All posts
AI Employee#how to implement AI in business#AI implementation#how to implement AI

How to implement AI in business: a 90-day roadmap for a company that already tried a chatbot

How to implement AI in business without a data team: a five-stage AI implementation roadmap, a one-page AI governance policy, how to measure AI ROI honestly, the change management that keeps a team on board, and the implementation challenges that stall most first projects.

Md Shohel· November 12, 2026· 13 min read
Listen to this post0:00 / 15:32
AI narration of this post, read in full

Most companies have already "tried AI". Someone on the team pays for a chatbot, two people use it to tidy up emails, and the owner asked it once to write a job ad. Six months later nothing about how the business runs has changed. The invoices still go out late, new enquiries still wait until Monday, and the weekly numbers still get pulled together by hand on Friday afternoon.

That gap, between people using an AI tool and the business actually running some of its work through AI, is what "implementing AI" means. It means deciding which work moves, who checks it, what it's allowed to do, and how you'll know it paid off. The model you pick matters much less. This post is a plain AI implementation roadmap for that: five stages over about 90 days, written for a company with ordinary apps and no data team.

We make Sanaf, an AI employee, so we have an interest in how this turns out. The roadmap works with any tool, and Sanaf only shows up near the end. If you haven't started at all, our AI readiness assessment is the better first read; it covers the first 30 days in detail. This post picks up from there and goes further.

The short answer

To implement AI in a business, pick one repeated job with a clear owner and a number that shows whether it improved. Write a short policy saying what the AI may do on its own and what needs a person's approval. Connect the AI to the apps where that job already happens, have a named person check its work for a few weeks, and measure the number against where it started. When the first job runs without much checking, add the next one. Most of the effort goes into choosing the work and writing the rules down. The technology part is usually the quickest.

Why most AI implementation stalls

Before the roadmap, it helps to know where first attempts usually stop. The common AI implementation challenges are rarely technical.

Nobody owns it

The AI was "for everyone", so nobody was responsible for checking its work or deciding what it should do next. Within a month people drift back to doing the task themselves.

It lives somewhere nobody goes

A separate dashboard or a browser tab that has to be remembered gets forgotten. Tools that sit inside the place the team already reads every day get used.

It can only talk

A chatbot can draft a reply, but someone still has to copy it into the email, look up the customer in the accounting app and update the spreadsheet. The person saves a little typing and keeps all the clicking. That's the difference our post on automation vs AI spends some time on.

There's no number

Without a starting figure, the review at the end of the trial turns into a debate about feelings. The people who liked it say it helped. The people who didn't say it didn't.

The rules were never written down

Either the AI is given access to everything and somebody panics after its first mistake, or it's given access to nothing and can't do the job. Both end the project.

Every stage below exists to head off one of these.

What AI transformation means for an ordinary company

"AI transformation" gets used for multi-year programmes at banks and insurers, with steering committees and a budget line of its own. For a company of ten or a hundred people, the useful version is smaller. It means that a growing share of the routine work, the follow-ups, the reminders, the data entry, the first draft of every report, is done by AI and checked by people, and that the people spend the freed time on the work that needs judgment.

You don't need a strategy deck to get there. You need a list of the jobs your team repeats every week, a rough idea of how long each takes, and the patience to move them across one at a time. An AI strategy for business at this size is mostly that list, in order, with a name next to each item.

The AI implementation roadmap: five stages in 90 days

The timings below assume one job at a time and one person giving it an hour or two a week. They stretch if the first job is messy and shrink if it's simple.

Stage 1, week 1: pick the job and the number

Write down every task your team repeats weekly. Next to each one, put who does it, roughly how many hours it takes, and what goes wrong when it slips. Pick the job that is repeated often, follows a pattern a new hire could learn from notes, and does limited harm if one instance goes wrong.

Good first jobs tend to look like these: replying to new enquiries within the hour, chasing invoices that are past due, sending booking reminders and confirmations, sorting a shared inbox, or writing a Monday summary of last week's sales. Poor first jobs involve negotiation, anything legal, or any decision where one mistake is expensive.

