AI for Small Business: What Running It Ourselves Taught Us.

Gravitas is a small business. A founder, a handful of senior specialists who join when the work calls for them, and a few products we run ourselves. So when a client asks where AI fits in a company with twelve people and no IT department, we do not answer from a slide. We answer from our own inbox, our own CRM, and our own mistakes.

This post is the honest version. Where AI does useful work for us every week. Where we tried it, watched it make a mess, and turned it off. What it costs when you run it yourself. And the plan we hand a new client on day one, which you can copy without hiring us.

We Are a Small Business Too

Before the lessons, the credentials, because they are the point. Alongside client work, Gravitas operates SUDO, a private AI product that gives small businesses AI employees with a handbook, an approval step, and a receipt for every task. We publish a weekly AI newsletter. And for clients like Cuvée, a luxury travel brand, we own the plumbing between call tracking, HubSpot, and campaign reporting.

None of that is a lab. It is payroll, hosting bills, and customers who notice when something breaks. That changes how you think about AI. You stop asking what it can do and start asking what you would trust it to do at 6 a.m. while you are not watching.

Start With a Job, Not a Tool

The most common small business AI story goes like this. Someone buys a few chat subscriptions, a couple of people use them for emails, the rest forget the login, and six months later the owner cannot say what changed. Nothing was wrong with the software. The problem was that nobody defined a job.

Every AI system we still run started as a sentence. Read the shared inbox and sort the messages that need a human today. Turn any inbound call longer than sixty seconds into a CRM contact with its source. Gather what shipped in AI this week and draft the first version of the newsletter. Each sentence names an input, an output, and a way to tell if the output was right.

That is also how SUDO is built. A hire has a handbook page written in plain language, a defined shift, and a receipt for every task it completes. When the handbook is vague, the work is vague. When the owner rewrites one paragraph, the very next shift behaves differently. We learned that from watching our own customers, and it holds for any tool you use.

Where AI Did Real Work for Us

Research and first drafts

Our newsletter, the AI Readout, goes to 2,172 subscribers. A pipeline we wrote gathers what changed across the major labs and open-weight models each week, checks prices and benchmarks, and produces a brief. A model drafts the issue from that brief. A person edits every line and checks every number against the primary source before it goes out.

That last step earned its keep on the first issue. A widely repeated benchmark score for a new model turned out to be wrong. The number circulating was 87.4. The published index said 57. The draft had the wrong one. The human caught it. Without the check, we would have sent a confident error to more than two thousand people who read us for accuracy. AI made the issue possible in the time we had. The review made it publishable.

Turning phone calls into records

For Cuvée, most serious inquiries begin with a phone call, and phone calls are where CRM data goes to die. We connected the call tracking system to HubSpot so that any inbound call over sixty seconds creates or updates a contact, with the marketing source attached. The sales team sees where the inquiry came from without typing anything. Along the way we cleaned 1,512 contacts so that reporting stopped counting the same person three times.

This is the least exciting AI story you will read this year, and it is the one that changed the most decisions. Attribution only works when the record exists. The automation makes sure it does.

Engineering

We build and maintain SUDO with AI-assisted engineering. Assistants write tests, propose fixes, and review changes before a person does. The person still decides. What that buys a small team is the ability to run a product and a client practice with senior people instead of a bench. It does not buy the right to skip review. The one time we let a change through on trust, we spent a Saturday finding it.

Where We Turned It Off

The Cuvée integration originally did more. Every qualifying call created a contact and a deal in the pipeline. It worked exactly as designed, and the sales team asked us to stop. Automatic deals cluttered the pipeline with conversations that were never going to close, and the team stopped trusting the numbers. We removed deal creation and kept the contacts. The lesson stuck: automate the record, keep the judgment.

The same rule shapes SUDO. A hire drafts the follow-up, but the owner approves it by email before a customer sees it. That sounds slower. In practice the approval takes a few seconds, and it is the reason owners keep the hire on. They know nothing leaves the building without a look.

Data location is the other place we say no. A law firm or a CPA practice cannot put client files into a shared cloud model and hope for the best. That is why SUDO ships two ways: a box that runs in the office where nothing leaves the building, and a cloud version where each business is isolated from every other. If a vendor cannot tell you where your data goes and who else can see it, that is your answer.

What It Costs, Honestly

Running AI yourself is cheap. An AI hire on SUDO is $199 a month. A general chat subscription is less. The pipeline behind our newsletter costs pennies per issue in model calls. If cost is the reason you have not started, cost is not the reason.

The expensive part is the thinking. Deciding which job, writing the handbook, agreeing what counts as done, and connecting the tool to the information it needs. When a business hires us for that, it is a scoped project with a beginning and an end, and we wrote up what that engagement costs separately. But the first month does not need us. It needs one job and a calendar.

The 30-Day Plan We Hand New Clients

This is what we tell a small business on day one. It works with any tool, including the ones you already pay for.

1. Pick one job that happens every day. Inbox triage, quote follow-up, meeting notes into the CRM, weekly reporting. Skip anything a customer would see without a person in between.
2. Write the handbook page. One page, plain language. What comes in, what should go out, what to do when unsure. If you cannot write it, the tool cannot do it.
3. Measure the baseline first. How long does the job take a person today, and how often does it get skipped? Count for one week before you turn anything on.
4. Run it with approval for two weeks. Every output gets a human look before it lands. Fix the handbook when the output is wrong. Most of the improvement happens here.
5. Decide with numbers. Completed work, time saved, or a customer action that happened sooner. If none moved, kill it and pick the next job. If one moved, take the approval step off the low-risk outputs and keep it on the rest.

Thirty days later you will know more about AI in your business than most of the vendors calling you.

If You Want a Second Opinion

We offer a free applied AI teardown. Send your website and a sentence about what you do. Within a week you get a short recorded walkthrough: the three places we would put AI to work first, what each one takes to build, and what we would leave alone. No call required. The founder records it, the same as the work.

Request a teardown, or read the companion pieces on what AI agents can do for a small business and the ten problems we find in almost every website audit.