The arbitrage is adoption, not technology
The tools are widely available and cheap. The scarce thing is a business that knows which task to apply them to, can integrate the result into an existing workflow and will stand behind the output. That is why the durable opportunities look like services and operations improvements rather than thin wrappers around a model everyone can access.
- Access to a model is not an advantage; applied judgement about a specific workflow is.
- Opportunities that require no domain knowledge are the first to be competed away.
- Distribution — knowing the buyers — is frequently the real asset.
Where margin persists
Margin tends to survive where the outcome is measurable, the buyer's cost of failure is high, or the work requires context that is expensive to acquire. It erodes fastest where the output is generic, the buyer can replicate it in an afternoon, or price is the only differentiator. Before building, ask what still needs a human to be accountable — that is usually where the fee lives.
- Measurable outcomes support pricing on value rather than hours.
- Regulated, sensitive or high-consequence work resists commoditisation longest.
- Recurring operational work beats one-off projects for stability.
Testing before you spend
The cheapest test is to sell the outcome manually first. Do the work by hand for a handful of paying customers, and only automate the parts you have proven someone pays for. This exposes the real cost of delivery, the objections, and whether the buyer you imagined actually exists — all before software, branding or a company structure.
- Find the buyer before building the product.
- Deliver manually to learn the true unit economics.
- Automate the repeated parts only after the offer is proven.
The obligations that come with it
AI-enabled services carry duties: accuracy of output, client confidentiality, disclosure of how work is produced, and compliance where a field is regulated. Getting these wrong converts a promising service into a liability. Treat data handling and disclosure as part of the offer, not as an afterthought.
- Do not feed confidential client data into tools whose terms you have not read.
- Be clear with clients about how the work is produced and reviewed.
- Regulated fields have rules about who can advise; a tool does not change them.