← All Field Notes

Field note 03 · Hyphos Inc. · · 9 minute read

10 Ways Small Businesses Are Actually Using AI Right Now

Five that take work off your week, and five that quietly turn you into a software company. What the second half costs, from someone who did it himself, and how to tell a demo from software you can run a business on.

Two different lists, and most people only notice the first one. The first five take work off your week. The second five quietly turn you into a software company.

Five That Take Work Off The Week

  1. Reading Supplier Invoices

    Pull the amounts, dates and line items off a PDF and code them to the right account, instead of someone retyping them into the accounting system.

  2. Turning A Conversation Into A Quote

    Notes from a site visit become a priced estimate in your own format, with your own margins, ready for you to check rather than write.

  3. Meetings That Produce Actions, Not Transcripts

    A wall of text nobody reads is not useful. What is useful: here are the four things somebody agreed to do, and who owns each one.

  4. Catching What Does Not Match

    Invoices against purchase orders, payments against invoices, hours against schedules. Software is patient about reconciliation in a way people are not.

  5. Finding The Thing You Know You Have Somewhere

    Search that understands "the Henderson quote from last spring" rather than needing the exact filename you cannot remember.

Five That Build Something That Stays

This is the half that has changed in the last two years, and it is the half that rarely comes with a warning. A year ago these needed a developer and a budget. Now they need an afternoon.

  1. Writing The Contract Yourself

    Service agreements, subcontractor terms, a privacy policy for the website, an employment letter. The work that used to mean a lawyer, or a template you did not really read.

  2. Wiring Your Own Systems Together

    Your booking system into your accounting. Your website form into your CRM. Most business software publishes an API precisely so it can be connected. What has changed is who is doing the connecting: an API key, a few lines of code, and what was a quote from a developer last year is running by Friday.

  3. Putting An Assistant In Front Of Customers

    A chatbot on the website that answers questions about your products, your hours, your policies, in your name, when nobody is watching.

  4. Building The Tool You Could Never Justify Paying For

    The job tracker, the customer portal, the dashboard, the little app your crew uses on site. Described in plain English, built in an afternoon, now part of how the business runs.

  5. Letting It Act Instead Of Advise

    Not "draft this". Send it. Book it. Order it. Post it. The moment it stops suggesting and starts doing is the moment the stakes change.

Why Most Of This Disappoints The First Time

Every business owner who has tried this has had the same experience. The demo is astonishing. You bring it into your actual business and it is confidently, fluently wrong, because it does not know your suppliers, your pricing, your customers, or the way you actually do things.

The Real Problem

That is not a smartness problem. It is an information problem.

A model with no access to your business tends to guess rather than say so, and it guesses in complete sentences, which is what makes it dangerous rather than merely useless.

You cannot fix that by asking it to be more careful. It is already trying. The fix is structural: give it your real information, stored so it can be retrieved and checked, and build in the discipline to answer "I don't know" instead of filling the gap.

That is most of the job. The interesting part was never the chatbot. It is the plumbing underneath it.

Half Of That List Is Software Development

Almost nobody sits down and decides to become a software company. You have a problem on a Tuesday, you describe it to an AI, and by Thursday something is running that your business depends on. There was no project, no budget, no decision. It just happened, and it happened because it finally could.

This is not a warning off it. Hyphos Inc exists because we built it this way, after twenty years of putting systems into pipelines and construction projects. That experience is why we know what the list below costs. We met most of it ourselves.

The Thing Nobody Says Out Loud

The building got easy. Knowing what you just took on did not.

So here is the part that usually goes missing. Five things you now own, whether or not anybody told you.

  1. Its Words Are Your Words

    Air Canada put a chatbot on its website. It told a grieving customer he could claim a bereavement fare after flying. He could not. When he took it to a tribunal, the airline argued the chatbot was a separate legal entity responsible for its own actions. The tribunal disagreed, and held the airline responsible for everything on its website, chatbot or not.

    The damages were $650.88. The principle costs considerably more. If it speaks in your name, you said it.

  2. Code That Compiles Is Not Code That Is Safe

    Veracode has been testing this steadily: over 150 models, eighty coding tasks, four languages. As of March this year, AI-written code is correct in form more than 95% of the time, and passes a security test 55% of the time. That number has not moved in two years. In Java it is 29%.

    Their conclusion is worth reading twice: the models have become excellent at writing code that compiles, and have not become good at writing code that is safe. Working and safe look identical from where you are standing.

  3. Your Privacy Obligations Do Not Transfer

    The moment you connect a system, customer information starts moving somewhere it was not before. Under Canadian privacy law you remain accountable for personal information you hand to somebody else to process, and you are expected to have satisfied yourself it is protected comparably over there. The obligation stays with you. It does not move to the vendor, and it certainly does not move to the model.

    Worth knowing before the API key goes in, rather than after.

  4. A Document Can Read Perfectly And Still Not Hold

    The contract, the policy, the terms on your website. They come back polished, properly structured, in the right voice. What reading them will not tell you is whether a clause means what you think where you actually operate, or whether the thing you needed is missing entirely.

    Which is the same failure as the rest of this note. A sentence that is wrong, you can argue with. The sentence that was never written, you only notice if you already knew it should be there.

