficult to leave. Individual effort cannot fix every structural disadvantage.
For your next team meeting, choose one repetitive task, name the person responsible for checking it and agree what evidence would justify a trial. Include a worker's learning opportunity and the customer's route to a human in that decision, alongside the cost.
Sources
[1] https://www.oecd.org/en/publications/generative-ai-and-the-sme-workforce_2d08b99d-en.html [2] https://www.ilo.org/publications/generative-ai-and-jobs-2025-update