A race car gives a driver extraordinary capability. It can accelerate harder, respond faster and achieve things an ordinary car cannot. But someone still needs to choose the line, read the conditions and know when to brake.
That is how I think about AI in a business.
My starting point is the capability it can put into the hands of the people already doing the work. Give a capable person better tools and they may be able to investigate more thoroughly, respond sooner and spend more time on the decisions that matter.
The investment needs to include the driver.
Start with the work your people could do better
In an SME, one person can carry several responsibilities. An account manager may look after customers, prepare proposals, update records and chase internal answers. An operations manager may plan delivery while assembling reports and dealing with exceptions.
Those roles combine judgement with preparation and administration. They are worth looking at task by task.
Where does someone spend time finding information before they can use their expertise? Which documents do they repeatedly assemble? What useful analysis gets postponed because preparing the information takes too long?
These questions give AI a practical purpose: helping people get to the important part of their work with better information and less avoidable effort.
Experience helps you judge the output
An experienced estimator can recognise when a proposed price overlooks a delivery constraint. A bid manager can spot a persuasive paragraph that fails to answer the buyer's question. An account manager can judge whether a technically accurate response will help a particular customer.
That knowledge belongs in the design of an AI-assisted process.
A tool might produce a useful first draft, extract details or suggest questions. The person needs enough context to assess the result, access to the underlying information and authority to change or reject it.
There is a wider distinction here between automating tasks and replacing whole jobs. The ILO's 2025 research on generative AI and jobs identifies job transformation as the more likely overall effect, because most occupations still contain tasks requiring human input. That is an assessment of potential exposure, rather than a guarantee about what any employer will do.
My principle is a choice about implementation: start by improving what your team can accomplish. Be honest that roles and tasks may change, and involve people in shaping that change.
What augmentation looks like in practice
Consider an illustrative tender workflow. A bid manager receives a specification and needs to establish the requirements, find supporting evidence, coordinate contributors and prepare a response.
An AI tool could help extract the requirements into a draft checklist, suggest relevant material from an approved evidence library and prepare a response structure. The bid manager checks the extraction against the original documents, confirms that the evidence fits and decides where input is missing.
Operations confirms what the business can deliver. The commercial owner approves pricing and commitments. The bid manager reviews the submission against the buyer's requirements.
The potential gain is more room for useful work: testing the delivery approach, improving the evidence and challenging weak answers. Establish whether that gain actually exists by measuring the whole process, including checking and correction.
The same principle can apply elsewhere. A customer-service colleague could start with a draft based on approved information. A manager could use AI to explore questions about a report, then verify the findings against its data. The task and the quality of the output determine whether the assistance is worthwhile.
Keeping a person involved has to mean something
Putting an approval button at the end of a process does not automatically create effective oversight. The reviewer needs time, relevant knowledge and a clear understanding of what they are checking.
Define the boundary. What may the tool prepare? Which actions may happen automatically? What requires approval? What should stop when information is missing or contradictory?
Make it possible to inspect the source material and correct the output. Assign someone responsibility for exceptions. For consequential decisions, design a review that tests the substance of the recommendation.
The race car metaphor has a limit: an AI system can produce plausible material that is wrong. Its behaviour needs testing against the actual task. A confident-looking answer is not evidence that the process is working.
Give people time to learn the controls
Buying access to a tool is only the beginning. People need practical examples tied to their jobs, clarity about which information they may use and a way to raise problems without feeling they have failed.
Work with the people who understand the exceptions. They can tell you why the apparently simple request becomes complicated, which records cannot be trusted and where a customer needs a conversation.
Give less experienced staff opportunities to build that understanding too. If they only receive completed answers, they may struggle to develop the judgement needed to challenge them. Use worked examples, source checks and discussion of corrections as part of learning.
Training should leave someone better able to do the job and explain their decisions. Knowing how to generate an answer is one part of that capability.
Decide what the released capacity is for
If a process becomes quicker, give the improvement a destination. It could mean faster customer responses, fewer late evenings, a backlog cleared, more thorough reviews or capacity to serve additional customers.
Measure those outcomes alongside quality, rework and staff experience. Include subscriptions, setup, maintenance and review effort in the assessment. A quick draft followed by extensive correction may offer little benefit.
Be clear with the team about the purpose of the trial and how its results will be used. An augmentation programme needs that intent reflected in the measures and management decisions around it.
Start with one capable person and one real task
Choose a recurring task with a clear output. Work through it with someone who knows it well. Agree what the tool will assist with, what the person will decide and what an acceptable result looks like.
Try representative examples, including awkward ones. Record the total effort and the corrections needed. Expand only when the process produces a useful, repeatable improvement.
A race car still needs a driver. In a business, the opportunity is to equip that driver properly: better information, more useful capability and enough control to make sound decisions.
If you want to explore where AI could support your team, Novologix offers practical AI support for business leaders. We can help you identify a relevant use, understand its limits and define a sensible trial, with implementation scoped separately where needed.