  5. You Are Now The Only Person Who Understands It

    It works. Nobody else can change it. There are no tests, so nothing tells you when a change breaks something quietly. There is no record of why any of it was built that way, because the reasoning happened in a chat window that has long since been summarised away.

    That is fine right up until you are away, or busy, or it goes wrong in front of a customer.

None of this is an argument for going back to doing it by hand. It is an argument for knowing which half of the list you are on, because the second half comes with obligations the first half does not.

What Separates A Demo From Something You Can Run On

Plenty of what is being built right now is genuinely useful. Almost none of it is built to be wrong in public, which is the test that matters. Your business does not need software that works when everything lines up. It needs software that behaves sensibly on the day an invoice is a photograph of a crumpled page, the customer’s name is spelled three ways, and the person who knew the exception is on holiday.

Four things make the difference. They are not clever, and that is the point.

  1. 01

    It Shows Where Every Answer Came From

    Not "the AI said so". A pointer to the invoice, the paragraph, the line of the transcript. If a number cannot be traced to something you could open and read yourself, treat it as a guess wearing a suit.

    In our own software: In Hyphos, every factual sentence either carries its source or is marked unsourced before you ever see it. In the funeral home software, a suggested form value is thrown away unless the exact words appear in the recorded conversation it claims to come from.

  2. 02

    Something Checks The Work That Cannot Be Talked Round

    A model can be persuaded. A test cannot. The check has to be ordinary code that either passes or fails, run every time, including on the AI’s own output.

    In our own software: Hyphos runs a second model over each cited sentence to ask whether the source actually supports it, then walks the relationships in the data to confirm claims about how things connect. Our own platform carries 22,561 automated checks. The funeral home software has 654.

  3. 03

    A Person Approves Anything That Leaves The Building

    Approval should be built into the data, not a habit somebody remembers. If it is a habit, it eventually gets skipped on the busy day, which is the day it matters.

    In our own software: In the funeral home software, AI-written content is created in a pending state. It can only move through approved and scheduled to published by a named person with the right role, and every one of those moves is recorded. A change to the public website needs the owner, signed in with two-factor. Nothing publishes itself.

  4. 04

    It Says When It Does Not Know

    The most dangerous answer is a confident one built on nothing. A system worth trusting will tell you what it searched, what it could not find, and where it is unsure.

    In our own software: Hyphos reports how much it actually read: so many of so many records, and what it deliberately excluded. Work that has not been verified is labelled unverified rather than quietly counted as done.

The Part Nobody Puts In The Demo

Our own platform is built by a system we wrote called Domas. AI plans each change, builds it in a separate copy, away from anything live, tests it, and a second AI reviews the work. A person approves every plan before any of it starts.

The morning this article was written, that system refused a change we tried to make to it, because the session had not yet proved it had read the records the change depended on. Not a warning. A refusal, twice, with the reason stated plainly.

That is inconvenient. It is also the entire point. A gate that never blocks you is decoration, and by the time you find out which kind you have, the wrong thing has already gone out to a customer.

Five Questions To Ask Anyone Selling You AI

Including us. And if you built the thing yourself, ask them of your own work, which is harder and more useful. You do not need to be technical, and the answers tell you nearly everything.

  1. Where did that number come from?

    Ask it about your own business, then ask it to show you the source. You should land on a document, a record, a line you can read. "It analysed your data" is not an answer.

  2. What happens when it is wrong?

    Not if. When. Ask who catches it, and whether that catching is a person remembering or something that runs automatically every time.

  3. What can it do without asking me?

    Send an email? Post publicly? Move money? Change a price? Ask for the list, and ask to see where that list is enforced.

  4. Show me it saying "I don’t know".

    Ask a question your data cannot answer. If it produces a confident, plausible answer anyway, you have learned the most important thing about it.

  5. What happens to my data, and where does it live?

    Whose systems, mixed with whose. Ask how your information is kept separate from another client’s, and get the answer in terms of how it is built, not a promise.

A good supplier will enjoy these questions, because the honest answers are the reason their thing costs more than a subscription. A bad one will tell you the model is very advanced.

What We've Built

Connected business software for Cornerstone Funeral Home. The system the business runs on: records and arrangements, estimates, documents, scheduling, accounting integration, the public website and marketing reporting, on one database, so one business stops retyping itself into six different tools. AI reads supplier invoices and drafts marketing, and every piece of it waits for a person to approve.

Wander. A production mobile app we built for a client from one codebase for iPhone and Android, live on the App Store, with AI-powered natural-language search.

Hyphos. Our own platform, and the deep end of this: the layer that gives AI the correct information from your business, with a way to show where every answer came from, and the discipline to admit what it does not know.

If You Have Already Built Something

This is the conversation we find most useful, and the one few people are offering. You connected the systems, or stood up the assistant, or had it write the agreement, and it works. The question is what you took on while you were doing it, and that is a much shorter conversation than people expect.

Forty five minutes. We go through what you have built, what it can do without asking you, where your customers’ information is going, and what happens on the day it is wrong. You get a one page write up at the end whether or not we work together, and if the honest answer is that you are fine, that is what it will say.

And if you have not built anything yet, the same session works in reverse: where AI would realistically save you time, and which of it is worth doing properly.

Ask about a 45-minute session or see the work.

Read next: When Your Website Becomes Part Of Your Business →