Setup help for your own AI agent
For firms that already run their own assistant, or are building one, with ChatGPT, Claude, n8n or whatever they use. The assistant stays yours. We write and test its instructions with you, what documents it asks for and when it hands over to a person, and we help with Meta and WhatsApp verification. It is a fixed first step, with scope and price agreed in writing before we start.
When the assistant talks well but does not do its job
Any motivated team can put an assistant together with ChatGPT, Claude or a tool like n8n. Getting it to chat fluently is the easy part. The hard part comes after, and it usually looks like this:
- It talks to customers nicely, but does not collect the details and documents you need to start the job.
- Someone on the team spends hours every week checking what it said.
- It misreads a photo (an 8 reads as a 3) and carries on regardless. And your customers send photos, not PDFs.
- It sounds like a robot, or filters everything and passes it all to a technician, which is exactly what you wanted to avoid.
- It answers in generalities when your technical criteria are already written down.
- And to take it to WhatsApp, Meta has its own steps (verifying the business, registering the number), and sometimes the account gets locked with no clear reason why.
Almost none of this is fixed by switching tools. It is fixed with instructions written case by case, a test set built from your real questions and clear rules for when it hands over to a person.
Your assistant stays the same, on your tools. What changes is how it works.
Five pieces, worked on with you
We start from what you already have and what your team already knows. Whatever is not written down, we write with you, in plain language. You build or adjust in your tool; we guide, write with you and review.
-
Step 01
We review what you have
The instructions, the knowledge base and a sample of real conversations, with whatever personal data we agree on. We find where it fails and why: what it does not know, what it does not ask for, where it makes things up and when it should have handed over to a person.
-
Step 02
We rewrite the instructions with you
Case by case, with examples, from your written criteria. One short rule per situation, with what it should do and, where it helps, an example of what not to do.
- “If they ask for a price, do not give one: say the team prepares it and ask for the details for the quote.”
- “If a photo is hard to read, do not guess: say which part cannot be read and ask for another.”
- “If the answer is not in the criteria, say so and pass it to a person.”
-
Step 03
What it must ask for, and how it checks
For each kind of enquiry, which details, documents and photos it has to gather before passing it to your team, and how it checks what it receives. Anything it cannot read or check with confidence it does not accept: it asks again or flags it.
-
Step 04
When it hands over, and what it never does
When and how it passes the conversation to your team, with the case summarised so the customer does not have to repeat it. And what it never does: give prices, give a technical verdict it cannot back up, or promise dates.
A persondecides whatever needs judgement.
-
Step 05
A test set of your real questions
We gather your real questions, the hard ones too, with what the assistant should do with each, and run the set every time the instructions change. It goes live when it passes. You keep the test set and learn to run it yourself.
Youdecide when it goes live.
The assistant, the accounts and everything we write stay with you.
One rule rewritten, and the tests that confirm it
An example of how we work on a single rule: the one about documents and photos. First it is rewritten around a concrete case; then the test set runs, with the easy questions and the hard ones.
“Ask the customer for the necessary documents.”
“For a new request, ask for the spec sheet, a photo of the manufacturer's plate and the owner's ID. Check every photo: if a number cannot be read in full, do not fill it in yourself. Say which part cannot be read and ask for another. Example: ‘I can’t read the last digit on the plate photo. Could you send another one, close up and without glare?’”
-
Real question · common
“What do I need to get started?”
PassAsks for the three documents on your list, in one message.
-
Real question · hard
Plate photo with glare
PassSays which digit cannot be read and asks for another photo. Does not fill in the number.
-
Real question · common
“How much would it cost?”
PassGives no price: asks for the details for the quote and passes it to your team.
-
Real question · hard
A case your criteria do not cover
PassSays so and hands it to a person, with the case summarised.
-
Kept in the test set
The same photo, with the previous instruction
FailedIt filled in the number with what it thought it saw. That is why the rule was rewritten.
The whole set runs again every time the instructions change, so fixing one rule does not break another. The pass mark for going live is agreed with you in writing.
The instructions say what it should do; the tests say whether it does.
Is this the right fit, or the full build?
