AI
Most small businesses have already bought AI. Far fewer have one that changed a working day.
The difference is almost never the model: everyone has the same ones. It is whether the thing was wired into the work: the inbox it should be reading, the calendar it should book into, and the person it should stop and ask.
Why most of it fails
The problem is not that small businesses have not adopted AI. They have. It is that adoption and results came apart.
MIT’s Project NANDA studied generative AI pilots and found roughly 95% produced no measurable return. The failures it describes are not technical ones. The models worked. What was missing was a definition of done, data the system could actually reach, and someone who owned it when the output drifted.
That is not an argument against AI. It is an argument about what you are buying, because a subscription is not an outcome, and a demo running on sample data is not an integration.
Source: MIT Project NANDA, 2025, as reported by Fortune and Yahoo Finance. We publish other people’s numbers only where we cannot yet publish our own.
The model is one layer of five.
Whatever you build runs the same five steps: your data, then retrieval to find the part that matters, then the model, then a guardrail deciding whether the answer is good enough to use, then the action it actually takes: the reply sent, the record written, the slot booked.
Four of those are yours. The model is the one you could swap this afternoon for a competitor’s and barely notice. That is why “which AI do you use” is a less useful question than it sounds, and why a tool that only does the middle layer leaves you holding the other four.
What we build
The useful part is where it stops.
An enquiry arrives, gets sorted, matched against what you have already answered before, drafted, checked, and sent. Five of those six run without anyone watching. The sixth is the point: when the check is not confident, the reply waits for a person instead of going out anyway.
Six things, in the order most businesses want them. Each one lists what it will not do, because that is the part you find out later otherwise.
First-response assistant
Answers the website chat and the enquiry that lands at 9pm. Books it, qualifies it, or hands a person the whole thread when judgement is needed.
Where it stops
Only as good as the material behind it. Where your own team answers a question three different ways, so will this.
Missed-enquiry recovery
An enquiry arriving out of hours gets acknowledged straight away, sorted, and queued with a draft reply already written for whoever picks it up in the morning.
Where it stops
Needs somewhere real to write into: an inbox or CRM you already run, not a new system to log into.
Inbox and document triage
Classifies what arrives, pulls the fields out of invoices and forms, and routes each one to whoever owns it.
Where it stops
Anything financial gets a review step. We will not release a payment off a parsed number.
Knowledge assistant
Answers questions from your own documents and cites the one it used, so an answer can be checked rather than taken on trust.
Where it stops
Depends on your documents being findable and current. Often that is the first job, and the assistant is the second.
Content operations
Drafts against a brief and your existing style, into a queue someone approves before anything goes out.
Where it stops
A drafting tool. It does not publish, and it does not replace whoever signs the work off.
AI readiness review
Fixed scope, fixed fee. Where AI would pay back in your business, where it would not, and what each one would cost to run.
Where it stops
The usual finding is that one or two workflows are worth automating and the rest are not. You get that in writing either way.
What we do not do
Four things we turn down.
- Training or fine-tuning models. The base models are not your bottleneck.
- Anything making a regulated medical or lending decision.
- Strategy decks with no build attached.
- Agents that take an irreversible action, sending money, deleting records, signing something, without a person in the loop.
If what you need is on that list you will hear it on the first call, rather than three weeks into an invoice. Some of it we can point you at someone else for, see how we work.
Before you ask anyone
Eight questions, and one of the answers is that you should not spend money on this.
It runs in your browser, nothing is sent anywhere, and there is no email box in front of it. If your work does not repeat, it will say so.
Take the diagnosticTell us what is broken. We will tell you what it takes to fix.
No pitch deck and no discovery retainer to find out. Describe the situation and you will get a straight answer about scope, timing and cost, including when the honest answer is that you do not need us.