Then choose the one number that shows whether the job got better. Time to first reply. Invoices more than 30 days overdue. Hours a week spent on the task. Write down where it stands today, before anything changes. This is the baseline your AI ROI depends on, and it can't be reconstructed later.

Stage 2, week 1 to 2: write the rules

This is the AI governance policy, and for most companies it fits on one page. It answers four questions.

  1. What may the AI read? Which apps and which parts of them. The shared sales inbox, yes. The owner's personal email, no.
  2. What may it do without asking? Usually things that can be undone or are internal: drafting, labelling, filing, updating a record, posting a summary to the team.
  3. What must it ask about first? Anything that can't be taken back or that a customer sees for the first time: sending an email to a new contact, issuing a refund, deleting anything, paying anything.
  4. Who approves, and how fast? A named person for each kind of approval, and what happens if they're away.

Add a line for your staff too: which AI tools the company pays for and approves, and what may not be pasted into tools it doesn't control (customer records, contracts, passwords). That one line covers most of what an AI use policy at work needs to say.

Write it in plain words and share it with the team before anything is connected. People are more relaxed about AI in the workplace when they can see what it isn't allowed to do.

Stage 3, weeks 2 to 4: connect it and watch

Connect the AI to the apps where the job already happens. If enquiries arrive in Gmail and customers live in HubSpot, it needs both. Give it access to that account or that folder only, never a shared password.

For the first two weeks, have it do the job while the named owner checks every result. Expect corrections, and treat each one as a missing line in the instructions. "Don't quote prices for jobs outside the city" is a rule nobody thought to write down until the AI quoted one. After a week or two, the corrections should slow down noticeably. If they don't, the job is less of a pattern than it looked, and it's worth stepping back to a smaller slice of it.

This is also where AI integration in business stops being abstract. The question is whether the AI can do the clicking as well as the writing: find the customer, update the record, file the email, put the date in the calendar. If a person still has to do all of that by hand, you've bought a drafting tool.

Stage 4, weeks 4 to 8: loosen and measure

Once the owner is approving most results unchanged, move the routine cases from "ask first" to "do it and log it". Keep the approval step on anything that can't be undone. The owner now reads a log instead of approving each item, which takes minutes instead of hours.

At the end of week 8, compare the number with the baseline from stage 1. Write down three things: what changed, what it cost, and what still goes wrong. If the number didn't move, find out why before adding anything. Usually it's one of the stalling reasons above.

Stage 5, weeks 8 to 12: add the next job

Go back to the list from stage 1 and pick the next job. It goes faster the second time: the policy exists, the apps are connected, and the team knows what to expect. By the end of 90 days, a company that started from nothing can reasonably have two or three jobs running, each with an owner, a number and a log.

That's the whole AI implementation framework: one job, one owner, one number, written rules, a period of checking, then the next job. It's a dull framework, and the dull version is usually the one still running a year later.

How to measure AI ROI without fooling yourself

AI ROI for a single job is simple arithmetic once you have the baseline. Hours saved per month, multiplied by what an hour of that person's time costs, plus whatever the faster result is worth, minus what the AI costs to run.

Here is an invented example to show the shape. A company spends about six hours a week chasing overdue invoices. After the AI takes over the reminders, the person who owned it spends about one hour a week reviewing the log and handling the replies that need a human. That's roughly 20 hours a month back. Separately, the share of invoices more than 30 days overdue falls, which brings cash in sooner. Set both against the monthly cost of the AI, and you have a figure you can defend.

A few habits keep the number honest:

  • Count the checking time. The hour a week spent reviewing is a real cost. Leaving it out flatters the result.
  • Count the corrections. If the AI gets one case in ten wrong and fixing it takes longer than doing it, that belongs in the sum.
  • Use the same measure before and after. If you counted overdue invoices at 30 days before, don't switch to 45 after.
  • Give it long enough. One month includes the setup period and usually understates the result. Two or three months is fairer.

If the number doesn't justify the cost after three months, stop that job. Saying no to one job is part of an AI adoption plan, and it makes the yes on the next one more believable.

AI change management: bringing the team with you

The quickest way to sink an AI project is to announce it as a cost saving. People hear "we're replacing you" whether or not that's the plan, and they find reasons the AI can't do the job.

