The playbook
How a small company actually uses AI.
Not predictions. The systems we run our own business on, in the order we built them, and what each one replaced.
We were paying four companies to not talk to each other.
Where we started
The problem was never intelligence.
I run Pivot XP, an experiential technology company. We build custom software and the hardware it runs on, then ship it to live events where it has to work the first time, in front of a crowd, with no second take.
Before any of this, the company ran on four separate things at once: an event and rental platform, a general CRM built for software sales cycles rather than gear that physically ships, a no-code stack somebody assembled to cover the gaps, and the spreadsheets, accounting software and email threads that covered the gaps in that.
Every one of them held part of the truth. None of them held all of it. A quote lived in one system, the equipment it promised lived in another, and a person retyped the difference. That is the actual problem most small companies have, and no chatbot fixes it.
Lesson one
Consolidate before you automate.
AI sitting on top of four disconnected tools gives you four disconnected answers. So the first build was not an AI feature at all. It was one system that finally held the whole business.
The commercial side
Clients, contacts, projects, quotes with line items, contracts, invoices and the financial reporting on top of them. One record from first conversation to final payment.
The physical side
Every item we own, down to the individual unit: what kit it belongs to, which case it ships in, what event it is reserved for, and when it is due back.
The field side
Shipments, equipment orders, tracker trails on cases in transit, repair issues raised from the field, and the maintenance each unit is due for.
One source of truth
A quote now draws on the same catalog the warehouse picks from. When something is promised, the system already knows whether it exists and where it is.
That replaced all four of the tools we had been paying for.
Lesson two
Give it real access, not copy and paste.
Most companies use AI by pasting things into a chat window. You copy a spreadsheet in, get an answer back, and retype the result into the system it came from. That is still a person doing data entry, just with extra steps.
Our system speaks a protocol Claude understands directly, so it is not pasted into. It is operated. I can ask what is going out the door next week, what a client has open, which cases are overdue back, or to draft a quote, and it works against the live business rather than a copy of it.
The part that makes that safe is boring and essential: access is scoped to the person asking. The system knows who I am and what I am allowed to see, and company-wide financials do not open up for someone whose job does not include them. AI inherits those permissions rather than bypassing them. The full version of this argument, including the questions to ask any vendor, is in is it safe to use AI with your company's data.
It reads the real thing
No exports, no stale copies, no screenshot of a dashboard from last Tuesday. The answer reflects the business as it is right now.
It respects who is asking
Permissions live in the system, not in the prompt. A person cannot ask their way into data their role does not cover.
People approve what matters
Drafting a quote is automatic. Sending it to a client is a decision a human makes, every time.
We own all of it
The system is ours. If we stopped using any particular AI model tomorrow, the business would keep running exactly as it does today.
Lesson three
Automate the assembly. Keep the judgment.
Selling a custom build involves a lot of thinking and a lot of typing. The thinking is the job. The typing is not, and that is the part worth handing over. I broke the proposal half of this down step by step in using AI to write proposals from your own past work.
Describe the job
A conversation about what the client needs becomes a structured quote, priced against the real catalog rather than a number someone half remembers.
It becomes collateral
The quote generates a branded document and a shareable link. Our catalog can be produced in a partner agency's own branding, so they present it as theirs.
It closes itself out
Approved work becomes a contract, then an invoice, then a project with gear reserved against it. Nobody retypes anything along the way.
The pricing judgment is still mine. The forty minutes of document assembly is not.
Lesson four
Build your website for the machines that read it.
This is the one I would do first if I were starting over, because it is the one most small companies get wrong by default.
Why we did not use WordPress
Nothing against it. The question I cared about was what the page actually is at the moment it arrives. We built the site as its own application that sends finished HTML, so the words are in the page when it lands. No theme, no stack of plugins assembling the content on the way out, nothing deciding on our behalf what the machine-readable version looks like.
That has a practical side too. There is no plugin to keep patched, no admin login to be brute forced, and no mystery about why a page renders differently this month than it did last month. For a company whose clients send procurement reviews before they sign, that matters more than a drag and drop editor.
One page per real question
The site is around seventy pages, and the structure is the whole point. There is a page for photo booth rental. A page for Starlink rental. A page for RFID wristbands. A page for event internet. Pages for leaderboards, fan games, sweepstakes microsites.
Most companies put all of that on one services page with six bullet points, because that is how their internal org chart looks. But nobody searches for your org chart. They search for the one specific thing they need, and the page that wins is the page that is actually about that thing. Match the structure of the questions, not the structure of your company.
Write the page as the answer
Our best performing page is not clever. It takes the questions that brand legal teams and procurement departments actually ask before signing, and answers them in plain language, each under its own heading. Who owns the guest data. What we capture. Where it lives. How long we keep it. Where AI is used and where it is not.
That format works twice. A search engine can lift a direct answer from it, and an AI assistant can quote it, because the answer is right there under a heading that matches the question. Burying the same information in a paragraph of marketing language would have made it invisible to both.
Tell the machines the facts outright
Every page carries structured data: a block of markup that states plainly what the organization is, what the site is, and what each frequently asked question and answer pair contains. It is not a ranking trick. It removes ambiguity, so nothing has to guess at what your business does from your adjectives.
The unglamorous parts count as much. One main heading per page with a real hierarchy underneath it. A canonical address so the same page is not competing with itself. Alt text on every single image, which on our homepage means all thirty of them. Most sites fail that last one, and it is free.
