AI Artificial Intelligence. Business woman using AI technology for data analysis

The short answer to how to implement AI in your business: start with one job that repeats, eats real time and produces something a person can check. Then:

  1. Measure how long it takes today.
  2. Give it to the person who already does it well and build a small first version.
  3. Test it on real work.
  4. Decide on purpose what the saved time is for.
  5. Do the next one.

For most companies, that’s a better place to start than trying to design a company-wide AI program first. The companies getting real value from AI tend to be the ones that picked a specific problem, changed how the work gets done and checked whether it paid off. In a short video I made in 2025, I put it as: stop playing with AI toys and start solving expensive problems. That still holds. What building with it since has added is that the technology is usually the easier part; deciding what the saved time is for, who owns the work and how you’ll know it worked is the harder one.

What the evidence says now

Four findings are worth knowing, mostly because they get quoted badly.

Business use is lower than the headlines suggest. The US Census Bureau found that in its survey period ending May 3, 2026, 19.8% of US businesses reported using AI in a business function during the previous two weeks. Use rose with company size: 32% of firms with 100 to 249 employees and 37% of firms with 250 or more. I used the same survey to estimate AI use by revenue and industry.

The “95% of AI pilots fail” number is shakier than it sounds. It comes from a preliminary 2025 MIT report on embedded and task-specific generative AI tools, not every AI project. The researchers found a steep drop from companies evaluating tools, to companies running pilots, to companies with something successfully in production, but the figures were partly based on interviews and used a demanding definition of success. The useful lesson is that getting from an interesting tool to measurable business value is much harder than getting people to try AI.

Most people feel faster. Few companies can find it in the numbers. McKinsey’s 2026 survey found that only about 37% of respondents attributed at least some effect on earnings (EBIT) to AI, essentially unchanged from the year before. About 6% qualified as high performers, meaning they attributed at least 5% of earnings to AI and reported significant value from it. The biggest difference: nearly three-quarters of the high performers had redesigned their workflows around AI, against about a quarter of the other respondents.

The saved time is leaking away. BCG’s 2026 survey found that 66% of frontline employees who regularly used AI got limited or no guidance on what to do with the time it saved them.

Put together, the evidence points less to tool access and more to what management does with it: redesigning the work, measuring results and deciding how the freed-up capacity gets used.

Step zero: find out what’s already happening

Before you choose anything, find out what your people are already using. In a lot of companies, somebody uses ChatGPT, somebody else bought a tool on a credit card, and leadership can’t tell whether any of it adds up to anything. That’s common, and it’s the real starting point.

Two things to do right away:

  • Ask. Which tools are people using, for what, and with what information? Make it safe to answer honestly. The goal is a map of what already exists.
  • Write a one-page data rule, before somebody pastes in something they shouldn’t. People using personal accounts for company work because nobody set the rules is a common problem.

The data rule doesn’t need to be long, but it should answer:

  • which AI tools and accounts are approved
  • what kinds of company and client information may go into them, and what may not (financials, personnel matters, anything a client contract restricts)
  • whether personal accounts are allowed for company work
  • which settings to use where tools offer them, like keeping your data out of model training
  • which meetings get recorded, who can read the transcripts, and where personal note-takers fit (a shared company recorder everyone knows about is different from someone’s personal note-taker, which shouldn’t feed company systems)
  • who may connect AI to company systems
  • which outputs a person must review before they’re used
  • where your contracts, regulators or industry rules override the general policy

For the level of detail to aim for, here’s what the answers might look like. Replace every one with your own, and have whoever handles security, privacy or compliance review the result before you adopt it.

