Short answer: To implement AI in a small business, name one person who owns it, set rules for what data goes into the tools, pick one workflow and measure how long it takes today, run a pilot with a person checking every output, then decide on the numbers whether to roll it out, fix it or stop. That sequence maps onto the four functions of NIST's AI Risk Management Framework (Govern, Map, Measure, Manage) and the generative AI principles Canada's privacy regulators published in December 2023.
Most owners I talk to have already tried AI. Someone in the office uses ChatGPT for emails, the estimator has played with a takeoff demo, and the accounting software switched on an assistant nobody asked for. That isn't implementation. It's a handful of experiments with no owner, no baseline and no decision at the end.
In my experience, experiments without an owner or a baseline rarely turn into results. In ServiceTitan's 2026 State of AI in the Trades survey of 1,017 U.S. residential and commercial trades contractors, 44% said understanding how to use AI in their business was a barrier and 43% pointed to a lack of clear use cases. In the same survey, high purchase cost ranked last among the barriers measured, at 17%. Cost is the smallest worry. Knowing what to do with the tools is the bigger one.
This is the sequence I use. I'm Scott Jarvis, I run Jarvis Strategies, based in Vernon, BC, and I spent 18 years in construction, including project management, coordination and estimating. The steps below work for a 10-person contractor or a 150-person shop. They don't need an IT department.
What do the frameworks say about implementing AI?
Two documents are worth knowing, and neither is long on theory. NIST's AI Risk Management Framework, released January 26, 2023, is a U.S. government framework intended for voluntary use. It sorts the work into four functions: Govern (build a culture of managing AI risk, which runs through everything else), Map (establish the context to frame risks related to an AI system), Measure (analyze, assess and monitor the risks) and Manage (put resources against the risks you've mapped and measured, on a regular basis). NIST added a Generative AI Profile, NIST AI 600-1, on July 26, 2024, and notes that the framework is now being revised.
On the Canadian side, the federal Privacy Commissioner and the provincial and territorial privacy regulators published joint principles for generative AI on December 7, 2023. The document says it is meant for organizations developing, providing or using generative AI, so it covers a contractor using ChatGPT as much as the company that built it. It also says many of its considerations will be required for an organization to comply with applicable privacy law. In other words, much of it reflects what privacy law already requires.
You don't need to read either document cover to cover. The seven steps below put both into the order a small firm actually works in.
How do you implement AI in a small business, step by step?
Start with an owner and rules, then one workflow, then a measured pilot, then a decision. The order matters more than the tool.
Step 1. Name one owner (Govern)
One person is accountable for AI in the business. In a small firm that's usually the owner or the operations lead, not the most enthusiastic user. Their job is to approve which tools get used, keep the rules current and make the final call on each pilot. The privacy regulators put it plainly: accountability for decisions rests with the organization, not with any automated system. My advice: put one person's name on it.
Step 2. Write the data rules before anyone pastes anything (Govern)
Write one page that answers three questions. Which tools are approved? What may never go into them? Who checks output before it leaves the building? The privacy principles say that where possible and reasonable, you should use anonymized or de-identified information in prompts rather than personal information, and only enter personal or sensitive information where that's authorized. My rule for a contractor: client names, home addresses, employee records and pricing history stay out of free chatbot accounts unless the owner has approved the tool for that data and you're sure the use is permitted under privacy law.
Step 3. Take stock of what's already switched on (Map)
List the software you already pay for and note which ones have added AI features. Your accounting, email and estimating platforms may already be doing more than you think. I walked through that inventory in the small business trust gap post, so I won't repeat it here. The point for implementation is that the cheapest first tool is often one you already own.
Step 4. Pick one workflow and time it as it runs today (Map)
Pick one job that repeats every week and annoys people: assembling a proposal, writing up site notes, answering the same supplier questions. If you're not sure which one, Where Do You Actually Start? covers how to find it. Then write down the baseline before you touch a tool. How long does it take now? How often does it come back for corrections? Who does it? Without that number, nobody can say later whether AI helped.
Step 5. Choose a tool that fits the systems you run (Map)
Fit beats features. In the ServiceTitan survey, 40% of contractors named integration with existing software as a barrier, and 61% of AI users said they'd prefer AI and automation built directly into their core platform. Before you buy something new, check whether the software you already run can do the job. If you do need a new tool, test it on your own past documents, not the vendor's demo files.
Step 6. Run a pilot with a person at the checkpoint (Measure)
Set the pilot by volume, not by mood: the next 10 proposals, or every daily report for one project. The tool drafts, a named person reviews, and nothing goes to a client, a sub or an inspector without that review. Log three things on every run: time taken, whether the output needed correcting and what kind of mistake it made. That's the Measure function in plain terms.
