Should We Implement AI? Three Decisions, Three Answers

For most businesses with 10 to 200 people, the answer is yes, but "implement AI" is really three separate decisions, and each one gets a different answer. Letting your team use general AI tools is a yes, with a one-page rule, because they're probably doing it already. Turning on AI features inside software you already pay for is worth a cheap, measured test, one feature at a time. Building AI into a workflow is where the money and the payoff are, and it's a yes only for a workflow that passes a handful of checks you can run before you spend anything.

In September 2026 we read the 21 pages ranking for "should we implement ai" and five close variants. Five said yes outright and none said no. Only 2 gave a test you could use to reach "no" or "not yet," none separated the three decisions above, and none mentioned that a business is legally responsible for what its AI tells customers. For an owner, those are the gaps that matter.

This page takes each decision in turn, looks at what other businesses are doing and what the evidence says about the payoff, and sets out when the honest answer is no. It sits above our guide to AI readiness assessment services, which covers what to do once you've decided a workflow is worth a closer look.

Key Takeaways

• "Should we implement AI?" is three decisions: staff using AI tools, turning on AI features you already pay for, and building AI into a workflow.
• Of 21 ranking pages we read in September 2026, none separated those decisions, and only 2 gave a test for saying no.
• Adoption figures range from 18% to 78% because they count different things. The Census Bureau put AI use at about 18% of US firms at the end of 2025.
• Over 90% of nearly 6,000 executives surveyed in four countries reported no effect of AI on employment over three years, and 89% reported none on productivity.
• The payoff depends on the task. In one field experiment, consultants were 19 percentage points less likely to get a task right when it was outside what AI does well.
• A tribunal held Air Canada responsible for its chatbot's wrong answer. It found it "makes no difference" whether the information comes from a static page or a chatbot.

Table of Contents

• The short answer, decision by decision
• What other businesses are actually doing
• What the evidence says about the payoff
• Decision 1: letting your team use AI tools
• Decision 2: AI features in software you already pay for
• Decision 3: building AI into a workflow
• AI that talks to customers carries your name
• When the answer is no, or not yet
• A worked example: a 60-person electrical contractor
• Where Avolis fits
• Frequently Asked Questions
• Continue Learning

Should We Implement AI? The Short Answer

Yes for most operations businesses, as long as you make it three decisions instead of one. The three differ in what they cost, what can go wrong, and who has to own them, so a single yes or no to "AI" will be wrong for at least one of them.

The first two decisions are cheap to get wrong and cheap to reverse. The third isn't, which is why most of this page is about it, and why "not yet" is a legitimate answer there even when the first two are an easy yes.

What Other Businesses Are Actually Doing

The honest answer depends on who's counting. Surveys put it anywhere from about a fifth of US firms to three quarters of US workers at firms that use AI, and far fewer businesses have it built into their core work. If you feel behind, it's worth knowing which number you're being compared against.

The spread isn't a contradiction. The Census Bureau asks a large sample of firms whether they use AI, and about 18% said yes at the end of 2025. The Survey of Business Uncertainty asks senior executives, mostly at bigger companies, and reports that 78% of the US labor force works at a firm that has adopted AI, which is a count of workers, not businesses. A Federal Reserve note on the gap put it down mainly to "differences in sampling distributions and units of analysis" (Allen, "Monitoring AI Adoption in the U.S. Economy", FEDS Notes, April 2026). The Goldman Sachs figure comes from alumni of its own small-business program, and even there only 14% said AI was "fully embedded in their core operations" (Goldman Sachs 10,000 Small Businesses Voices, March 2026). ERGO NEXT, an insurer, surveyed 1,500 owners in April 2025 and found use falling, from 42% in 2024 to 28% (ERGO NEXT, May 2025).

Spending tells a more useful story than any of those rates. The Federal Reserve Bank of Atlanta surveyed 748 financial executives with Duke University and the Richmond Fed. Around half of the smaller firms in the sample invested in AI in 2025, and nearly 60% of smaller firms planned to spend less than $20,000 on it in 2026 (Baslandze and others, "How Might AI Change the Workplace?", Federal Reserve Bank of Atlanta, March 2026). A follow-up found that more than half of respondents expected to spend no more than $200 per employee in 2026, while the top 10% planned at least $2,800 per employee (Barrero and others, "How Much Are Firms Spending on AI?", Federal Reserve Bank of Atlanta, May 2026). In other words, most firms are spending modestly, and a small group is spending a lot.

