ARTICLE 01 · AI TRANSFORMATION · 2026-11-22

How to implement AI in your company: 7 steps from day one to your first working system

A guide for whoever has been asked to bring AI into the organization. Seven steps, in the order the work really happens: pick one task, map how it's done today, check data and access, set the rules, choose who builds it, pilot with a small team, then open it to everyone. The one-page checklist at the end is ready to print for the kickoff meeting.

How to implement AI in your company: 7 steps from day one to your first working system
The short version
  • Start with one task the team repeats every week, with a clear owner and a number you can count. Pick the tools later.
  • Before anyone uses it, know three things: how long the current way takes, where the data lives and who may see it, and what kind of data staff may put into AI.
  • Pilot with three to five people first, have a person check every output, and compare against the old numbers every week.
  • The first system is done when the team runs it without the builder sitting next to them: there's a manual, an owner, and someone reading the numbers every month.

The job landed on your desk. Where do you start?

The owner says in a meeting that this is the year the company starts using AI, and the job lands on one manager. That manager's first question is usually which tool to buy. Implementations that work start from other questions: which task goes first, who needs to be in the room, and what must be finished before the next step. Implementing AI here means putting AI inside a work process that has an owner, uses the right data and has someone accountable for the result.

If you want the big picture first, what changing the whole company means and what the first 90 days should deliver, read What is AI transformation? Where to start. What follows is the hands-on part, one step at a time.

1 Pick the task2 Map the work3 Data and access4 Rules5 Who builds it6 Small pilot7 Roll out and run

1. Pick one first task you can measure

A good first task has four traits. It repeats at least weekly. One person can say whether it was done right. You can count its time or volume today. And if the AI gets it wrong, a person still catches the mistake before it reaches a customer.

Tasks that often qualify: drafting replies to customer emails that ask the same things, turning meeting notes into action items, reading invoices and keying them into the accounting system, and searching internal policy documents for new staff. Tasks that shouldn't go first: anything where a wrong result can't be undone, such as approving credit limits, and anything where nobody has agreed on the right way to do it.

Keep it to one task. If every department head proposes one, write them all down as a queue and pick the one that meets the most of the four traits. The rest wait for the next round.

2. Write down how the work is really done before you open any tool

Sit with the people who do the task for a day or two and write it down step by step: where the information comes from, who decides what, and which cases mean calling the supervisor. The procedure in the binder rarely matches what people actually do, and the gap is exactly where AI will stumble.

Meanwhile, collect a baseline for at least two weeks: average time per task, tasks per day and how often work has to be redone. Two months from now, these numbers will tell you whether the system really helped or only felt like it did.

This step ends with a one-page workflow diagram that the people doing the work read and nod at. If they shake their heads, fix the diagram before you move on.

3. Survey the data and who can access it

AI answers only as well as the data it reads. Go through each source the task uses: which drive the files sit on, which ERP or CRM holds the records, whether the data is current or months stale, and who in the company may see what.

Four levels are enough: public, internal, confidential, and personal data under Thailand's PDPA. Each level has its own rule for which AI tools it may go into. The thing to get right from the start is that the AI sees no more than the person asking is allowed to see. A salesperson shouldn't ask the chatbot a question and get accounting's salary figures back.

IT will want to know how many layers the data passes through and who looks after each one. The details are in The 7 layers of an enterprise AI system, and what IT decides in each.

For the development team · Permissions and usage trails

Tie access to user groups in the SSO the company already runs, such as Microsoft Entra ID or Google Workspace, instead of creating separate accounts in the AI tool. If document search uses RAG, filter permissions at retrieval time (permission-aware retrieval), because filtering after the answer is generated lets content the asker may not see slip into it first. Keep an audit log of who asked what and which documents the system pulled to answer, so you can trace back when someone raises a question about a data leak.

4. Set the AI rules before anyone starts

Good rules are short enough for staff to remember. They say which tools are allowed, which levels of data must stay out, which outputs need a person's check before leaving the company, and who to ask when unsure.

This step gets skipped because it looks like paperwork. Without rules, though, staff use personal accounts on free apps anyway, and the company has no idea where its customer data has gone.

