AI agents and automation
Someone on the team is still moving leads from the ad account to the CRM by hand. I build agents that do it themselves - with timestamps and checks.
What ends when the agent starts working
Nobody moves leads from the ad account to the CRM anymore, copies numbers into a report, or wonders in the evening whether some client was left without a reply. A program does it, with three parts: a trigger, access to the systems and the logic that decides what to do with each case.
Most automations do not need a language model: if a lead from an ad form has to go into the CRM with the right source and owner, that is rules, not thinking. A model is needed where unstructured text comes in: a client's email, an invoice in PDF, a free-form note. A flow without a model is cheaper and faster, and it makes nothing up.
The third variant is an AI chatbot for the website: the same agent with one channel. Without access to real data on prices, availability and orders, a chatbot is an expensive FAQ page. And if the process is not written down on paper, there is nothing to automate.
Trigger: a website form, a webhook, an incoming email, a schedule or a new record in a database
Access: CRM, email, Google Sheets, the website, the ad account, the accounting system
Logic: rules where the answer is unambiguous; a model where text has to be understood
Log: every run with time, input data and result, so that what happened can be checked
Human: one owner who receives error notifications and can stop the flow
What one real flow looks like, with timestamps
The Tenter ad lead path. Before automation, a salesperson saw a new lead only when someone opened Meta Business Suite and downloaded a file: sometimes the same day, sometimes the next.
Now less than five minutes pass between the form being filled in and the notification on the salesperson's phone. The most important thing is not the speed but that no lead gets lost and every one has a known source: then the ad budget can be split by results, not by feel.
14:32 - a person fills in and submits a Meta lead form on their phone
14:35 - a Cloudflare Worker picks up the new lead IDs from the Graph API every five minutes and requests the full data
14:35:02 - a duplicate check by phone and email, so the same person does not create two records
14:35:03 - a new lead is created in the Odoo CRM with all UTM fields: source, campaign, ad
14:35:04 - the lead is assigned to an owner by product group
14:35:05 - the salesperson receives a notification with the name, the phone number and which ad the person came from
08:35 the next morning - a daily summary: how many leads, from which campaigns, how many are still without a reply
Automations that are running right now
Below is work that runs every day, not demos. Each started with one process in which a person moved data from one place to another every day. Some started with connecting website forms to a CRM; the published work is in the portfolio section.
Clients: Tenter, Estire, MLM Cargo, ARM Metals, Tavasdurvis, Box Latvia, Universal Solutions, Improvement. Not everyone needs an agent - for some, one correctly connected form and a report that arrives by itself is enough.
Tenter: leads from Meta lead forms reach the Odoo CRM through a Cloudflare Worker. The flow runs 24/7, finds new ad forms by itself, checks for duplicates and keeps the UTM fields.
Tenter: Odoo process automation - lead stages, owner assignment, reminders about deals that are standing still, notifications in a Telegram channel.
Report generation: data from the Meta and Google Ads APIs is collected, compared with the previous period and turned into a finished document. A human reads the conclusions, not copies the numbers.
Connecting website forms to a CRM: a request from a WordPress or React site lands in the CRM with its source, not in an inbox where it gets lost between messages.
My own bookkeeping: monthly subscription invoices are created by themselves, the PDF goes to the client, and I only see what has not been paid.
MCP integrations: I manage 22 WordPress sites in one interface - content changes, updates and diagnostics without logging into each site separately.
How much does AI agent development cost
The price depends on how many systems have to be connected and how many exceptions the process has. Two systems with clear logic have one price; five, two of which have no proper API, quite another.
Payment 50% at the start, 50% after handover, due in 5 business days. The price does not include third-party costs: hosting for the flows, model API tokens, an n8n or Cloudflare subscription. For a medium-sized flow these are usually 5-40 EUR per month, and I calculate them already in the audit.
| Service | Price | What is included | Timeline |
|---|---|---|---|
| Process audit and automation plan | from 400 EUR | A step-by-step description of the process, a list of stages that can be automated, a cost and savings calculation | 5-7 business days |
| One agent or automation (one process from start to finish) | 900-1800 EUR | Development, tests with real data, documentation, two weeks in which the automation runs in parallel with the manual path | 2-3 weeks |
| Complex integration (CRM, reports, several data sources) | from 2500 EUR | Several interconnected flows, data synchronization, report generation, a permissions model | 4-8 weeks |
| Monitoring and maintenance | from 100 EUR per month | Error notifications, fixes after API changes, a monthly summary of runs | subscription |
| Separate work and consulting | 60 EUR/h | Changes to an existing flow, team training, technical consulting | by agreement |
Where automation pays off
Five conditions I check before making a proposal. If the process meets at least three, the conversation is worth continuing.