What the Evidence Says About the Payoff

The evidence says AI pays off on specific tasks and hasn't yet shown up in most companies' overall results. That's not a reason to skip it. It's a reason to decide task by task instead of company-wide.

Start with the company-wide view. Researchers at the Atlanta Fed, the Bank of England, the Deutsche Bundesbank, and Macquarie University put identical questions to nearly 6,000 executives in the US, the UK, Germany, and Australia between November 2025 and January 2026. "More than 90 percent of executives report no effect of AI use on employment over the past three years, and 89 percent report no impact on labor productivity" (NBER Digest, "Global Evidence on Business Use of AI", May 2026, summarizing Yotzov and others, NBER Working Paper 34836). The Atlanta Fed survey found the same pattern from the other side: firms reported a 1.8% productivity gain from AI in 2025, but the gain implied by their own revenue and staffing figures was "much smaller across all major industries," and headcounts barely moved.

Now look at single tasks, where the picture is sharper. In an experiment with 453 college-educated professionals given short writing tasks from their own jobs, ChatGPT cut the time taken by 40% and raised quality by 18% (MIT News on Noy and Zhang, Science, July 2023). In a field experiment with 758 Boston Consulting Group consultants, those using GPT-4 on tasks within what the researchers called AI's "jagged technological frontier" completed 12.2% more tasks and finished 25.1% faster. Consultants given a task just outside that frontier were 19 percentage points less likely to reach the correct answer than colleagues working without AI (Dell'Acqua and others, "Navigating the Jagged Technological Frontier", Harvard Business School Working Paper 24-013, 2023).

A third study shows who gains. When 640 small-business owners in Kenya were given an AI business mentor over WhatsApp, the ones who were "already performing well saw considerable improvement," while those who weren't "experienced a downturn" (Harvard Business School Digital Data Design Institute on Otis, Clarke, Delecourt, Holtz, and Koning, "The Uneven Impact of Generative AI on Entrepreneurial Performance", 2024). These were very small firms getting advice, not operations businesses automating work, so treat it as a caution rather than a forecast.

Our reading of the evidence: the gains are real, they're specific to the task, and they go to businesses that already know how their work runs. That's why "should we implement AI?" gets a better answer as "which piece of work, measured how, and checked by whom?" A company-wide yes spreads effort across tasks where AI helps and tasks where it quietly makes things worse.

Citation-ready summary: In 2026 surveys, most executives reported no effect of AI on employment or productivity so far. Controlled experiments found large gains on specific tasks (40% less time on writing tasks) and losses on others (19 percentage points fewer correct answers outside AI's frontier). The decision to implement AI is best made one workflow at a time.

Decision 1: Letting Your Team Use AI Tools

Say yes, and write it down. Your people are very likely using AI tools already, so the real choice is whether that use follows a rule you've set or one each person has made up.

About 41% of people in the Real-Time Population Survey reported using generative AI for work by November 2025, according to the same Federal Reserve note. In an office of estimators, project managers, and coordinators, it's safest to assume someone already drafts emails or summarizes documents with a chat assistant, possibly on a personal account. That's mostly fine. What isn't fine is a customer's drawings, a bid number, or an employee's pay details going into a free account whose terms nobody at the company has read.

A one-page rule covers it, and it should answer four questions:

• Which accounts. One company account, or a short list of approved tools, so the business can see and control what goes in.
• What stays out. Customer documents, pricing and bid figures, and anything about employees' pay, health, or performance, unless the approved account's terms allow it.
• What gets checked. Anything that goes to a customer, a supplier, or an inspector is read by the person sending it, every time.
• Who owns the rule. One named person who updates it when a new tool shows up.

If you want to take stock of what's already in use before you write the rule, our guide to AI readiness assessments for SMBs includes a keep, standardize, or stop inventory built for this.