Step 7. Decide: roll out, fix or stop (Manage)
At the end of the pilot the owner makes one of three calls, on the log rather than on how people feel about the tool. Roll it out and write it into the procedure and the training. Fix it and run a second, shorter pilot. Or stop, and record why, so nobody buys the same tool again next year. Then pick the next workflow and go back to step 4.
A one-page AI implementation record
This is the sheet I'd keep for every workflow you try. It fits on one page and answers the questions a partner, a client or an insurer is likely to ask later.
- Workflow: what the job is, in one line, and who does it today.
- Owner: the person who approves the pilot and makes the final call.
- Baseline: hours per occurrence and how often it's corrected, measured before the pilot.
- Tool and data: which tool, which data it may see, and which data it may not.
- Checkpoint: who reviews each output and what they check for.
- Pilot size: how many runs, and the date it ends.
- Result: hours per occurrence and corrections during the pilot, against the baseline.
- Decision: roll out, fix or stop, with one sentence on why.
How do you know if the implementation worked?
It worked if the measured time came down without corrections going up, and if the saved hours went somewhere useful. Time saved and money made are not the same thing. In the ServiceTitan survey, 66% of contractors using AI said they save at least three hours a week, but fewer than half reported improved profitability (47%) or increased revenue (43%). Bluevine's 2026 survey of 942 U.S. small business owners found a related gap: 52% of those using AI reported a return on their investment, and 24% said they hadn't seen one yet.
My read on that gap is that saved hours leak unless someone decides where they go. If the estimator gets four hours back a week, decide up front whether that buys more bids, better follow-up or a shorter week. Write that on the implementation record next to the result.
A three-question stop test
Run this at the end of every pilot. If any answer is no, don't roll the tool out yet.
- Is it faster? Time per run, including the review, is lower than the baseline.
- Is it as accurate? Corrections are no higher than before, and no mistake reached a client.
- Can someone else run it? A second person can follow the procedure without the person who set it up standing beside them.
Where AI doesn't belong in an implementation
Some decisions stay with people no matter how good the tool gets. The final price on a bid and the risk you're pricing in are judgment calls, which I cover in AI for Construction Estimating. So are hiring, discipline and anything else that decides something about a person. For those decisions, the privacy regulators say to clearly tell any affected party whether a generative AI tool will be used as part of a decision-making process, and to consider whether its use is necessary and proportionate. Safety sign-offs, inspection records and anything you'd defend in a dispute need a human signature that means the person actually checked.
Trust is the other limit. Only 11% of contractors in the ServiceTitan survey placed high or complete trust in the AI tools they use, and 58% reported little or no trust. That isn't a reason to stop. It's a reason to keep the checkpoint in step 6 until your own log shows the tool has earned a lighter review.
If you'd rather run this with help, a paid pilot follows the same sequence on one workflow. For construction-specific workflows, the construction page and proposal turnaround solution show where this usually starts. How long each step takes depends on scope, which How Long Does AI Implementation Take? breaks down.
Key facts
- NIST's AI Risk Management Framework was released January 26, 2023, is voluntary, and organizes AI risk work into four functions: Govern, Map, Measure and Manage.
- Canada's federal, provincial and territorial privacy regulators published generative AI principles on December 7, 2023, covering organizations that develop, provide or use generative AI.
- In ServiceTitan's 2026 survey of 1,017 U.S. trades contractors, 44% named understanding how to use AI as a barrier, 43% a lack of clear use cases and 17% high purchase cost.
- In the same survey, 66% of AI users save at least three hours a week, but 47% report improved profitability.
- In Bluevine's 2026 survey of 942 U.S. small business owners, 52% of AI users reported a return on investment and 24% had not seen one yet.
FAQ
What is the first step to implementing AI in a business?
Name one person who owns AI in the business and write a one-page rule on which tools are approved and what data may never go into them. Do that before anyone picks a tool. This sits under what NIST's AI Risk Management Framework calls the Govern function, and Canada's privacy regulators say accountability for decisions rests with the organization, not the software.
How do I choose which workflow to try AI on first?
Pick one job that repeats every week and frustrates people, such as assembling proposals or writing site notes, and time how long it takes today. Without that baseline, you can't tell later whether AI helped.
How do I measure whether AI is working?
Log time per run, including review, and how often the output needs correcting, then compare both against the baseline you took before the pilot. Also decide where the saved hours go. In ServiceTitan's 2026 trades survey, 66% of AI users saved at least three hours a week but only 47% reported improved profitability.
Can I put client or employee information into ChatGPT?
Only if the owner has approved that tool for that data and the use is authorized under privacy law. Canada's privacy regulators say to use anonymized or de-identified information in prompts where possible and reasonable, and to enter personal or sensitive information only where that is authorized.
Do I need an IT department to implement AI?
No. A small firm needs an owner, a one-page data rule, one measured workflow and a person who reviews every output during the pilot. Start with AI features in software you already pay for before buying something new.