The honest version of "optimize for AI"
People ask whether you now have to optimize for Google or for AI assistants. The answer is that it is mostly the same work, and claiming otherwise is usually someone selling a new service.
When a buyer asks an assistant to recommend a vendor, the answer is assembled from two things: what the model absorbed in training, and what it can fetch while it answers. Both reward the same qualities. A page a crawler can actually open. Words that live in the HTML rather than being assembled afterward. Headings that match real questions. The answer in the first sentence instead of the fifth. That is the same list that has always made search work. You do the work once, and it serves both.
The three files almost nobody adds
A robots file that explicitly welcomes AI crawlers by name rather than quietly blocking them. A sitemap so nothing has to be discovered by luck. And a plain text summary at /llms.txt that states who you are, what you offer and how to reach you, written for a machine rather than a browser.
Be honest about what each one buys you. The robots file and the sitemap do real work. The llms.txt is a cheap bet on a convention that coding agents read today and assistants may read later, and one large log study found most published files are never requested at all. This site has all three, and they took about twenty minutes. If you only do one thing from this section, open your own robots file and make sure you are not blocking the assistants your buyers are asking.
Check that any of it is reaching them
Then verify, because the embarrassing failure is doing all of this correctly and still being invisible. Security settings, bot protection and firewall rules can all end up serving a challenge page to crawlers instead of your site, and nothing on your end looks broken when it happens. Your pages load perfectly in your own browser.
Test it the way a crawler sees it, not the way you see it. Check what your site returns to a search engine or an assistant specifically, and confirm in your search console that your pages are actually indexed. I check this on every site I build, including this one.
Lesson five
The boring data is the valuable data.
Everyone wants AI to write things. The bigger wins at our company came from the least glamorous corner of the business: knowing where our equipment is and what condition it is in.
Cases carry trackers, so a shipment has a trail rather than a guess. Gear is reserved against specific projects, so double-booking the same unit on two events in one weekend is something the system catches instead of something a client discovers. Technicians raise repair issues from the field, and maintenance comes due on a schedule rather than when something fails on site.
None of that sounds like artificial intelligence. All of it is the structured operational data that makes AI useful later, because a question like whether we can take on a job that weekend has an actual answer sitting underneath it.
Still in progress
What we have not finished.
We are building an internal voice agent to take routine phone work off the team. It is genuinely useful and genuinely not finished, and I would rather say that than present it as a result.
That is worth a note of its own. Plenty of AI case studies are written about pilots that never reached daily use. Everything else on this page is in production and gets used every week. This one is not yet, and when it is, I will say what it actually saved.
Start here
What to do first, in order.
If you want to do this in your own company without hiring anyone, this is the order I would follow.
1. Find the retyping
Spend one week noting every time somebody copies information from one system into another. That list is your roadmap, and it is almost never what you assumed.
2. Fix the worst one properly
Consolidate before you automate. One place where that information lives correctly beats a clever assistant stitching together three places where it does not.
3. Open your robots file
Twenty minutes, no budget required. Confirm you are not blocking AI crawlers, add a sitemap, and test what your site returns to a crawler rather than to you.
4. Rewrite one page as answers
Take the questions buyers actually ask you and answer them under their own headings, in plain words. That single page will outperform most of what you have written.
Questions
Asked and answered.
Do I need to build a custom CRM to use AI in my business?
No. We built one because event production did not fit anything off the shelf: our product is partly software and partly physical gear that ships to a venue. If an existing tool genuinely fits your business, use it and spend your effort connecting it properly instead. The lesson that transfers is consolidation, not custom software. AI on top of one accurate system beats AI on top of four partial ones.
Is WordPress bad for SEO and AI visibility?
No, and anyone who tells you that flatly is oversimplifying. A well built WordPress site can do everything described here. What matters is the output: does the page arrive with its content in the HTML, is there one clear heading structure, is there structured data, can a crawler fetch it. We chose a custom build because it gave us direct control over all of that, plus fewer plugins to patch. The principles matter more than the platform.
How do I get my business to show up in ChatGPT or Claude?
Mostly by doing the same things that have always made a site work in search, then verifying a crawler can actually reach you. Put your content in the HTML. Give each real topic its own page. Answer questions under headings that match how people ask them. Add structured data. Allow AI crawlers in your robots file and add a sitemap. Then check what your site returns to a crawler, because bot protection and firewall rules can block assistants without anything looking broken to you.
What does it mean to give AI "real access" to a system?
It means the AI can query and act on the live system directly rather than you pasting exports into a chat window. In practice the system exposes a set of defined actions, and the assistant uses them. The critical part is that access is scoped to the person asking, so the AI inherits that person's permissions instead of bypassing them, and anything that reaches a customer still goes through a human.
Isn't it risky to let AI touch your business systems?
It is, if you do it carelessly. The safeguards that make it reasonable are unglamorous: give it access to only what the task needs, scope that access by role, keep a human approval step in front of anything that reaches a client or moves money, and log what happened. Those are the same controls you would want around a new employee.
Where should a small business actually start?
Track where your team retypes information from one place into another for one week. That list is your roadmap. Fix the worst one by consolidating the data properly, before adding any AI on top. Separately, spend twenty minutes on your robots file and sitemap, because that one is nearly free and most companies have never checked it.
Want this
in your business?
I build these systems for other companies now. The first conversation is thirty minutes and costs nothing, and I will tell you plainly if AI is not the answer for you.