Buyers were searching for 26 problems we fix. Our website answered four.
Every week those pages didn't exist, those searches went to whoever had bothered to write them. Fixing it meant research, new page titles across the site, three new pages, an FAQ and a careful deploy. For a small firm, that's the job that sits on the list for a month. Our own AI team did it in one evening, and a person approved every step that touched the outside world. This page shows how that team is built, because it's the same thing we'd build for you, sized down to the one job that's costing you most.
What it was costing
Small firms rarely lose work to one big failure. They lose it to the list. The website page you meant to write in August. An enquiry reply that waited until Thursday. A blog that slipped three weeks running and then quietly stopped.
Ours looked the same. Chrysis works with contractors, engineering practices, law firms and accounting offices, and those buyers search for something specific: "law firm intake follow-up", "CPA firm offboarding", "stop retyping supplier invoices". When we checked in late October, our site gave a real answer to 4 of the 26 searches that fit our work. The other 22 sent buyers somewhere else, every day, while we were busy.
And the obvious shortcut, an AI tool that writes and posts on its own, carries a loss of its own. One confident wrong sentence emailed to a prospect, or one broken page pushed live on a Friday night, costs more trust than a month of saved hours buys back. We needed the work done fast. Nothing could go out unchecked.
"I'd been meaning to write those pages since August. The agents had them in a pull request before dinner, and all I had to do was read."
Tyler Fameli, founder
The software already in place
Nothing exotic. Microsoft 365 for mail, with Outlook as the place every outgoing message waits. A code repository where every website change arrives as a pull request somebody has to approve. Leads and accounts in a self-hosted CRM. And access to two families of AI models: Claude from Anthropic and GPT from OpenAI.
What we added sits on top of that: a task board the agents share, a scheduler for recurring jobs and a written set of rules about who may do what.
How it works
In plain English: one coordinator takes a request and breaks it into tasks. Four specialists do the work. A person approves anything that leaves.
The specialists each have one job.
Research. Search results, analytics and the CRM. It runs on GPT, which handles wide, fast lookups well.
Writing. Every word a customer will read, from page copy to email drafts. It runs on Claude.
Engineering. Code, website changes and pull requests, also on Claude.
Review. Advisory only. It puts the same question to two different AI models side by side and reconciles the places where they disagree, so a big call gets a second opinion before it reaches a person.
Work moves on a task board that survives a restart. A task waits until the tasks it depends on have finished, tries again when something flaky fails, and leaves a record of every run: who did it, with which model and what came back. Schedules start the recurring work, so nobody has to remember it.
One evening, start to finish
Here's the website job from the top of this page, as it ran on November 1, 2026.
The research agent built the list: 26 searches, each one a trade plus a problem, checked against what the site already said. Next, the writing agent turned that into page titles, search descriptions and FAQ answers across the site, plus three new pages for law firms, accounting and tax offices, and dental offices. Then the engineering agent put every change into one pull request. Sixty files. A separate reviewing agent read the whole diff before anyone else did.
Then Tyler read it and said go.
The coordinator merged it, took a backup, deployed, cleared the cache, opened every changed page on the live site to check it, and resubmitted the sitemap. Same night. From request to live took under five hours, and Tyler's part was about 20 minutes of reading.
What we automated
The legwork. Research, first drafts, code, the review pass, the deploy and the checks after it. Schedules run the jobs that repeat: three blog articles a week, each one researched, written, built as its own pull request, reviewed, deployed and announced to the search engines, with a LinkedIn post and a Reddit post drafted alongside. When a deploy fails its checks, the backup goes back on its own.
What still needs a person
Anything that leaves the building. Nothing is emailed, posted, published, merged, deployed or paid without Tyler, and he approves the exact thing, not a category of things. Outgoing mail lands in his Outlook Drafts, written and addressed, and he presses send himself. The LinkedIn and Reddit drafts arrive in his inbox and he posts them by hand. Website changes wait as pull requests he can read line by line, with an independent code review already attached.
The agents can't approve their own work. And approving one email never approves the next one.
Results
Since the team went live in August 2026:
- Website pages answering a real buyer search: 4 before that night, 29 after it.
- Blog articles: three a week, every week since September, 31 published so far.
- Runs recorded on the task board in October alone: more than 1,400, each with its record.
- Outgoing emails or posts sent without a person's approval: zero.
- Founder time on research and first drafts: from about 15 hours a week to about 3.
"The agents do the typing. I do the deciding. That split is the whole product, and it's the part most AI tools skip."
Tyler Fameli, founder
Same method, one job at your firm
You don't need five agents. Most firms need one, pointed at the pile that's costing them most: the shared inbox nobody triages until Thursday, supplier invoices retyped from PDFs, or bid requests that need a summary before the estimator looks. We build it the same way. The agent reads, sorts and drafts inside your own accounts. Anything it isn't sure about waits for a person, and nothing reaches a client until someone on your side approves it.
That's our Practical AI service, and under every agent sits plain automation like our access review build.
What it costs
Reviewed AI is $4,500 fixed for one AI workflow built in your accounts, about 15 business days once data access and security sign-off are done. AI model usage fees and software subscriptions are separate.
Misses the accuracy bar in the acceptance test we agree in writing? One free in-scope fix of up to 3 hours, then a retest. If it misses again, you choose 10% of the fee back as a partial refund or credit, up to $450.
Not sure a model belongs in the work at all? A $350 paid discovery answers that first, credited to the build if you go ahead.
Asked and answered
Can a small firm use AI agents without them emailing clients or posting on their own?
Yes, when every outbound step needs a person's approval. Chrysis runs its own AI agents this way: they research, write and build, outgoing email waits in Outlook Drafts, website changes wait as reviewed pull requests, and nothing is sent, posted, deployed or paid until a person approves it. A one-workflow build for a client is Reviewed AI at $4,500 fixed, plus model and software fees.
Should a small business use one AI model for everything?
Usually not. Chrysis picks the model for the job: its own writing and engineering agents run on Claude and its research agent on GPT, with an advisory reviewer comparing two models side by side when a decision needs a second opinion. For a client's workflow, the model is chosen against an accuracy bar the client sets before the build.
Who built it
Chrysis Networks is an automation firm based in Marietta, Georgia, led by Tyler Fameli. The agents, the task board and the approval rules are ours, built and run by us every day. We show them here because a firm buying AI should see how the people selling it keep it on a leash.
Find the job your team keeps pushing to next week, before a competitor does it first
Tell us the one job that keeps sliding and roughly how often it comes up. Within one business day you'll get a written reply: whether an agent fits, what a fixed price would cover and what would still wait for your approval.
No passwords or client files needed. Not even a sample document.