Artificial Intelligence is becoming increasingly available to a growing number of organizations. For a small business, the decision to adopt AI systems is not driven by the technology’s perceived popularity. It is not an enhancement of the digital workplace. Small business owners have limited resources; therefore, they cannot risk adopting an unreliable and unproven technology. A company’s needs must dictate the AI’s potential use, and not the other way around.
The rise and growth in investment in the Artificial Intelligence Market is a result of the growing investment in areas such as machine learning, generative AI, natural language processing and computer vision. At the same time, OECD research shows that AI adoption among small and medium-sized enterprises is lower than it is among larger businesses, with factors such as skills, data availability, financial resources and the suitability of the technology influencing adoption choices.
For a small business, such a distinction is important. The mere fact that a technology is becoming more capable does not mean that it should be adopted by every organisation. An AI application can be beneficial when it addresses a clearly defined problem, enhances an existing process, or aids employees in using information in a way that they would not otherwise do. Without a clear purpose, however, even an impressive technology can become an expensive burden.
Begin With the Problem You Want to Solve
That shift in thinking can ease the burden of a decision. Almost all businesses have processes that eat up time without adding value. An employee might spend an hour dealing with a similar enquiry, copying data from one system to another, reviewing documents, compiling routine reports, or working through back office records.
It’s worth exploring, as you could identify wasted time without changing your business too dramatically. The best place to begin is to ask the employees engaged in these tasks daily – they will know where the bottlenecks are as they experience them. Processes that appear to run quite smoothly from the management level may actually contain several complex manual steps.
It is also helpful to differentiate between occasional inconvenience and a genuine business issue. Spending ten minutes on a single task each month might not merit time spent on automation. Spending several hours every week duplicated by five employees might.
Look Closely at Everyday Workflows
Work out what workflow you want to improve before you choose an AI application.
Let’s say you run a small accountancy practice and prepare reports for customers on a regular basis. You might need to gather the relevant information from various sources, check the numbers, copy the data into a spreadsheet, work out the common metrics, write the report, and review the finished item.
On the surface, this sounds like a perfect candidate for generative AI to do the writing. Dig deeper, and you might find that the most significant issue is the copying and pasting of data, meaning you get more value by fixing that problem rather than adding generative AI.
That’s the reason for creating workflows: so you don’t apply the most advanced technology to the part of a process that doesn’t actually need it. It may also turn out that different parts of a process will benefit from different solutions: one may require traditional automation, one generative AI, and another a human. The best solution doesn’t have to involve AI at every stage.
Understand Where AI Can Actually Help
AI is most likely to be useful where a task involves: detecting patterns; reading language; analysing large amounts of data; forecasting or content generation for an individual to check. Document processing, such as payroll, applications, forms and invoices, is one example. A business that already processes a large volume of documents might be able to deploy AI to pre-process, identify and categorise data for a human to review.
Customer service applications are another example. A business receiving many similar queries about opening times, delivery methods, products and processes might be able to deploy an AI system to find them for the customer service rep to process, or even prepare an initial response, removing some of the drudgery from some of the most routine work. In the same way, a business might deploy AI to help employees find and summarise information from many documents and policies that have been accumulated over time, as the organisation grows, provided that the material is correct and well-managed.
Forecasting demand and planning supplies, inventory, customer behaviour and workflows could all be valuable applications of AI, if the organisation has a good range of historical information to test with. But it’s generally more complex and requires more preparatory work than productivity applications.
Do Not Assume Every Repetitive Task Needs AI
A repetitive task is often claimed to be the perfect candidate for AI, and that can be a trap. There are some processes that are better suited for traditional automation.
Say, for instance, that a company wants a task to happen every time a certain event happens and that nobody needs to actually do anything. There may not be an AI system necessary; a rules-based workflow can perform the task more reliably and simply.
The guidance is simple: if you can establish the rules by which something operates – the ‘if this, then that’ conditionality of most AI systems – traditional automation is likely to suffice. When processes require prediction, inference, natural language processing or any other element of human cognition, then AI is going to be more useful.