Decision 2: AI Features in Software You Already Pay For

Test these one at a time, for a month, against a number you count before you switch the feature on. Most field-service, accounting, CRM, and project-management vendors now offer AI features, some included and some as an add-on, and a feature that's switched on isn't the same as a feature that helps.

The cost is usually small, which is exactly why these features pile up unmeasured. Pick one feature tied to one task, such as summarizing requests for information on a job, drafting follow-up emails from the CRM, or reading supplier invoices into the accounting system. Count how long that task takes, or how often it goes wrong, for two weeks before the test, and again during it. Check the data settings while you're there, since some features send your records to an outside model and the default isn't always the one you'd choose.

At the end, keep it, turn it off, or stop paying for it. Our page on how businesses assess readiness for AI adoption turns this into a step-by-step trial with a keep-or-stop table, and it explains why a tool trial measures the tool rather than the work.

Decision 3: Building AI Into a Workflow

Build AI into a workflow only when the workflow passes a short screen, and run the screen before you talk to a vendor. This is the decision with real money and real payoff in it, and it's the one where "not yet" is most often the right answer.

The basics are four questions you can answer at your desk. Does the work repeat many times a week? Is it done roughly the same way each time? Does the information it needs already land in a system, an inbox, or a form? And is there one named person who'll own it? Our guide to assessing your own AI readiness turns those into a table with a way to check each one yourself, so we won't repeat it here. A workflow that fails one of them usually needs a process fix before it needs AI.

Two more checks decide whether the answer is yes, and they're the ones owners most often skip:

• Mistakes get caught, and catching them is faster than doing the work. This is where the Dell'Acqua experiment bites. AI output that looks right but isn't does the most damage where nobody checks it, and the check only saves time if it's quicker than doing the job. A drafted quote that an estimator reviews in three minutes passes. A drafted load calculation that an engineer has to redo from scratch doesn't.
• You can count it today. If nobody knows how often the work happens or how long it takes, you'll never know whether the AI helped. Count it for two weeks before you decide anything.

A workflow that clears all six is worth a closer look. For that, the pillar guide covers what a readiness assessment examines, our readiness assessment methodology explains why you score the workflow rather than the company, and the DIY guide's step 6 shows how to set a spending limit from measured hours. Our analysis of why AI projects fail covers what tends to happen when these checks get skipped.

AI That Talks to Customers Carries Your Name

Anything an AI tells your customers, you've told them. That's the plain reading of the best-known tribunal decision on the question, one every owner thinking about a website chatbot or an AI phone assistant should know, and none of the 21 pages we read mentioned it.

In November 2022, after their grandmother died, Jake Moffatt asked Air Canada's website chatbot about bereavement fares. The chatbot said they could apply for the reduced rate after booking, "within 90 days of the date your ticket was issued." Air Canada's employees later told Moffatt the airline didn't permit retroactive applications, and it wouldn't pay the difference. When the case reached British Columbia's Civil Resolution Tribunal, Air Canada argued it couldn't be held liable for information from its agents, including a chatbot. The tribunal member called that "a remarkable submission." The chatbot "is still just a part of Air Canada's website," the decision said, and "it makes no difference whether the information comes from a static page or a chatbot" (Moffatt v. Air Canada, 2024 BCCRT 149, paras. 27–28, February 2024).

The award was small, CAD $812.02 including interest and fees. The principle isn't. A small-claims decision in one province isn't binding precedent elsewhere, but the reasoning is ordinary negligent misrepresentation, and it's hard to see an owner arguing their way around it. If your chatbot quotes a price, promises a service window, or describes a warranty, the business has made that statement.

The risk runs the other way too, from vendors. In August 2025, the Federal Trade Commission sued Air AI, which marketed "conversational AI" that it claimed could replace customer service staff, alleging that its false promises left small businesses with losses of up to $250,000. "Companies that market AI-related tools with false promises of unrealistic investment returns and guaranteed refunds harm hardworking small business owners," said the director of the FTC's Bureau of Consumer Protection (FTC, "FTC Sues to Stop Air AI", August 2025). In March 2026, Air AI and its operators agreed to a proposed order banning them from marketing business opportunities (FTC, March 24, 2026).

The practical rule is simple. Put AI behind your staff before you put it in front of your customers, and when it does face customers, limit it to answers you've written and checked, with an easy route to a person.

