How to add AI to your business, the short answer
Pick one repetitive, measurable job, try AI on it with the tools you already pay for, and only build custom software once the numbers show it pays. That is the whole method for adding AI to a business. Most of the money people lose on AI comes from skipping one of those three steps, usually by buying a platform before choosing a job, or by building before measuring anything.
I'm a forward deployed engineer, which means I build AI into small companies' own systems and stay until it works on real data. I wired five AI agents and usage-based billing into the Tryneth backend, and built a sales CRM where Claude drafts every outreach email and a person approves it before it goes out. What follows is the plan I give owners who ask how to add AI to business operations they already run, with the costs I actually see on projects.
Most small businesses have not added AI yet
If you feel late, the data says otherwise. The US Census Bureau's Business Trends and Outlook Survey put AI use at 19.8 percent of American businesses as of 3 May 2026. Among firms with four or fewer employees it was under 20 percent, and for firms with fewer than 20 employees the rate barely moved between December 2025 and May 2026. Larger companies are further along, with 37 percent of firms with 250 or more employees using AI (Census Bureau, May 2026).
The NFIB's technology survey tells the same story from the owner's side. Nearly a quarter of small business owners used AI tools such as ChatGPT, Copilot, Grammarly or Canva, rising from 21 percent of firms with under 10 employees to 48 percent of firms with 50 or more (NFIB, June 2025). Among the owners already using AI, 30 percent said it raised productivity, and 98 percent said it had not changed their headcount.
Why so many AI projects still go nowhere
Adoption is slow partly because early results were poor. MIT's Project NANDA reported in 2025 that only about 5 percent of corporate generative AI pilots produced rapid revenue growth, while most stalled with little measurable effect on profit and loss (Fortune's summary of the study). The sample was small and the bar was narrow, so read it as a direction rather than a precise rate. Its main finding still matches what I see on client work. Projects failed at the workflow, not at the model. Tools bought from specialist vendors or built with a partner reached production about 67 percent of the time, far more often than tools companies built entirely alone.
Three levels of adding AI to a business
Every AI integration for business I have worked on falls into one of three levels. Start at the lowest level that can do the job, because each step up costs more and needs more care.
Level 1, switch on the AI in tools you already pay for
The cheapest way to use AI in your business is the assistant already sitting in your software. Microsoft 365 Copilot Business lists at about $21 per user a month, ChatGPT Business and Claude Team start at around $20 to $25 a seat, and Google Workspace business plans now include Gemini. Accounting, CRM and store platforms such as QuickBooks, HubSpot and Shopify ship their own assistants too. This level covers writing, meeting notes, summaries, spreadsheet formulas and first drafts of almost anything.
The risk at this level is rarely the tool. It's staff pasting customer data into personal free accounts that may train on it. Buy the business plan, switch off training on your data where the setting exists, and write down what may and may not go into a chat window.
Level 2, connect AI between your apps with automation
Automation platforms such as Zapier, Make and n8n let you put an AI step in the middle of a workflow. A new enquiry arrives, the AI classifies it and drafts a reply, and the draft lands in your CRM or help desk for a person to send. Zapier's Professional plan starts at roughly $20 to $30 a month, though AI steps use the same task allowance as everything else, so busy workflows climb fast.
Level 2 is where many small businesses get their first real time savings. It is also where things quietly break. A renamed form field or a changed API can stop a workflow for days before anyone notices, and every step sends your data through one more vendor. Once you have more than a handful of AI workflows touching customers, somebody needs to own them.
Level 3, build AI into your own software
At Level 3 your own backend calls a model through its API, reads your database and documents with proper permissions, and acts inside your systems. This is how you add AI to existing software without rebuilding it. A support assistant that answers from your help centre and the customer's order history belongs here. So does an agent that prepares quotes from your price list, or a pipeline that turns supplier invoices into accounting entries.
Model usage is cheaper than most people expect. Take an assistant answering 1,000 questions a day, each sending about 3,000 tokens of context and getting 400 tokens back. On a small fast model priced near $1 per million input tokens and $5 per million output tokens, that comes to roughly $150 a month, and a mid-tier model lands nearer $450. Prompt caching cuts the input side a lot, which matters because a careless prompt can triple the bill. The build is the bigger cost. My AI projects start at $5,000, while agencies usually quote $20,000 to $80,000 for a small custom AI build.
| Level | What it looks like | Typical cost | First result | Main risk |
|---|---|---|---|---|
| 1. AI in tools you already use | Copilot, ChatGPT Business, Claude Team, Gemini in Workspace, or the assistant inside your CRM or accounting app | About $20 to $30 per user a month | Same day | Staff pasting customer data into personal accounts |
| 2. AI inside your automations | Zapier, Make or n8n with an AI step that sorts, extracts or drafts | From about $20 a month, plus model usage | A few days | Workflows that break silently and that nobody owns |
| 3. AI built into your software | Your backend calls a model API with your data, your permissions and proper logging | From $5,000 to build, plus $150 to $450 a month in usage for a busy assistant | Days to two weeks for a working version | Building before the workflow is proven |
Where to add AI first in a small business
The best first project is boring. It happens many times a week, it is mostly text, a person can check the output in seconds, and a mistake costs little. Score your candidates against those four tests and the right one usually stands out. These are the places where I see small businesses get a return first.