A plainer approach tends to work better. Say which task is moving, and who will check it. Say what the person who did it will do with the time, and make it something they'd rather do. Let the person who used to own the task write its instructions, since they know where the exceptions are. Invite the team to report mistakes, and fix the instructions in front of them when they do.

AI change management at this size is mostly about trust. People trust what they can see. An AI that works in the same team channel, says what it did, and asks before anything it can't undo is easier to trust than one that works out of sight. That's also why the written policy from stage 2 matters: it gives people something concrete to point at when they're unsure.

Tools, agents and AI employees: which kind to implement

There are three broad ways to put AI to work, and they suit different companies.

OptionWhat it isSuitsWatch for
AI features in your appsThe AI built into your email, CRM or accounting appOne task inside one appStops at that app's edge
Automation builders and agent platformsYou design the workflow or agent yourselfA team with someone who enjoys buildingSomebody has to maintain it
An AI employeeA hired AI that takes whole jobs across your appsA team with no builder and work spread across several appsLess control over each individual step

Many companies end up with more than one. Our post on AI for business automation walks through the builder route in more depth, and what is an AI employee covers the third.

Where Sanaf fits

Sanaf is an AI employee. You hire it into your team's Slack or Microsoft Teams, connect the apps it should work in, and give it jobs from a roster: the inbox, lead follow-up, booking, invoicing and others. It does the work inside the apps you already run, so it can do the clicking as well as the writing.

It follows the roadmap above by design. Each job is a separate roster entry, so you add them one at a time. It asks before anything it can't undo, and you decide which actions need your approval. It keeps a record of what it did, which is the log your owner reads in stage 4 and the evidence for your ROI sum. Because it works in the team channel, the team sees its work where they already are, which helps with the change management part.

It isn't for every company. If you already have a working automation stack and someone who maintains it, keep it, and use Sanaf only for work that stack can't reach. If your first job involves decisions where one mistake is costly, it should stay with a person.

You start with $120 of free credit and no card, to use in your first 30 days, which covers stages 1 to 3. After that you buy credits from $20 when you need them, and credit you buy never expires. You pay for the work, not the seat. The pricing page has the details, what an AI employee costs sets it against hiring, and hire an AI employee walks through the first day.

Questions people ask about implementing AI

How do you implement AI in a business?

Pick one repeated job with an owner and a number that shows whether it improved. Write down what the AI may do on its own and what needs approval. Connect it to the apps where the job happens, check its work closely for two or three weeks, then loosen the checks on routine cases and measure the number against where it started. When that job runs well, add the next one.

What are the steps of an AI implementation roadmap?

A practical AI implementation roadmap has five stages: choose the job and record a baseline number, write the rules for what the AI may and may not do, connect it and check every result, loosen the checks and measure, then add the next job. For most companies with ordinary apps, the first three jobs fit in about 90 days.

How long does it take to implement AI in a business?

A first job can be running in two to four weeks, including a period of close checking. Two or three jobs running with light oversight is realistic in about three months. Large-scale AI transformation at a big company takes much longer because of the number of systems and approvals involved.

How do you measure the ROI of AI?

Record the task's time and result before the AI starts. After two or three months, compare hours spent and the result itself, include the time spent checking and fixing the AI's work, and subtract what the AI costs to run. Use the same measure before and after, and give it long enough to get past the setup period.

What should an AI governance policy include?

For most companies, one page that says which apps and data the AI may read, which actions it may take on its own, which actions need a person's approval first, who approves them, and which AI tools staff may use and what they must not paste into tools the company doesn't control.

What are the biggest AI implementation challenges?

The most common are not technical: no named owner, a tool that sits somewhere the team never looks, an AI that can write but can't act inside the apps, no baseline number to judge it by, and no written rules about what it may do. Each one is fixable before you start.

Do you need a data team to implement AI?

No. Modern AI can work from the email, documents, spreadsheets and app records a company already has. What it needs is a clear job, access to the right apps and a person who checks the results. A data team matters for building custom models, which most companies never need to do.

Get the next one by email

We write when we have learned something worth passing on, so nothing arrives unless it is worth your time.

Unsubscribe in one click, any time. Privacy policy