When the Answer Is No, or Not Yet

The answer is no when AI would add cost and risk without removing enough work to matter, and not yet when the workflow is worth it but something it depends on isn't in place. Neither is a failure. The US government's AI risk framework says as much: "AI systems may not necessarily be the right solution for a given business task or problem," and a standard practice is to decide "whether its development or deployment should proceed" (NIST, AI RMF Playbook, MANAGE 1.1).

The answer is no when:

• Checking takes as long as doing. If a person has to redo the work to trust it, the AI has added a step, not removed one.
• The fix is a process change. If the work stalls because two systems aren't connected, or because a form asks for the wrong things, fix that first. It's often cheaper, and it may be all you need.
• The reason is a vendor's promise. A pitch built on guaranteed returns or replacing staff is a reason to ask for evidence, not a reason to start. The Air AI case is what happens when nobody asks.

The answer is not yet when:

• It would face customers unchecked. Keep a person between the AI and the customer until you've seen its answers hold up over weeks, not days.
• It fails one of the four basics. Low volume, a process that changes every month, information that lives in one person's head, or no owner. Our page on how businesses assess readiness describes what each of those looks like in practice.

Waiting on the third decision costs little if the workflow isn't draining you, and the surveys above suggest your competitors mostly haven't built much yet either. Waiting on the first decision is different, because your team won't wait with you. Write the rule now, even if you build nothing this year.

A Worked Example: A 60-Person Electrical Contractor

This is an illustrative composite, not a client, and its figures are examples rather than benchmarks. Picture a 60-person commercial electrical contractor doing tenant fit-outs and service work, with about 40 electricians in the field, four project managers, three estimators, and an office of six. The owner wants to "do something with AI." The co-owner, who runs operations, got burned by a scheduling tool three years ago and thinks it's a distraction.

They're both partly right, and the three decisions show where. The estimators already paste bid documents into personal chat accounts, so decision 1 isn't optional. The project-management software has a new AI feature that summarizes requests for information, which is a cheap test. A vendor has pitched a website chatbot to book service calls and quote availability, which would speak for the company to customers with nobody checking. And the co-owner mentions, almost in passing, that change orders are a mess: foremen text photos of extra work to the project managers, who write change-order requests when they get to it, and some extra work never gets billed at all.

That last one is the build candidate, because it clears the screen. It repeats every week, it follows a pattern a project manager could write down, the inputs already arrive as texts and photos, and the co-owner wants to own it. A project manager reviews every request before it goes to the customer, and reviewing a draft takes minutes where writing one from scratch takes much longer. The only check it fails is the count, since nobody knows how many change orders go unwritten.

What settles the argument between the partners isn't a vote on AI. It's the two-week count, owned by the skeptic, on the one workflow where the money is leaking.

Why the skeptic should own the count: it turns a disagreement about AI into a disagreement about a number, and the partner least likely to believe the number is the one both will trust. If the count shows little unbilled work, the build is a no, and the owner hears it from the person who ran the count rather than from a vendor. If it shows a lot, the co-owner has made the case themselves.

Where Avolis Fits

Our two-week diagnostic answers decision 3 for the workflows in scope. We work through each one with the people who run it, measure volumes, time per task, and handoffs, map where the work stalls, and rank what's worth building by impact and effort. Every engagement with us starts with that diagnostic, and the ranked result is yours whatever you decide to do next, including deciding that nothing is worth building yet.

If you're weighing who to work with on a build, start with which AI consulting company to choose. If you want to see where your business sits on the usual maturity scales first, our AI maturity model page explains the levels and why, at this size, what you've shipped and measured counts for more.

Frequently Asked Questions

Should we implement AI in our business?

For most businesses with 10 to 200 people, yes, but as three separate decisions. Let staff use AI tools under a one-page rule. Test AI features in software you already pay for, one at a time. Build AI into a workflow only where the work repeats, has an owner, can be checked faster than it can be done, and can be counted today.

When should a business not use AI?

Don't use AI where the task happens too rarely to repay setup, where a process change would fix the problem, or where checking the output takes as long as doing the work. Hold off where the process changes monthly, nobody owns it, or AI would answer customers with no one checking. NIST's AI risk framework treats not deploying as a legitimate outcome.