The topics a policy needs and the parts PDPA touches are in An AI usage policy your staff will follow. For the first system, add two or three rules specific to that task. In the first month, for instance, every reply the AI drafts goes through a person before it reaches a customer.

5. Choose who builds it

There are three routes: build it with your own IT team, hire an outside team, or mix the two, with the outside team building the first system while teaching your people. Which one fits depends on whether anyone on your team has connected AI to real business systems before, and whether that person has the time.

If you hire outside, ask these five questions before you sign.

  • Show us a system you delivered that the client still uses. A demo alone isn't enough.
  • Who will sit with our staff, and does that person build the system or hand it to someone else?
  • Where will our data be stored, who can reach it, and how is it deleted when the work ends?
  • After handover, who holds the code, the AI provider accounts and the manual?
  • If the results miss the numbers we agreed on, what happens next?

DNA Maker does this work through our AI implementation service. Whoever you choose, the five questions above work with every vendor, including us.

6. Pilot with a small team, and check every output

Pick three to five users who do the task every day, and include one person who still doubts AI. This group will find problems the builder never thought of. This trial period is the pilot, and every output passes a human check before it's used (human-in-the-loop).

Review once a week; half an hour is enough. Look at three things: the numbers against the baseline, cases where the AI was wrong or couldn't answer, and work that users quietly went back to doing the old way. That last one says a lot. If people reopen Excel, some step in the system is still slower than the old way.

Write down the pass criteria on day one of the pilot, such as how much time per task must drop and the largest share of work that may still need fixing. If the numbers still fall short after four to six weeks, stop and look at whether the problem is the data, the process or the task itself. Stopping what isn't working and moving to the next task beats dragging it on because of the time already spent.

7. Roll out, hand over and keep it running

Once the pilot meets its criteria, expand to the whole team one group at a time. There's no need to switch on the whole company at once. The first pilot users teach their colleagues better than any manual.

Before the builder steps back, four things must exist: a manual for daily use and basic troubleshooting (a runbook), one system owner on the company side, a channel for reporting problems that someone answers, and a monthly report of the numbers against the baseline. AI systems shift with the data and with model versions. An answer that's right today may drift in three months, and checking the numbers monthly catches it early.

Then go back to the queue from step 1 and pick the second task. The second round is usually faster than the first, because most of the rules, permissions and data connections are already in place.

A one-page checklist before your first AI system goes live

Print this for the project kickoff. Every item should be ticked before the whole team gets access.

Before you start (steps 1–4)

  • One first task chosen, with a named owner on the company side
  • A workflow diagram that the people doing the work have confirmed
  • At least two weeks of baseline numbers collected
  • You know where the data lives, how confidential it is and whether it includes personal data
  • The AI sees no more data than the person asking is allowed to see
  • AI rules that staff have read and acknowledged

During the pilot (steps 5–6)

  • The builder answered all five questions before signing
  • Three to five pilot users, including one who still doubts AI
  • Pass criteria written as numbers before the pilot starts
  • A person checks every output before it's used
  • A weekly review with a list of the cases the AI got wrong

Before rolling out to the whole team (step 7)

  • The runbook is ready, and someone in the company can open it and fix basic problems alone
  • One system owner on the company side
  • A channel for reporting problems, and someone who picks them up
  • A monthly report of the numbers against the baseline
  • You know who holds the accounts, the code and the contracts with AI providers

A parts distributor that started with quote requests

Sample data (names and figures invented): a mid-sized distributor of machine parts has six office salespeople and receives about 40 quote request emails a day. Each one means opening the catalog, checking stock in the ERP and typing a reply. The owner wants AI across the company, and the team has proposed seven tasks.

Step 1 picks drafting replies to quote requests, because it happens every day, the sales manager can say which replies are right, and every email still goes through a person before it's sent. Step 2 finds that most of the time goes into matching part numbers that customers write differently from the catalog. The typing takes less time than anyone thought. The baseline is about 12 minutes per email.

Step 3 finds that each customer's special prices live in one manager's Excel file, so the team agrees that the AI drafts standard prices only and a person fills in the special ones. The pilot runs with three people for five weeks. After a table of alternative part names is added in week three, time per email falls to about 5 minutes.