We do both. The difference is who builds it and how much work your team puts in.
Your assistant, working as it should
For firms that already have an assistant, or want to build one and run it in house. You keep everything we write and test together, and the method to keep doing it.
We build it
For firms that would rather not build it. See the WhatsApp customer service AI agent.
If you later decide you would rather we build it, what we wrote and tested together still counts. And if not, that is fine: the assistant, the instructions and the tests are already yours.
Meta verification, without losing weeks
Taking an assistant to WhatsApp through the API involves Meta steps that have nothing to do with AI and can stall a project for days. We help you prepare them, check them and read what each error is telling you.
- We do it with you, in your accounts and in your company's name. If you prefer, we do it for you in a shared session.
- The decisions are Meta's: approving verification or lifting a restriction. What we can do is make sure you reach them with everything in order.
- Verify the businessMeta asks for the business details (name, address, phone, email and website), a way to confirm your connection to the business and, if it cannot find the business, documents that prove it (Meta documentation).
- Register the numberThe number is added in WhatsApp Manager, its ownership verified and then registered with an API call. A number already in use with WhatsApp cannot be registered until that account is deleted (registration, phone numbers).
- How many numbersA new business portfolio starts capped at two registered numbers; Meta raises the cap to 20 once the business is verified or reaches a messaging limit of 2,000 (Meta documentation).
- Locked accountMeta documents error 131031 for a WhatsApp Business Account that has been restricted or disabled for a policy violation, or when it could not verify data in the request against the account, such as a wrong two-step verification PIN (error codes).
Nothing is locked to Odysi
We do not ask you to switch tools or open accounts in our name. We work on what you already use, and what we write together is yours from day one.
If you would rather do it yourself, here is the method: how to write instructions for AI agents, with a template to start from.
- The assistantYours, in the tool you already use: ChatGPT, Claude, n8n, Make or another.
- The accountsIn your company's name: Meta, the AI model and the orchestration tool.
- The instructionsA document of yours, in plain language, that your team can change without us.
- The test setYours. We teach you to run it and to add new cases as they come up.
A fixed first step, priced in writing
It is not an open bank of hours. Before we start, we agree in writing what the first step includes, how many working sessions and the fixed price. After that, extra sessions only if you want them.
- A short callWhat you have, what fails and what you want it to do.
- Scope, in writingReview, instructions, documents, hand-off, tests and a few sessions, at a fixed price.
- Working sessionsYou build or adjust in your tool; we guide, write with you and review.
- TestsThe test set runs, and we fix things until it reaches the agreed pass mark.
- You run itWith the instructions, the tests and the method to keep going.
Why test before going live: proof of concept vs production.
The method comes from an agent we built end to end, at Fixlab
Fixlab, a vehicle homologation lab, is not a setup help case: we built its WhatsApp agent ourselves, from start to finish. What carries over to your assistant is the method.
Before launch we agreed a written acceptance test with Fixlab: 50 real scenarios, with zero critical failures and at least 80 % passing as the bar. It went live only once it cleared that bar. It is the same discipline we bring to your test set.
Read the Fixlab case- 279conversations in the first three months
- 51 %closed with no person involved; most hand-offs are price questions, which the agent passes to the team
- 20modification requests created end to end in Fixlab's platform, with documents and data
- 0critical failures across the 50 acceptance-test scenarios
FAQ: setup help for your own AI agent
We already have an assistant built with ChatGPT or another tool. Can you work on it?
Do we have to change tools?
Who owns the instructions and the tests?
Can you help with Meta verification or a locked account?
We tried AI and it did not reach the level we need. Why would this be different?
How long does it take? How many sessions?
How much does it cost?
Is the assistant a subscription or pay per use? Does my ChatGPT plan cover it?
What if we later want you to build it?
Sources, checked 8 Oct 2026: Meta for Developers, “Become a Tech Provider” (business verification), “Register a business phone number”, “Business phone numbers” and “Error codes”; OpenAI, “Managing billing for ChatGPT and the API platform”. Fixlab figures: Fixlab case study. The rules, questions and tests on this page are examples.