There might be a synergy between systems, too. A traditional automation may be able to gather a great deal of information, which an AI system then makes sense of, before a member of staff checks through the report or the results. This could be more feasible than creating an entire workflow with AI.
Consider the Potential Return Before Spending Money
Once you’ve decided that you have a potential use case, then the next question is whether it is worth implementing. You may not need a complex financial model to answer this question. An educated approximation can be just as effective.
For example: say employees spend 15 hours a week collectively creating a certain type of document, and a new system you are considering would be able to cut that time by one third.
You can estimate the savings over the course of the year and compare that amount with the cost of the software, training, implementation, ongoing maintenance, and supervision.
Cost savings are not the only reasons you may want to implement a certain technology. For one thing, a promising use case may also help reduce errors, improve response times, increase the number of employees available to do tasks, or enable a business to accommodate more work without raising administrative costs at the same ratio. Even a modest cost saving may be accompanied by other benefits such as better consistency. In some cases, the key benefit may be a very low failure rate rather than low cost.
Conversely, a technology might seem to be very inexpensive but end up needing significant integration work, staff time, and ongoing monitoring. These factors should be taken into account in the calculation. The aim is not to generate a rather precise forecast. It’s to determine whether the proposed use case is worth pursuing.
Check Whether the Business Has the Right Data
Data is one of the most important considerations in implementing AI, but one that’s often overlooked in early conversations. An AI system can only be effective if it has access to the right data. If that data is lacking, inconsistent, out of date,e or simply not readily available, then the system will likely not be up to the job – regardless of how advanced the underlying AI model might be.
An AI forecast system for an example business might find that its previous sales data has been stored in different formats across different locations. That’ll have to be cleaned and homogenised before a forecast system can work.
Perhaps one of the most important takeaways is that an AI solution might not be the right place to start – or that it may be more appropriate to invest first in better quality data systems.
Think About Risk Before Testing the Technology
Not all applications of AI are the same. This AI assistant that helps an employee summarise an internal meeting is quite different from an application making decisions about employment, financial eligibility, healthcare, or other decisions which could significantly impact individuals. The US National Institute of Standards and Technology’s AI Risk Management Framework recommends organisations consider reliability, safety, security, transparency, privacy and fairness when managing AI risks.
This doesn’t necessarily mean complicated governance structures for small businesses, but it does mean identifying where the potential impact may mean heightened risk and oversight.
If an incorrect AI output may result in a serious financial, legal or personal consequence, human review should arguably always be included, and there should be processes and guidelines in place to redress errors and to hand over to humans when the system is unsure.
Be Careful With Confidential Information
Have a Privacy and Confidentiality policy in place before employees start adding business data to external AI too.ls Small business data can contain everything from customers’ personal data to employee records and contracts, and internal business data, contracts, and financial records.
There is value for businesses to consider how input information is used and stored by external AI tools before employees start plugging business data into them. Businesses should review questions regarding the duration of storage of input data, access, and security, where it is processed, and whether it will be used for model improvement purposes.
This is relevant as employees will often use popular consumer-facing AI tools on their own when they believe doing so will be more efficient. Internal policies on what information can be entered into external AI tools can be helpful for employees to refer to.
Give Employees a Role in Choosing the Use Case
It is easier to develop an AI solution that works if the staff who are affected by the system play a part in designing the solution. Those who work with a process or data every day will spot issues more easily than people watching from afar, and will be able to spot exceptions that might not seem obvious during the first evaluation. Involving staff in this way can improve the quality of a use case and make implementation more straightforward, as staff will understand why a change is being made.
This does not necessarily mean training every staff member to become an AI developer, but staff do need to understand how a new system may affect their job, and when they should challenge its output.
They should be trained in practical skills, such as how to check AI output, when to challenge this output, and how to deal with sensitive information. The OECD identified skills and workforce readiness as key to SME adoption of generative AI.
Test One Specific Use Case First
A controlled pilot will often be more revealing than a wide deployment once a potentially valuable use case has been identified.
