← Guides

        Guide
        
        Getting started

How to implement AI in your company step by step

Implementing AI in a company means choosing one piece of repeated work where it clearly pays, writing down the rules your people follow today, and building a small system that does that work while a person checks the results. You test it on past cases against a pass mark agreed in writing, run a supervised pilot, measure what it saves, and only then move to the next process. You can do it with your own team or with an outside partner such as Odysi, an AI studio that finds where AI is worth it, builds only that and leaves your team able to run it.

Thomas Trincado Updated 6 Oct 2026

Topic Getting started

Read 10 min

Where AI is worth it in a company

Start from the work, not from the technology. AI is worth it where three things meet: the work repeats many times a week, it follows rules you could explain to a new hire, and it has a cost you can measure in hours, delays or errors. In most companies that means jobs like these:

Answering the same customer questions on WhatsApp, email or the web, and collecting what is needed to open a request. See our AI agent for WhatsApp.
Checking documents and case files against a checklist, and asking for everything missing in one go. See AI document review.
Sorting email that arrives in several inboxes, and drafting the reply for a person to approve.
Answering your team from your own procedures, so the same questions stop landing on the same person.
Replacing a spreadsheet that has outgrown itself with a small tool built to your criteria, such as a quote or pre-qualification tool.

It is rarely worth it for decisions that need an expert's judgement, for work that happens twice a year, or where a mistake cannot be undone. Those stay with a person. There are more examples in AI process automation.

To check one idea in two minutes, run it through the AI Project Scorecard. The five questions behind it are explained in is an AI project worth building?

How to implement AI, step by step

These are the steps we follow, and the ones we would ask any partner to follow. They are the same whether you build in-house or with outside help.

Step 1Find where AI is worth it. List the repeated tasks in one team: how many times a week each one happens, how long it takes and what a mistake costs. Choose by cost, not by how impressive the idea sounds.
Step 2Pick one workflow. One process, one owner who decides what correct looks like, and one result you can measure, such as hours a week or time to first reply. A company-wide "AI programme" is where most projects stall.
Step 3Write the rules and the acceptance test. Sit with whoever does the work today and write down what they look at, what they decide and when they pass it on. Then collect real cases and agree in writing what counts as a pass. At Fixlab it was 50 real scenarios, zero critical failures and at least an 80% pass rate.
Step 4Build a small version. Only the part that matters, on your data and connected to the software you already use. Start by reading from your systems, and write back into them only once that has been tested. If a product already does the job, buy it instead: see build vs buy vs wait.
Step 5Test it on closed cases. Run it on past cases whose right answer you already know and score it against the test from step 3. It goes no further until it passes.
Step 6Pilot with human review. A few weeks on real work, with a person reviewing and approving every result before it counts. This is the step between a demo and production; we explain the gap in proof of concept vs production.
Step 7Measure. Compare with the numbers from step 1: hours saved, reply times, errors, and how often and why it hands work to a person. Count what it costs to run, not only what it cost to build. More in the ROI of AI workflow automation.
Step 8Extend. Only once the first workflow pays: widen what it covers, or move to the next process on your list. A second project usually costs less than the first, because the underlying platform is reused.

What usually goes wrong

Most AI projects that fail do so for reasons decided before anyone builds anything:

Starting from the tool ("we should do something with AI") instead of an expensive problem.
Trying to automate a whole department at once.
Mistaking a demo for the product, and running out of budget in the edge cases.
Taking people out of decisions that need judgement.
Building something nobody on the team can run once the builder leaves.

We go through them, and what the projects that work do instead, in why most AI projects fail.

What to do in-house and what to do with a partner

Some parts of an AI project are always yours, whoever builds it. Others depend on whether you have engineers with time to spare.

Job	Your team	A partner
Choosing the process	Knows where the hours and the errors are.	Brings an outside view, and should be willing to say "not worth it".
Writing the rules	Always yours: whoever does the work today.	Runs the interviews and writes the rules down with you.
Building and connecting it	Possible if you have engineers with time for it.	Usually the reason to bring one in.
The acceptance test	Decides what counts as correct.	Builds the test set and runs it.
The pilot	Reviews and approves every result.	Fixes what fails.
Running it afterwards	Owns it, and changes rules and answers.	Upkeep if you agree it; documentation either way.

Odysi summary of how the work splits on the projects in this guide.

If AI will be a permanent, central function and you can hire and keep the right people, building a team makes sense. If the need is a project, a partner is usually cheaper. We compare the two in in-house AI vs an AI studio.

Whichever you choose, train the people who will use it. If your company operates in the EU, Article 4 of the AI Act asks companies that use AI systems to take measures to support the AI literacy of the staff who use them, and it has applied since 2 February 2025. The European Commission says no certificate is needed, an internal record of training is enough, and it also covers staff who use tools such as ChatGPT for everyday work (AI literacy questions and answers, updated 27 July 2026).

Who can help you implement AI: what kind of company to look for

Ask "which company can help me implement AI?" and you will be pointed to very different kinds of firm. None is better in general; each one fits a different job.