This round worked because the team started by watching where the time went and only then configured the tool. The other six tasks are still in the queue, and the second one reuses nearly all the same permissions and rules.

The 1-1-1 rule for your first AI system

Before every round, write these three lines on a single sheet of paper.

  1. One task: a task repeated every week, named so that someone outside the department understands it.
  2. One owner: the person who can say whether the result is right, and who has the authority to change how the work is done.
  3. One number: something you can measure today and will measure the same way every week.

If you can't write one of the lines yet, don't move on to step 3.

Questions executives ask about time and data safety

How long does the first AI system take?

It depends on how ready the data is and how many systems need connecting. For a narrowly scoped first task, many companies plan around 90 days from step 1 to full rollout. The slowest parts are usually preparing data and waiting for access approvals. Setting up the tool tends to be quicker than expected.

Will the AI provider train its models on our company data?

That depends on the plan and the contract. Most major providers' enterprise plans state that they don't train on customer data by default, while free apps and personal accounts may have different terms. Read the terms of the plan your company actually pays for, and write into the step 4 rules that personal accounts may not be used with company data.

Can customer data that includes personal information go into AI?

Yes, when there's a lawful basis under PDPA, you send only what the task needs, and the provider has a data processing agreement stating where the data is stored and for how long. Many tasks work just as well with names or ID numbers removed before sending. Have your DPO or legal team review steps 3 and 4 from the start.

Do we need our own IT team before we start?

Not a big one. You do need one person on the company side who can manage user accounts and permissions, and after handover a system owner who can read the runbook and fix basic problems. If you don't have that person yet, have the builder train them alongside the work.

If the pilot misses its criteria, was it wasted?

Not if you know why it missed. The workflow diagram, the baseline, the rules and the data permissions all carry over to the next task. Pilots that miss usually stall on data that isn't ready or a process that hasn't settled, and both can be fixed before you try again.

DNA MAKER · AI TRANSFORMATION

Where we help in these seven steps

Your team knows which work is done right. Sales knows which customers get special prices, and accounting knows which documents get sent back. Our job is to listen to those people and organize what they know into a process and workflow, rules, decision points and data a system can read. Steps 2 and 3 are our main work at the start.

Then we design and build the first system to work with the tools your company already has, connect it to your ERP or shared drives, set permissions to match your org chart, and stay with the team through the pilot and the runbook handover. Our team in Bangkok designs and builds everything in-house, with no subcontractors. If you'd like to know which task should be your company's first, send us the list of tasks your team repeats every week. We'll sort them with the step 1 criteria before we talk about anything else.

Software engineering glossary

TermWhat it isA simple exampleWhat executives should ask the development team
WorkflowThe sequence of steps in one task, from the information arriving to the work being done, including the points where a person decidesReceive a quote request, find the part number, check stock, draft the reply, supervisor checks, sendHave the people who do this work confirmed the diagram?
BaselineNumbers for the current way of working, collected before you start so there's something to compare against12 minutes per email (sample data)How many weeks of numbers do we have, and will we measure the same way after launch?
PilotA trial period with a small group of users before the whole team gets accessThree salespeople use the system for five weeksWhat are the pilot's pass criteria, and who decides?
Human-in-the-LoopA person checks or approves the AI's output before it's usedThe sales manager approves each email before it leaves the companyAfter full rollout, which points still need a person's check?
Access ControlRules for who can see or change which data; the AI system must follow the same rulesA salesperson asks the chatbot and can't see salary dataDoes the AI use the asker's permissions or an administrator's?
RAG (Retrieval-Augmented Generation)A method where the AI searches the company's documents first and answers from themAsk about annual leave, and the system pulls the relevant page of the employee handbookDoes each answer link to its source document?
Audit LogA record of who did what in the system, and whenYou can see who asked about one customer's contract last FridayHow long are logs kept, and who can open them?
RunbookA short manual for running the system and fixing basic problems after handoverIf the AI gets a part number wrong, which table you add the alternative name toIf the builder is gone, what can our people fix themselves with this manual?
Try this tomorrow: Ask each department head to list three tasks their team repeats every week, then use the 1-1-1 rule to narrow them down to one. If you can write down an owner and a starting number, that task is ready for step 2.