  1. Approved tools: company accounts on our chosen assistant, for drafting internal documents and summarizing our own meetings. Its integrations and extensions need separate approval.
  2. What may go in: public information, our own marketing copy, internal meeting notes. Client information only where the client’s agreement allows it and the owner has approved the use.
  3. What may not go in: passwords and keys, financials, personnel and disciplinary matters, health information, and anything a client contract restricts.
  4. Personal accounts: no. Use your company account so the data stays under our settings.
  5. Settings: data not used for model training, shortest available retention, sharing limited to our company. Turning off training alone doesn’t make data confidential.
  6. Meetings and transcripts: client and internal meetings via the shared company recorder, with everyone told. Transcripts readable by the people in the meeting and their managers. Personal note-takers stay personal and don’t feed company systems.
  7. Connecting systems: only with the operations lead’s approval, with the minimum access the job needs. No changes to live records without separate approval.
  8. Human review: anything sent to a client, published, or touching money, contracts or people decisions, checked against the original source.
  9. What overrides it, and what to do after a slip: client contracts with data clauses and any rules for our industry come first. If you share restricted information with an AI tool, tell the owner of this rule straight away; deleting the chat may not remove it from the provider.

It’s a starting point for an operating rule, not legal advice. Where your contracts, regulators or industry rules say something different, they win.

Step one: find the expensive, boring problems

The best first uses of AI are rarely the impressive ones. They’re the jobs that repeat every week and that nobody enjoys.

The way to find them is to ask the people who do the work, and three questions get you most of the way: what do you do all day, what part do you hate, and can you show me? Leadership’s guess at the best first project and the front line’s answer are often different, and the front line can see problems leadership misses.

A good first candidate:

  • repeats often and takes meaningful time
  • has inputs you can actually get at
  • produces something a person can check
  • belongs to someone who wants the help

When you have more candidates than you can pilot, score each one from 1 to 5 on those four things plus how much time or money it costs you today, after first ruling out anything that fails your security or legal requirements, and mark the more complicated ones down. It doesn’t need to be sophisticated. It just stops the loudest idea from winning by default.

Boring is usually a good sign. Meeting notes are the classic first job, where your recording and data rules allow it. After that: proposal first drafts, call reviews, the weekly report someone assembles by hand, research before an important meeting. Another good one is a second reader: checking a proposal against the brief, or a report against the numbers it’s built on, before it goes out. It isn’t proof the work is right, especially when AI wrote the first version too, so someone still checks important claims against the source.

Leave some things for later: anything customer-facing that runs without a person reading it, and anything that touches money or contracts before you’ve seen you can trust the output. A useful test is whether you’d let a brand-new intern do the job without checking every last bit. If not, AI shouldn’t do it unchecked either.

Check what you already pay for, too. Many tools you already use have AI features, and getting more out of software you’ve already bought is often a better first move than buying or building something new. And notice when AI isn’t the answer. If the steps are fixed and the rules never change, a template or conventional automation will be cheaper and more predictable. If the real problem is that nobody agrees on how the job should be done, settle that first. AI will just do the disagreement faster.

Step two: measure the job before you change it

Write down how the job works today: how long it takes, who reviews it and what good output looks like, described by the person who does it best.

This is the easiest step to skip, and it’s the one that lets you tell the difference between “felt faster” and “was faster.” People’s impressions of how much time a tool saves them are a poor substitute for a measurement. Without a baseline, you can’t tell which kind of result you got. What to write down, and how to score the result each month, is in how to measure AI ROI.

Measure four kinds of payoff separately, because they get blurred together:

  • Capacity back. Hours freed up. It only saves money if something changes because of it.
  • Cash saved. A cost that actually goes away, like a subscription or outside spend.
  • Better work. Higher quality, fewer errors, better decisions.
  • Faster delivery. A proposal that goes out two days sooner is faster delivery. It doesn’t mean two days of labor were saved.

Measure what happens before and after the task, too. Making one step faster doesn’t help much if it creates more review, rework or waiting somewhere else. Then count what AI adds: setup, review, corrections, training, upkeep, the licenses themselves, and usage charges once tools connect to your data, which can climb fast if someone gets enthusiastic.