Instead of trying to implement AI across an entire department, a business can select one narrowly defined task and apply the technology to it. This might be, for example, using AI to help with one specific type of customer query. Before the trial starts, the business might measure the average time taken to respond to each query, the amount of employee effort required, and the proportion of queries that need to be escalated to another team.
By comparing the pilot to those benchmarks, a business can evaluate whether the technology saves time, the outputs are accurate enough, the employees don’t end up spending more time correcting mistakes than saving, customers are unaware of any difference,e and any privacy and security issues are acceptable.
The point of a pilot should not be to show that AI can do a task; it should be to determine whether the AI does it well enough.
Measure Quality as Well as Productivity
Speed is a weak indicator of usefulness. Generating an answer quickly isn’t particularly useful if your employee then needs to spend minutes correcting it. Businesses should try to measure quality in parallel with productivity. Depending on the task, this could include accuracy, error rates, correction time, customer experience,nce or the proportion of cases that require human input.
Considerations around generative AI: Generative AI can result in a convincing-sounding response that is actually wrong, so employees should be given the right opportunity to review some of these outputs before they’re relied upon.
The correct approach will depend on the task, but the concept remains the same: speed is not everything.
Recognise When the Right Decision Is Not to Use AI
A good AI plan will acknowledge where to draw the line, and say no. A use case may not be worth the cost, either because the gains would be negligible, the data needed for it does not exist, the risks are too high, or it could be replaced by a simpler solution.
Perhaps the use-case is too rare to pay off in savings. Or the business does not have enough technology to get it going. Sometimes, better foundational systems are the starting point for future gains.
A study by the OECD found one of the most common reasons SMEs gave for not doing generative AI was “this type of AI is not appropriate for the work they do”.
That is key for businesses. Say no and stick with less complexity. If the AI is not going to be solving a big problem, then steering clear may be the best way forward.
Review the Decision as Technology and Needs Change
Treat AI implementation as an ongoing commitment. AI use cases that don’t seem viable now may suddenly be useful when the prices of technology change or processes evolve in the business. The reverse is also true: a process that has performed well in an initial pilot may need rethinking when a business scales up or changes strategy.
Consistent review of an application may reveal that it is no longer adding value. For example, a business may want to check that staff are still using it properly, that the quality of output has not changed, that the data source is stillfit-for-purpose, or that new risks have emerged.
The NIST framework supports a lifecycle approach to managing the risks of AI by considering how AI should be governed, mapped, measured and managed, instead of a one-off project to get AI up-and-running. For small businesses, that is even more crucial as running costs have to be weighed against other urgent business priorities.
A Practical Way to Approach AI Investment
The most logical path for any small business should be to view AI as a potential component of a wider process of business improvement. Rather than seeking ways to apply a specific technology, first identify the problem. Focus on the work that is being done, and where the costs of that work are highest, then explore whether AI, traditional automation, better software,e or simply a straightforward process change would be the most suitable solution.
Where AI looks promising, consider the potential benefit and how much data exists, and is of sufficient quality to use, what the implications are in terms of privacy and security, whether staff are ready to work alongside the technology, and what operational risk is involved.
A small pilot project will then provide evidence that it makes sense to go further, or that a different solution would be preferable.
Looking Ahead
AI will become more accessible for smaller organisations in the near future. Researchers from the OECD identified AI adoption by SMEs using it for business tasks at 6% of respondents in 2020, with the differences across sectors and countries being significant. They also recognise that variations in digital maturity, skills, data and financial capacity are factors that influence adoption. As the tools get easier to access, the difficulties will be about selection rather than simply the supply of AI-enabled products.
There will be many more tools, many more applications and many more reasons to see how we can use AI in our work. This makes a systematic assessment process all the more important.
Organisations that approach AI with well-defined needs, realistic expectations and tangible performance goals can better position themselves to assess where it has genuine utility. This enables them to pilot, hypothesize, and experiment less expensively, learn lessons and steer clear of putting technology in merely because of other people’s excitement.10 In the end, selecting the optimal AI use case is less about hi-tech appropriateness than about operational need, data fit, risk and value for the business.