When an owner finally decides to try AI, the instinct is almost always the same: reach for the most visible thing. A chatbot on the website. An AI "assistant" that customers talk to directly. It is the version of AI you have seen most often, so it is the version you picture when you imagine doing it yourself.
It is also, more often than not, the worst possible place to start.
The projects that actually pay for themselves early are the boring internal ones nobody makes a video about. They do not impress anyone at a dinner party. They just quietly give you time and money back. This post is about how to spot them, and why the flashy customer-facing project tends to be the one that disappoints first.
Where the value actually sits
You do not have to take this on faith. McKinsey's widely-cited study on the economic potential of generative AI, which examined 63 uses across 16 business functions, found that around three-quarters of the value the technology could create concentrates in a handful of areas: customer operations, marketing and sales, and the general knowledge-work of drafting, summarising and writing. The same analysis estimated that today's tools can meaningfully assist with tasks that currently absorb 60 to 70% of the time employees spend working.
Translated for a business like yours, that finding is not about a clever customer gimmick. It is about the routine, language-heavy work happening inside your business every single day: the quoting, the admin, the correspondence, the sorting of information. That is where the value is. It is unglamorous, and it is enormous precisely because it is so ordinary and so constant.
The shape of a good first project
The candidates that pay off early share a recognisable shape. When you are scanning your own business for a first project, look for tasks that are:
- Repetitive and high-volume. Something you do many times a week, so that even small savings compound into real hours.
- Made of words or documents. Drafting, reading, sorting, summarising. Language is exactly what these tools are best at.
- Low-stakes when occasionally wrong. A slightly-off first draft that a person fixes costs you nothing. A wrong answer sent straight to a customer costs you trust.
- Already yours. A task you already do and understand well, so that you can tell good output from bad at a glance.
In practice, at your scale, that points to work like this:
- Turning enquiries into first-draft quotes or proposals.
- Triaging incoming customer emails and drafting routine replies for a person to approve.
- Pulling structured data out of invoices, purchase orders or forms, instead of someone retyping it.
- Getting a plain answer out of your own numbers, such as "which customers slipped last quarter," without building a report by hand.
- Turning a pile of reviews or survey responses into the three themes that actually matter.
Every one of these is internal, bounded, and forgiving. That is not a coincidence. It is the profile of a project that works.
Why the flashy project flops first
The customer-facing chatbot is, in fact, the hardest possible first project, for reasons that have nothing to do with how clever the underlying model is.
It is high-stakes: every output goes straight to a customer, so being confidently wrong is expensive and public. It is data-hungry in a way that catches businesses out: a useful customer assistant has to know your products, your prices and your policies, and if that information is scattered across systems or quietly out of date, the assistant will be too: a model can only ever be as reliable as the information you point it at, and it should never be treated as the final word in the first place. And it is open-ended: customers ask anything, whereas a narrow internal task has edges you control.
Start there, and you are most likely to produce the exact thing that gives AI a bad name: an impressive demo that embarrasses you in front of a paying customer three weeks later.
Boring first, visible later
The counterintuitive move is to earn your confidence on the low-risk internal work first, where a mistake is cheap and the volume is high, and let the customer-facing ambitions wait until your information and your instincts are ready for them.
This is, quietly, what the businesses pulling ahead have done. They did not open with the moonshot. They took one tedious internal task, proved it saved real time against a real baseline, learned to trust the result, and only then took on the next one. Boring, measured, repeated. It is far less exciting than a chatbot launch, and far more likely to still be delivering value a year later.
The hardest part of all this is not technical. It is resisting the shiny option long enough to pick the boring one that actually pays. That single choice, which task to start with, is where most of the value, and most of the wasted money, is quietly decided. It is also exactly the kind of judgement that is easier to get right with AI consulting from someone who has watched it play out across a lot of businesses that look like yours.
References
- McKinsey & Company: The economic potential of generative AI: The next productivity frontier (June 2023): around 75% of the value concentrates in customer operations, marketing and sales, software engineering and R&D; today's tools can assist with tasks absorbing 60 to 70% of working time. https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier
- James Ryseff & Anu Narayanan, RAND Corporation: The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed (2024): most AI projects fail for non-technical reasons, chiefly a failure to pin down the problem being solved. https://www.rand.org/pubs/research_reports/RRA2680-1.html
Written by JP, Sixees Labs. Last reviewed July 2026.