The process repeats at least 20 times a month and the steps are the same every time
A mistake costs money: a missed lead, a wrong invoice, a forgotten reminder
The data is already digital and the systems have an API or a webhook
The process has been stable for at least half a year and nobody intends to change it
There is one person who owns this process and can say how it should work
Where I advise not to spend the money
Half of the conversations about AI agents end with automation not paying off: the process happens five times a year, the data is in scanned PDFs without structure, the process still changes every month, or a human makes the decision anyway. I have written separately about when I am not the right choice in general.
When data flows on its own, the team notices a mistake before the client does, not after.
How I measure the savings
Before development I write down two numbers: how many minutes the process takes now and how many times a month it happens. After the switch I measure the same ones again - without the first measurement the second means nothing.
An example with real volumes: 40 leads a month, each manual transfer to the CRM with source tagging 6 minutes. That is 4 hours a month; at an hourly cost of 15 EUR it is 60 EUR a month or 720 EUR a year. An automation at 1200 EUR with 100 EUR monthly maintenance never pays off on this calculation alone.
So I measure two more things. Response time: when a lead reaches the salesperson in five minutes rather than the next day, a larger share of people start the conversation. Lost leads: if two out of 40 leads a month get lost and the average deal is 800 EUR, that is a different calculation than 60 EUR for time.
Security and data processing
Automation means a program has access to company data, so security is not an add-on at the end but part of the architecture.
If the company requires that data must not leave the EU, that can be solved: either without a model at all, or with a model running in an EU region. This has to be known before development, not after.
Every system gets its own token with minimal permissions. An agent that writes data into the CRM cannot delete the database
Tokens live in environment variables or a separate key vault, never in code and never in the repository
I send the model only the fields the task needs. Name, phone and address are usually not among them
Every run is written to a log: time, input data, result, error. Without a log there is nothing to diagnose
Deletions and actions involving money require human approval. The agent prepares, the human approves
In case of an error the flow stops and sends a notification, instead of staying silent and continuing to write wrong data
GDPR: I describe where the data goes, how long it is stored and who the sub-processors are. That is one document, not a separate project
What I work with
I choose the tool by what the company already has, not by what is more convenient for me to sell: if the team wants to see the flow and edit it themselves, I build in a visual tool; if the volume is large or the logic complex, I write code. The code and the accounts stay with the client.
n8n and Activepieces, when the team wants to see the flow and make simple changes themselves
Cloudflare Workers and Python, when the flow is complex, the volume is large or API limits have to be handled
Odoo, WordPress and Shopify - the systems where the client's data already lives
Meta Graph API, Google Ads API, Google Sheets and Docs API, email APIs
Claude and OpenAI APIs where text understanding is really needed
MCP integrations, when an AI assistant has to work directly with the company's systems
How the work goes
In five steps, with a pause after the second: after the audit you can stop and build nothing, and the audit document can be implemented without me too. I credit the audit against the development price if we continue the work.
- 01
Conversation
45 minutes
You tell me where the time goes. I tell you whether there is something there to automate.
- 02
Process audit
from 400 EUR
I write down the process step by step with time spent, mark the stages that can be automated and calculate the costs.
- 03
Prototype for one process
Works with real data, but in parallel with the manual path.
- 04
Switch-over
For two weeks both paths run at the same time, we compare the results, then we switch the manual one off.
- 05
Monitoring
Error notifications, a monthly summary, fixes when a platform changes its API.
This page is about internal processes
AI agents here mean a company's internal work: data, flows, CRM, reports. If the question is how your company, and not a competitor, gets into ChatGPT and Google AI Overviews answers, that is different work - SEO, GEO and AEO. I do not mix the two in one proposal.
Where to start
Write to me at connect@gatisdesign.com and describe one process that takes the most time from the team. It is enough to describe what happens step by step and how often. I reply within one business day with an estimate: whether it is worth automating, how much it will cost and how long it will take.
I work alone - 18 years of experience in web development and design - so the number of projects taken on at the same time is limited.
Questions
How much does AI agent development cost?
How long does development take?
Do I really need AI, or is ordinary automation enough?
Will the automation work with my CRM?
What happens if the automation breaks?
Does my data end up in AI model training?
Can I take the agent with me if we end the collaboration?
Other services
Write to connect@gatisdesign.com and describe one process that takes the most time from the team. I reply within one business day with an estimate: whether it is worth automating, how much it will cost and how long it will take.