Kind of company	Good for	Watch for
Large consultancies and IT integrators	Company-wide programmes, many departments at once, strategy and change management.	Cost and pace for a single workflow; the people who sell are not always the people who build.
Software development houses	Building to a specification you already have, and integrations with your systems.	They build what you ask for: deciding which process is worth it stays with you.
Freelancers and automation specialists	Small, well-defined automations, often with no-code tools.	Continuity: who maintains it if they move on, and whether it is documented.
AI studios, such as Odysi	Finding where AI pays in a company and taking that one workflow to production.	Small teams: check their capacity, and check that they hand over so you are not dependent.
Software products with AI built in	A common need that a product already covers well.	It works to the product's rules, not yours; check what it does with your data.

Odysi comparison. We are one of the kinds of company in the table, so read it as an interested but honest view.

Questions to ask before you hire anyone

Which process would you start with, and why that one?
How will we know it works? Ask for the test and the pass mark in writing before the build.
What stays with a person, and how does the system hand work over?
What will we own at the end: the code, the accounts it runs on, our data, the rules and prompts, and documentation our team can follow?
Who does the work? Are the people on the call the people who build?
What will it cost to run once it is live: model usage, hosting and upkeep?
Can you show us something in production, with numbers?

Red flags

They start from a tool or a model, not from your process.
They promise to automate a whole department, or full autonomy with nobody checking.
There is a demo, but no written test.
You would not own the code, the rules or the data, so leaving means starting again.
A package price before anyone has looked at how the work is done.
Nothing in production to show you, only demos.

What Odysi does, and how it works

Odysi is an AI studio that helps companies implement AI. We find the few places in your operation where AI pays for itself in money or time saved, build only those, and leave your team able to run them without us. The founders, Thomas Trincado and Mike Tucci, do the work themselves, fully remote, in English and Spanish, with clients wherever they are. Most of our work today is AI agents on WhatsApp, AI process automation and AI document review.

How it works: a short call about the work your team repeats. If AI will not pay for itself there, we say so on that call. If it will, we send a written proposal with the price and the assumptions behind it. Then come the steps above: we write your rules down with the people who do the work, agree the test with you in writing, build, run the pilot with your team reviewing, and hand over with training and documentation. Your rules, documents and data stay yours, and we do not share them with other clients.

What that looks like in production: at Fixlab, a vehicle certification lab, our AI agent on WhatsApp passed a written test of 50 real scenarios with zero critical failures before it launched. Between 26 June and 24 September 2026 it handled 279 conversations from 218 people, with a median first reply of 16 seconds; 51% closed without a person, and it opened 20 reform requests end to end in Fixlab's platform. For Urban Capital, a mortgage advisory firm, we built a pre-qualification tool to the firm's own criteria, now live in production.

We are a small team: two co-founders, plus James part-time. If you need a programme across many departments at once, a larger firm is a better fit. If you need one workflow done properly and handed over, that is our work. More about us on the studio page.

Checklist before you start

One process chosen, with an owner on your team.
Its numbers today: how often it happens, how long it takes, what a mistake costs.
The rules written down, as you would explain them to a new hire.
A set of real past cases whose right answer you know.
A pass mark agreed in writing.
What always stays with a person.
Where the data goes, written down for whoever handles data protection.
What you will own at the end, agreed before the build.
Who runs it once it is live.

Sources

European Commission: AI literacy, questions and answers (updated 27 July 2026)
Odysi's projects: the Fixlab and Urban Capital case studies

Read on 6 October 2026. Check rules at the source before you decide. This guide is not legal advice.

FAQ

FAQ: implementing AI in a company

How do you implement AI in a company, step by step?

Find where AI is worth it, pick one workflow, write down its rules and an acceptance test, build a small version, test it on past cases, run a pilot with a person reviewing every result, measure what it saves, and only then extend it. Most of the work is deciding and testing, not the AI model.

Where should a company start with AI?

With one piece of repeated work that follows rules and has a cost you can measure, such as answering the same customer questions, checking documents against a checklist or sorting email. Score the idea with the AI Project Scorecard before you spend anything.

Which company can help me implement AI?

It depends on the job. Large consultancies fit company-wide programmes, software houses build to a specification you already have, freelancers fit small automations, and AI studios such as Odysi find where AI pays, take one workflow to production and hand it over to your team. Ask any of them for a written test, what you will own at the end and something they have in production.

How much does it cost to implement AI in a company?

It depends on how much of your rules has to be written down, which systems it connects to, how much testing it needs and what it costs to run afterwards; the AI model is rarely the largest part. A first project on one process costs far less than a company-wide programme. We send a written proposal with the price and the assumptions after a short call. What moves the cost is explained in how much a custom AI tool costs.

How long does it take to implement AI?

It depends on the rules and the systems involved, so plan it in phases rather than to one date: writing the rules and the test, the build, a pilot of a few weeks with a person reviewing every result, then production. One workflow is one project; a company-wide programme is a series of them.

What can AI automate in a company?

Work that repeats and follows a rule: answering repeated customer questions, collecting and checking documents, sorting email across inboxes and drafting replies, answering staff from internal procedures, and turning an overgrown spreadsheet into a small tool. Final decisions, rare cases and anything that cannot be undone stay with a person.

Do we need an in-house AI team?

Only if AI will be a permanent, central function and you can hire and keep the right people. For one or two workflows, a partner who hands the result over to your team is usually cheaper. Either way, the rules and the review stay with your people. More in in-house AI vs an AI studio.

Do we have to change the software we use?

Usually not. The system connects to what you already use, through an API, database access or exports. If your software cannot be connected, the result reaches you by email or the channel you already use, with the data and documents in order.