What we mean by consulting
Most AI consulting for small business is a paid workshop, a slide deck of opportunities, and a recommendation to come back for implementation. We do not do that. We build assistants and automations for a fixed price, and the thinking that goes before a build we give away as a free fit check: one task, one page, two working days, a price or an honest no.
This page is the method behind that page. It is what we work through when your request arrives, written so you can run it yourself first. If you run it and conclude nothing is worth automating yet, you have saved both of us a week.
Step one: list the candidate tasks
Write down every task in your business that someone does more than once a week and that follows roughly the same steps each time. Do not filter yet. Typical lists from stores and service businesses include answering order-status emails, processing returns, chasing suppliers, writing product descriptions, checking invoices, routing enquiries, producing the Monday report, updating stock from supplier spreadsheets, and answering the same twelve questions on the website.
Ask the people who do the work, not only the owner. The owner knows what is expensive; the staff know what is tedious, and tedious is where automation holds up.
Step two: score each task
For each task, estimate three numbers and multiply them:
- Frequency: how many times a month it happens. 600 order-status emails. 4 Monday reports. 150 invoices.
- Time: minutes per occurrence, honestly, including finding the information and switching between tabs. An order-status reply that "takes a minute" usually takes four.
- Error cost: what it costs when it is done wrong, in dollars. A wrong return decision might cost a shipping label and a customer. A wrong invoice match might cost an overpayment. A late supplier chase might cost a stockout.
Frequency times time gives hours a month. Multiply by what an hour of that person's time costs you, add a share of the error cost, and you have a rough monthly value. A task worth $800 a month pays back a $4,500 build in six months. A task worth $90 a month never pays back a build at all, and belongs in an off-the-shelf tool or on the list of things you simply accept.
Step three: check the data exists
An automation can only act on information it can reach. For each high-scoring task, ask:
- Where does the input live? In a system with an API (your store, your helpdesk, your accounting software), in email, in spreadsheets, or in someone's head?
- Where does the output need to go? Back into a system, into an email, into a document?
- Is the rule written down anywhere? Your return policy is. The way Priya decides which leads are serious probably is not.
Tasks whose input and output are both in systems with APIs are the easiest and cheapest to build. Tasks that start from email or PDFs need a model to read them, which is fine, but costs more and introduces a confidence threshold. Tasks that start in someone's head need the rule written down first, and writing it down is sometimes the whole improvement.
Step four: estimate the exception rate
Of every hundred times this task is done, how many follow the standard path, and how many need a judgement call? Order status is nearly all standard. Return eligibility is mostly standard with a clear policy. Complaints are nearly all judgement.
A task with an exception rate under about 15% is a good first automation: the standard cases run themselves and the exceptions go to a person with the data prepared. A task with an exception rate over 40% is not an automation candidate yet; a model can draft, but a person will still touch every one, and the saving is small. Between those two, the answer depends on how expensive the exceptions are and how cleanly they can be detected.
What AI is bad at
The marketing around AI implies it can do anything with text. Our experience building these systems says otherwise, and you should know the limits before you buy:
- Arithmetic and matching. A model will confidently add numbers wrongly. Totals, matching, date maths and lookups belong in plain code; the model's job, if any, is reading the text around them.
- Facts it has not been given. An assistant answering from your content is reliable about your content. Asked about something outside it, a model will often produce a plausible answer rather than none. Every system we build restricts the model to your sources and makes it say "I don't have that" otherwise.
- Judgement with consequences. Deciding a refund, handling a complaint, choosing between two suppliers: a model can prepare the information, but a person should decide, and the system should make that split explicit.
- Consistency over time. Models change when providers update them. An automation without a test list drifts quietly. This is why every build we do comes with a list of real cases it must pass, re-run after every change.
- Rare, high-stakes events. Automation earns its keep on the frequent and ordinary. For the thing that happens twice a year and matters enormously, keep a person on it.
What a fair budget looks like
For a store or service business doing a first project, the numbers we think are fair:
- A fixed build price between $3,500 and about $8,000, written down before work starts. An assistant on your own content from $3,500; support or order automation from $4,500; one internal workflow from $3,500; each extra system connection $600 to $2,000.