Step three: build something small, with the person who knows the job

The people who know what good looks like should shape the tool. Your senior estimator knows what a good estimate looks like. Your best account manager knows what a client update should say. When AI gets handed to someone who doesn’t know the job, the usual result is a tool that technically works and that nobody uses.

Three habits keep the first build honest:

  • Small before big. Build a reusable set of instructions and examples for one task, like drafting a proposal in your format or reviewing a sales call, before you try anything autonomous.
  • Prove it by hand first. Spreadsheet first, automation later. Prove the logic works in something cheap and visible before you build the automated version.
  • Read-only first. Let it suggest before you let it change records or send anything. Read-only can still expose information, and a model update or new integration can change how a tool behaves, so give it the minimum data the job needs and widen access only when there’s a real need, testing behind it and someone accountable. Before connecting AI to a core system, ask the vendor what can go wrong, what backups exist, and what they’ve seen fail for other customers.

Expect the first version to be rough. Its job is to show you where the instructions, examples and review process fall short, before you spend money automating it. Anything that connects to company systems, handles sensitive data or acts directly with customers needs technical and security review from the start, and any custom integration or software the business will rely on needs technical review early.

Where AI agents fit

An agent is AI that takes actions, like changing records, sending messages or working through a series of steps, rather than only answering in a chat window. They’re worth knowing about, and they’re rarely the right first move. I’d get comfortable in the chat window first, then move to tools that can act on your files and systems once you’ve seen how often the output needs correcting. The more a tool can do, the more damage it can do when it’s wrong.

What one implementation looks like

One we did at my agency, and wrote up this year: rebuilding our buyer personas from our own sales calls.

  • The problem. Our personas were about four years old. The company had changed and they hadn’t, and because they fed the way we write our own marketing, the drift showed up everywhere downstream. Meanwhile we had years of recorded sales calls and client conversations that nobody had ever mined.
  • The first version. We built something to read through all of it and pull out the exact words buyers used for their problems, goals and objections. The first version was wrong. It over-weighted whichever customers happened to have the most recorded conversations, so two talkative accounts could pass for a company-wide pattern.
  • The fix. We rebuilt the scoring around one rule: count distinct customers, not conversations. A theme only made it into our marketing once several different customers had said it in their own words. When we applied that honestly, a batch of our favorite in-house phrases fell apart, because no actual customer had ever used them.
  • The human check. We never let the tool decide what was true. It proposed, and a person decided. I reviewed what it found, and so did our salespeople, against what they actually hear on calls.
  • What it took. Real time, most of it the unglamorous part: feeding it conversations, watching where it went wrong and correcting it. It surfaced a kind of buyer we’d been under-serving in our messaging.

Notice the shape: a real job, a first version that failed in an instructive way, a person who knew what good looked like catching it, and a decision about what to trust.

Step four: pilot it on real work, with a small group

Run the pilot on real work, with the smallest group that has a specific job for it. Licenses are an easy way to waste money, and wanting someone to use a tool doesn’t mean they will.

Before the pilot starts, agree what would make you keep it: a minimum improvement in total time or cost, a quality standard it can’t fall below, and the most extra review or upkeep you’ll accept. Say, as a made-up example, a weekly report takes four hours. With AI, drafting takes 45 minutes and checking takes another hour, so it saves two hours and fifteen minutes a week before setup and upkeep. Then check that it’s at least as accurate and that people still use it.

Compare the results with your baseline, review time included. Train the people using it, because the first few weeks of use are where most of the fixing happens. Then decide: widen it, change it or stop it. A pilot that gets stopped has still told you something useful.

Before a pilot becomes part of normal work, make sure it has:

  • a named owner
  • agreed access and data sources
  • tests that include realistic failures
  • a way to catch and undo mistakes
  • a fallback for when the tool is down or wrong
  • a date to review its cost, accuracy and usefulness

Step five: decide what the time is for, and who owns it

This is the leadership part, and it’s where the evidence points.

