Agentic AI implementation
AI that does the work, not just answers questions.
A chatbot answers. An agent reads what came in, decides what to do, does it in your systems, and hands a person anything it isn’t sure about. I build custom AI agents for small businesses in Dallas–Fort Worth, at a fixed price, with approval by a person as the default.
What an agent is
Two sentences, no jargon
An agent is software that reads something messy, like an email or a PDF, decides what should happen, and then does it in the tools you already use. When it isn’t sure, it stops and asks a person instead of guessing.
Three of mine are running in production today: transaction classification from bank and card exports; bill intake from an email inbox, with PDF extraction; report and invoice generation. Each one has a human approval queue, scheduled runs, a memory of past decisions.
Rules-based automation vs. an AI agent
Most businesses should start with rules
A rules-based automation is cheaper, faster to build, and easier to live with. An agent earns its keep only where the input genuinely varies. The table is meant to help you tell which one you need.
| Question | Rules-based automation | AI agent |
|---|---|---|
| Best for | Repetitive work whose inputs look the same every time. Most businesses should start here. | Work where every input is a little different and someone has to read it and make a call. |
| What the input looks like | A form submission, a spreadsheet row, a file with fixed columns. | An email, a PDF in any layout, a drawing, a voicemail transcript. |
| What happens when the input changes | It stops or gives a wrong result, usually visibly, and needs a fix. | It usually copes and flags what it is unsure about. It can also drift quietly, which is why it is monitored. |
| Who checks the work | Nobody has to once it is tested. The output is predictable. | A person approves anything that matters, by default, until you loosen it. |
| What it costs to change later | Low. A rule is a line someone edits. | Moderate. The decision boundary and learned rules are editable, but changes are tested against real examples first. |
| Ongoing maintenance | Minimal. It breaks loudly when a tool changes. | Ongoing. Accuracy is measured monthly. Ninety days of monitoring is included, and I recommend keeping it. |
Where an agent earns its keep
Five things an agent does that a script can’t
All five come from the builds on my services page. Each is a before and an after, not a result I am claiming for a client.
- 01
Reads an emailed bill or statement in any layout and codes it to the right job and account.
- Before
- Someone opens each bill, reads it, and keys the vendor, amount, job, and account by hand.
- After
- The agent reads the bill in whatever layout it arrived, codes it, and queues it for a one-click approval.
- 02
Drafts a quote from the customer’s own email or drawing.
- Before
- Quotes wait until the owner has an evening to read the email or drawing and price it.
- After
- A draft quote is waiting in the morning, built from that email or drawing, for the owner to adjust and send.
- 03
Triages the shared inbox: answers the routine ones from your own files, escalates the rest with a summary.
- Before
- Whoever has time works the inbox, and the same questions get answered by hand every week.
- After
- Routine questions are answered from your own documents; the rest arrive with a two-line summary and a suggested reply.
- 04
Spots the exception in a PO or an invoice that doesn’t match and tells a person why.
- Before
- A PO or invoice that doesn’t match what was ordered gets found at month end, if at all.
- After
- The mismatch is flagged the day it arrives, with the reason spelled out, before anyone pays it.
- 05
Writes the weekly summary in sentences, not just numbers: what changed, what needs a decision.
- Before
- The weekly numbers are a spreadsheet someone assembles on Monday morning.
- After
- A written summary arrives: what changed, what is overdue, and the two decisions that need you.
How I build them
Custom AI agent development, in six steps
This is the same shape every time, whether the input is a bill, an email, or a drawing.
- 01
Ingest
Email, upload, export, or API, into one place.
- 02
Canonical record
Everything converted to one consistent structure.
- 03
Decide
The model makes the call, and remembers how you decided last time so it needs asking less often.
- 04
Review queue
Anything low-confidence, high-dollar, or new waits for a person.
- 05
Write back
Into the tools you already use, not a new system.
- 06
Run log
Every run, every decision, every reason, kept and readable.
Control and safety
What it can’t do without you
- 01
Approval is the default
Nothing sends, pays, or deletes on its own. Anything involving money, a customer, or a deletion waits in a review queue for a person.
- 02
Every decision is logged with its reason
Each run records what came in, what the agent decided, and why, in plain language you can read.
- 03
Learned rules are visible and editable
The rules it has picked up from your decisions are kept as readable entries you can change, not buried in a model.
- 04
You know what leaves your business
Before the build starts I put in writing which data a step sends out and which provider processes it, and under what terms.
- 05
Your accounts, your data
It runs on your accounts. Your records stay in your systems, and nothing is used to train anyone’s model.
- 06
You own everything I build
Outright, on final payment: the code, the configuration, the learned rules, and the logs.
Why agents need maintenance
A script breaks loudly. An agent degrades quietly.
When a tool changes under a script, the script stops and someone notices the same day. When your inputs drift under an agent, new vendors, new formats, a new person writing the emails, it keeps running and gets a little less right each week. Nobody notices until the numbers are off.
That is why every agent build includes 90 days of monitoring with accuracy reported monthly, and why I recommend keeping monitoring after that. It is not an add-on. It is the cost of running one honestly.
Pricing
Fixed price, agreed before I start
- AI agent build
- from $3,000one AI agent, fixed price, 90 days of monitoring included
- Connected systems
- from $4,000several automations or agents working together
- Monitoring & care, agent builds
- from $600/moagent builds · month to month, cancel anytime
- AI & automation opportunity audit
- $1,500bigger jobs only; credited toward the build
The exact price is fixed in writing before I start. Every tier, including rules-based builds, on the services page.
Questions owners ask
Fair questions about AI agents
Answered the way I would on the walkthrough call.
Will it make things up?
It can, which is why nothing it produces is treated as final on its own. Anything involving money, a customer, or a deletion waits in a review queue for a person. Every decision is logged with its reason, so when it is wrong you can see exactly why and correct the rule it learned.
What data leaves my business?
Only what a given step needs: the text of the email or document being processed and the fields needed to make the decision. I tell you in writing which providers process it and under what terms before the build starts. Nothing is used to train anyone’s model, and your accounts and records stay in your systems.
What happens when it gets something wrong?
The first weeks run in shadow mode and then approval-first, so mistakes are caught before they touch anything. When one gets through, the run log shows the input, the decision, and the reason. We fix the rule, and the fix applies from then on. Accuracy is measured and reported monthly, not assumed.
Do I have to buy new software?
Usually not. Agents write back into the tools you already pay for: your accounting system, inbox, spreadsheets, or CRM. If a build genuinely needs something new, it is named and priced in the scope before I start.
Can my staff change how it behaves?
Yes. The rules it has learned are kept as plain, readable entries your team can see and edit. Approval thresholds are settings you control, and the runbook explains each one.
Who owns it?
You do, outright, on final payment. The code, the configuration, the learned rules, and the logs are yours, running on your accounts. If we stop working together, everything keeps running.
How is this different from the AI already built into QuickBooks or my CRM?
Built-in AI works on that vendor’s own data, inside that vendor’s product, and it is often worth turning on. The work that eats your week usually crosses between tools: an email becomes a bill in accounting and a line on a job. That crossing is what I build, and I use the vendor’s built-in features wherever they already do the job.
Next step
Bring one messy job to a free walkthrough.
Thirty minutes is enough to tell whether it needs a rule, an agent, or neither. Dallas–Fort Worth, Texas, on-site or on a call.