- Running costs on your own accounts: $40 to $200 a month in model fees for an assistant, more for automations running on every order, plus hosting usually under $20 a month.
- An optional care plan of $250 to $1,000 a month, cancellable any month, for a named engineer who watches the logs and fixes things when a vendor changes something.
- Nothing under $2,500. A task that small is an off-the-shelf tool or a platform setting, and a firm that takes the job anyway is pricing in the time it will take to argue about the result.
Competitor pricing for the same work ranges, on pages we have read, from a few thousand dollars to the tens of thousands, and some firms quote only monthly retainers of several thousand dollars. For a business your size, a retainer before anything is built is the wrong shape of deal.
How to buy
Whoever you buy from, including us, these five conditions protect you:
- A fixed price in writing for a defined scope, before work starts. Hourly billing for a first AI project puts all the uncertainty on you.
- Everything in your own accounts: the model provider, the hosting, the code repository, the automation tool. If the firm hosts it for you, you are renting your own process back.
- A test list of real cases the system must pass before launch, agreed up front, that stays with you as the acceptance test for every later change.
- A log you can read without an engineer, showing what the system did and what it handed to a person.
- A named person who owns it on your side. Not a department. One person who reads the log weekly and can switch any part of it off.
If a proposal fails more than one of these, ask why.
What to do next
Run the four steps above on your own list. If one task comes out with a clear monthly value, data in reachable systems, and an exception rate you can live with, put that task into the free AI fit check. You will get a one-page plan with a fixed price in two working days, or a note that says which tool to use instead.
If no task passes, you have your answer for now, and it cost you an hour. Come back when the list changes. Or email info@riantsoft.com with the list itself, and we will tell you which one we would look at first.
Questions people ask
Why don't you offer a paid AI strategy or roadmap session?
Because for a business your size the roadmap is short, and we would rather spend the hours building the first item on it. A paid half-day session produces a document; the fit check produces a plan with a price for one task, which is what you need to make a decision. If after the first build you want a second and third task scoped, we do that the same way, free, one page at a time. Firms that sell roadmaps make money whether or not anything gets built. We only make money if something does, so our incentive is to tell you quickly when a task is not worth it.
How do I know which task to put in the fit check?
Use the scoring method on this page: frequency, time per occurrence, cost of an error, and then check that the data exists and the exception rate is low. The task that scores highest on the first three and passes the last two is your candidate. If two tasks tie, pick the one whose output you can check most easily, because a first automation you can verify builds confidence for the second. If you are stuck, send us two or three; the plan will say which we would do first and why.
Is AI worth it for a business with five or ten people?
Sometimes. The arithmetic does not care about headcount; it cares about hours. A five-person business in which one person spends two hours a day answering order-status emails has a stronger case than a fifty-person business where the same task is spread thinly. The risk for a small business is paying for something with no clear owner. If nobody will read the log once a week, the automation will drift and you will stop trusting it. Decide who owns it before you decide whether to build it.
Should we just use ChatGPT or the AI features in our existing software?
Try them first, and we mean that. A general assistant on a staff member's screen helps with writing and summarising, and the AI features inside your helpdesk, store platform or accounting software may cover your task for the price of your existing plan. Where they fall short is anything that needs your own data, your own rules, and a record of what happened: looking up an order, applying your return policy, matching an invoice. The fit check will tell you if the built-in feature is enough; we would rather lose a job than build something redundant.
What does a reasonable first project cost, all in?
A fixed build price between $3,500 and roughly $8,000 for the first task, depending on the number of systems connected and whether a step needs a model. Running costs on your own accounts of $40 to $200 a month for an assistant, more for automations running on every order. An optional care plan of $250 to $1,000 a month. Your own time: perhaps four to six hours across the build, mostly checking test output. If anyone quotes you a first project in the tens of thousands for a business your size, ask what, specifically, you are getting that costs that.
What if the fit check says AI is not the answer?
Then you have a one-page document that says so, with the reason and the alternative: a named off-the-shelf tool, a change to your process, a form that fixes the input, or simply "keep doing it by hand". It is yours to keep and there is no follow-up pitch. A reasonable share of fit checks end this way, and we would rather you remember us as the firm that told you the truth than the one that sold you an automation you switched off after three months.