What the time is for. Some companies use AI to do the same work with fewer people. Others use it to do better work with the people they have, or to grow without adding as many. That’s your call. My own preference is putting the time saved back into better work. Whatever you choose, make the call on purpose and say it out loud, because your team is guessing either way. The BCG finding shows how often nobody does.

Who owns it. One person accountable for adoption in each area, with time set aside to do it. When it’s everyone’s side project, it becomes nobody’s. Someone also has to own the instructions and examples after the first version works, or the tool quietly drifts out of date.

How you’ll judge it. Pick the measures before you start, and check them every month.

What usually breaks

  • Volume over quality. A common failure is a team that gets faster, produces far more and lets quality slide without anybody quite noticing, because every piece looks finished. AI is very good at producing pounds and pounds of mediocre work. If the goal is better work, the time saved has to go into specificity and quality.
  • Approval that turns into a rubber stamp. “AI proposes, a human approves” is the right rule. But as volume climbs, approval can quietly become rubber-stamping. Real approval means someone with the authority and the time to say no.
  • Tools nobody uses. Seats bought for everyone, used by a few. Start small and widen based on use.
  • Nothing written down. AI can’t reliably follow company-specific processes nobody has documented. If the real process lives in people’s heads and personal spreadsheets, writing it down is the first project.
  • Copies that drift. Keep your examples and standards in one maintained place the tool reads from, so an update to your brand or your process reaches every tool at once.

The people side

Getting AI into habits across a whole company takes longer than the first pilot, and much of it is people work.

You don’t need an AI class for the whole company before you start. Teach the people in the pilot how to use the tool for the work you’ve chosen: what good output looks like, what it gets wrong and what they’re still responsible for checking. Broader training can follow once people have something real to apply it to.

Start with directors and managers where you can, alongside the people who know the work best. Managers have the judgment to tell good output from bad, they can change how the work gets done, and they become the people their teams follow. Leaders also need to understand the tools and back them visibly. If leadership won’t use AI themselves, a broader rollout probably doesn’t make sense yet.

And take the fear seriously. People are often nervous about what AI means for their jobs, including your most experienced people. Say plainly what the time is for, and have those conversations in person where you can.

A first 90 days, roughly

Every company moves at its own pace, and security approval, technical dependencies or purchasing can stretch it, but a sensible first quarter often looks like this:

  • Weeks 1 to 2. Find out what people already use. Write the one-page data rule. Interview the people who do the work.
  • Weeks 3 to 4. Shortlist a few candidate jobs, pick one, name its owner and measure the baseline.
  • Weeks 5 to 8. Build the first small version with the person who knows the job, and test it on real work with a small group.
  • Weeks 9 to 12. Compare against the baseline, decide whether to widen, change or stop, and decide on purpose what the saved time is for. Pick the next job.

Do we need custom AI software?

Usually not at first. In my experience, most of the value is available through off-the-shelf tools you can subscribe to today, used well. Build custom only when a proven use case needs it.

Your first month

Ask your team what they already use, write the one-page data rule, and pick one boring, expensive job to measure this month. That’s the start of an AI adoption plan you can actually judge.

Choosing that first job, running the pilot and making the leadership calls around it is what I do as an AI implementation consultant. Tell me where your team is with AI.

Rodney Warner, seated, in a flat cap and dark t-shirt

Rodney Warner

Marketing and business advisor to owners and leadership teams. Based in Houston, working nationwide.

Talk it through with me

Bring the decision you’re stuck on. We’ll work out whether you need advice, an agency, or neither.

Start a conversation

Free scoping call. Reply within one business day.

It can be, but not as a claim. Every firm says it has great service, so the words give a buyer nothing to go on....

Use AI to make your team better first, then decide what that means for staffing. If AI reliably does the work at the same quality...

It should change who your marketing targets, what it promises, which objections it answers and how it says things. In a business where people close...