Is a business responsible for what its AI chatbot tells customers?

In the best-known decision on it, yes. In Moffatt v. Air Canada (2024), British Columbia's Civil Resolution Tribunal rejected the airline's argument that its chatbot was responsible for its own answers. The decision said it "makes no difference" whether information comes from a static page or a chatbot, and ordered Air Canada to pay CAD $812.02.

How much do small businesses spend on AI?

Most spend modestly. In a 2026 Federal Reserve Bank of Atlanta survey of financial executives, nearly 60% of the smaller firms planned to spend less than $20,000 on AI that year. A follow-up found more than half of all respondents expected to spend no more than $200 per employee, while the top 10% planned at least $2,800 per employee.

Will implementing AI mean cutting staff?

So far, mostly not. Of nearly 6,000 executives surveyed in the US, UK, Germany, and Australia, more than 90% reported no effect of AI on employment over the past three years. Looking ahead, they expected a 0.7% reduction over three years, mostly through slower hiring. At this size, AI mostly takes paperwork off people.

Continue Learning

"Should we implement AI?" is the wrong size of question. Ask it three times, once for each decision, and the answers get easier.

The decision that matters most comes down to one workflow and a two-week count.

Deciding and assessing:

• AI readiness assessment services
• AI readiness assessment methodology
• AI maturity model
• AI readiness assessment for SMBs
• How can I assess my organization's AI readiness?

Choosing who does it:

• Which AI consulting company should I choose?
• Is it worth hiring an AI consultant for a small business?
• Why do AI projects fail?

---

Sources

All sources retrieved 2026-09-25.

• Jeffrey S. Allen, "Monitoring AI Adoption in the U.S. Economy," FEDS Notes, Board of Governors of the Federal Reserve System, April 3, 2026, retrieved 2026-09-25 — https://www.federalreserve.gov/econres/notes/feds-notes/monitoring-ai-adoption-in-the-u-s-economy-20260403.html
• Salomé Baslandze and others, "How Might AI Change the Workplace? Evidence from Corporate Executives," Federal Reserve Bank of Atlanta Policy Hub, March 25, 2026, retrieved 2026-09-25 — https://www.atlantafed.org/research-and-data/publications/policy-hub-macroblog/2026/03/25/how-might-ai-change-the-workplace-evidence-from-corporate-executives
• Jose Maria Barrero and others, "How Much Are Firms Spending on AI, and What Will Happen to Headcounts?" Federal Reserve Bank of Atlanta Policy Hub, May 6, 2026, retrieved 2026-09-25 — https://www.atlantafed.org/research-and-data/publications/policy-hub-macroblog/2026/05/06/how-much-firms-spending-on-ai-and-what-will-happen-to-headcounts
• NBER Digest, "Global Evidence on Business Use of AI," May 2026, summarizing Ivan Yotzov and others, "Firm Data on AI," NBER Working Paper 34836, retrieved 2026-09-25 — https://www.nber.org/digest/202605/global-evidence-business-use-ai
• Goldman Sachs 10,000 Small Businesses Voices, "Survey: Small Businesses Embrace AI, But Need Training and Support to Fully Harness It," March 17, 2026, retrieved 2026-09-25 — https://www.goldmansachs.com/pressroom/press-releases/2026/small-businesses-embrace-ai-but-need-training-and-support-to-fully-harness-it
• ERGO NEXT, "Small business AI adoption declines to just 28%," May 28, 2025, retrieved 2026-09-25 — https://www.nextinsurance.com/blog/ai-for-small-business/
• Shakked Noy and Whitney Zhang, "Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence," Science 381, July 2023, as reported by MIT News, July 14, 2023, retrieved 2026-09-25 — https://news.mit.edu/2023/study-finds-chatgpt-boosts-worker-productivity-writing-0714
• Fabrizio Dell'Acqua and others, "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality," Harvard Business School Working Paper 24-013, September 2023 (SSRN 4573321), retrieved 2026-09-25 — https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4573321
• The Harvard Crimson, "Harvard Business School Partners with BCG on AI Productivity Study," October 13, 2023, retrieved 2026-09-25 — https://www.thecrimson.com/article/2023/10/13/jagged-edge-ai-bcg/
• Harvard Business School Digital Data Design Institute, "The Uneven Impact of Generative AI on Entrepreneurial Performance" (Otis, Clarke, Delecourt, Holtz, and Koning), 2024, retrieved 2026-09-25 — https://aiinstitute.hbs.edu/the-uneven-impact-of-generative-ai-on-entrepreneurial-performance/
• Civil Resolution Tribunal of British Columbia, Moffatt v. Air Canada, 2024 BCCRT 149, February 14, 2024, retrieved 2026-09-25 — https://decisions.civilresolutionbc.ca/crt/crtd/en/item/525448/index.do
• Federal Trade Commission, "FTC Sues to Stop Air AI from Using Deceptive Claims about Business Growth, Earnings Potential, and Refund Guarantees to Bilk Millions from Small Businesses," August 25, 2025, retrieved 2026-09-25 — https://www.ftc.gov/news-events/news/press-releases/2025/08/ftc-sues-stop-air-ai-using-deceptive-claims-about-business-growth-earnings-potential-refund
• Federal Trade Commission, "Air AI and its Owners will be Banned from Marketing Business Opportunities to Settle FTC Charges the Company Misled Many Entrepreneurs and Small Businesses," March 24, 2026, retrieved 2026-09-25 — https://www.ftc.gov/news-events/news/press-releases/2026/03/air-ai-its-owners-will-be-banned-marketing-business-opportunities-settle-ftc-charges-company-misled
• National Institute of Standards and Technology, AI RMF Playbook, MANAGE function, retrieved 2026-09-25 — https://airc.nist.gov/airmf-resources/playbook/manage/

