Many AI projects start when someone sees an impressive demonstration and asks where the technology could be used. This is usually where the conversation moves towards products before anyone has established which business problem needs attention or what a worthwhile result would look like.
“The model is the easy part. The system you build around it, your data, your rules, your verification, your owner, is what makes it work or fail.”
For AI to become a reliable business system, it needs to work from your facts, not the internet’s, and operate within your systems. A person must check the output before anything counts, with ownership assigned to a name, not a committee.
AI is probably already inside the business
Employees are already using readily available AI tools to summarise email threads and interrogate lengthy documents, often without an approved platform or formal record.
Client emails and contracts may be pasted into consumer tools, along with sensitive operational information. This leaves the business with data it no longer controls, creating substantial risk, while providing no audit trail and little visibility of who is relying on what.
“Nobody reaches for these tools at random. They reach for them at the exact points where the existing process is slow or painful. It is initiative, not sabotage. The tool works and it is one tab away.”
So what should businesses be doing? Banning the tools has the potential to drive their use out of sight without addressing the underlying need which is that organisations need a sanctioned path that is easier to use than the shadow one, supported by clear guidance about which information may leave the business.
Start with the cost of the problem
An impressive AI demonstration often prompts a business to ask, “Where could we use this?” A useful starting point is to identify a painful, high-volume process that is already costing the business time or money.
The right first use case will often involve repetitive work where the inputs already exist and the output is easy to check. If a competent person can tell within a minute if the answer is right or wrong, the organisation can run the system safely with a human in the loop. If the output cannot be checked easily, the business is not ready to start there.
This immediately rules out tasks where an incorrect result would be expensive and difficult to detect. Generative AI can make a weak argument sound complete, allowing the polish of the language to outrun the quality of the thought. An answer may also be correct in general but wrong for that particular organisation.
Understand the work before building
Espresso is applying this approach with a manufacturer whose estimating process has become a bottleneck. Each order begins with a design drawing, which an experienced estimator uses to determine what must be manufactured and calculate the price.
Producing an estimate can take weeks and depends on knowledge held by a small group of employees. Different estimators may also calculate different prices for the same job.
Espresso began by reading the historical estimates and studying the actual workflow. The team needed to understand what information gets used, where human judgement enters and what correct means for this business.
A focused AI agent is now being developed to interpret a drawing and produce a draft estimate using the manufacturer’s own data. An experienced estimator reviews every output, while the system cannot send a quotation or make the final decision.
The team will track how quickly each estimate is produced and how closely the AI-generated draft matches the approved version. This will show if the system understands the business or is simply producing wrong answers quickly.
Make the first commitment specific
Espresso recommends one workflow, one team, one quarter, and one number agreed in advance. This makes the project small enough to fail safely and real enough for people to notice when it works.
Once the problem is understood, the technology decision becomes much clearer. The solution may involve an enterprise product, Espresso’s LAIQ or CREMA platforms or a bespoke build. Models will continue to change, while the system built around the chosen model is what the business actually keeps.
A board or AI committee should leave its first discussion with a one-page AI policy and one named pilot. Before buying or building anything, the business also needs to know “where its people are already using AI without permission. That’s your requirements list, written for free.”