On the page review. "The 21 pages" means the distinct organic results (44 result slots in all, before removing duplicates) for "should we implement ai," "should my business implement ai," "should my company use ai," "should small businesses adopt ai," "is ai right for my business," and "when not to use ai in business," fetched and read on 2026-09-25. They included pages from IBM, the US Small Business Administration, BizTech, HBS Online, FIU, Microsoft, MIT Sloan Management Review, MIT Sloan, two law firms, several agencies and IT firms, ERGO NEXT, Goldman Sachs, and nibusinessinfo. Five more (CompTIA, EarthLink, Medium, Fast Company, Forbes) couldn't be retrieved and aren't counted. "A test for saying no" means criteria you could check against a business, such as task frequency or acceptable error rate, rather than general cautions. Microsoft's page separated kinds of AI, but as a ladder of its own products. It's a snapshot of one day's results, not a market survey.

On the adoption figures. The four figures in the first chart answer different questions and come from different samples. The Census figure is the share of US firms using AI, as reported in the FEDS Note. In late 2025 the Census survey widened its question from AI used in "producing goods or services" to AI used "in any of its business functions," and the note flags that change. The Survey of Business Uncertainty figure is a share of workers, not firms. The Goldman Sachs figure is from participants in its own program, a self-selected group. The Atlanta Fed survey's "smaller firms" are as defined by the survey's authors, who don't give the cut-off in the blog post.

On the experiments. The Noy and Zhang figures are as reported by MIT News. The Dell'Acqua figures are those reported in the working paper's abstract and by The Harvard Crimson. The working paper's quality result is reported differently across versions, so we've left it out. Both studies used short professional tasks with 2023-era models.

On the Air Canada decision. The quotes are from paragraphs 15, 27, and 28 of the tribunal's reasons, and the award from paragraph 44. The tribunal is a small-claims body in British Columbia, and its decisions don't bind courts elsewhere.

On the electrical contractor. It's an illustrative composite, not a client. Its figures (about 40 electricians, four project managers, three estimators) are representative, not measured.

On first-party claims. Descriptions of how the Avolis diagnostic works describe our own method. They are not independent research and are not offered as benchmarks.

---

About Avolis Research Group

Avolis Research Group is Avolis's in-house research practice, focused on how operations-heavy small and mid-sized businesses actually adopt AI. It synthesizes primary economic research, government survey data, and results from real implementations into practical, vendor-neutral guidance.

More about Avolis and how we work · Get in touch