- Inbox and enquiry triage. AI reads each incoming message, tags it by type and urgency, and drafts a reply from your templates. A person still presses send.
- Customer support answers. An assistant answers routine questions from your own help pages and policies, then hands anything unusual to a person along with a summary of the conversation so far.
- Sales follow-up. In the Leads Oracron CRM, a scraper audits each prospect's website and Claude drafts an outreach email from the findings. The sales team previews every draft before it sends, so the pace goes up without the brand voice drifting.
- Documents and data entry. Invoices, receipts, purchase orders and forms become structured fields in your accounting or inventory system, with low-confidence fields flagged for a person to check.
- Internal knowledge search. Staff ask questions in plain English and get answers from your procedures, contracts and past projects, with a link back to the source document.
- Weekly reporting. A summary of sales, support volume and overdue invoices, drafted from your CRM and accounting data before Monday's meeting.
Where the AI shows up matters as much as what it does. A feature buried in a separate tab gets ignored by the second week, which is the point of my note on shipping AI features that actually help.
Where not to start
Keep AI away from setting prices, approving refunds, giving legal or medical advice and screening job applicants until you have run it somewhere safer first. Hiring and credit decisions also count as high-risk under the EU AI Act, with strict rules starting on 2 December 2027. My guide to the EU AI Act for small businesses covers which features fall where.
A 90-day plan to add AI to your business
Ninety days is long enough to get one workflow live and measured, and short enough that nobody loses interest. The plan assumes one owner on your side, usually the founder or an operations lead, with an hour or two a week to spare.
Days 1 to 30, pick one workflow and measure it
- List the repetitive jobs your team does every week and score each one against the four tests above.
- Pick one. Write down how many times it happens each week and how many minutes it takes today, because that is your baseline.
- Collect a few dozen real examples, including the awkward ones. They become your test set, and they are worth more than any vendor demo.
- Decide the data rules, meaning which customer data may go to which tool and who signs off on that.
Days 31 to 60, run a pilot on real cases
Start at Level 1 or Level 2 if the job allows it. Run the AI on real work with a person approving every output, and log each mistake in a simple spreadsheet with a note on what went wrong. By the end of the month you should know three numbers. You'll know how long a task takes now, how often the AI output needed fixing, and what the tools cost.
Days 61 to 90, harden it or stop it
If the numbers are good, make the workflow production grade. That means access limited to what the job needs, logging, spending caps, a fallback to a person and a named owner who checks it every week. If the numbers are poor, stop, write down why, and move to the next workflow on your list. Stopping a pilot after two months is a good outcome compared with the usual alternative, a demo that stays almost ready for a year. I wrote about that trap in why AI pilots stall before production.
Guardrails that keep AI from hurting the business
Most AI damage to small businesses comes from the plumbing around the model, not from the model itself. Customer data ends up in the wrong place, a bot answers something it should have passed on, or an API bill arrives that nobody expected. Five rules prevent most of it.
- Know where the data goes. Use business plans with a data processing agreement, and check where each provider stores and keeps prompts. If you serve customers in Europe, my notes on which LLM providers store EU customer data will save you a few evenings.
- Give AI the least access that works. Let it read before it writes, and route its actions through your own backend instead of handing it database credentials. On Tryneth that meant row-level security on 97 of 98 production tables. Skipping this step is how thousands of quickly built apps exposed their databases this year.
- Keep a person in the loop at first. Anything that reaches a customer, moves money or changes a record gets a human approval step until your error log says otherwise.
- Tell people when they are talking to AI. For customers in the EU, Article 50 of the AI Act has required this since 2 August 2026, and it is simple good manners everywhere else.
- Cap and watch the spend. Set monthly limits with each provider and alert on unusual usage. On one production endpoint I cut about 10 million uncached tokens, and the saving showed up on the very next bill.
How to tell if AI is paying off
Measure it in hours and errors. Here is a worked example with round numbers. A small shop gets 400 customer emails a month and each reply takes six minutes, which adds up to 40 hours of work. With AI drafting, each email needs about two minutes of checking and editing, so the same inbox takes around 13 hours. Twenty-seven hours saved at $25 an hour is roughly $670 a month, against perhaps $40 in automation and model costs. That is an easy yes.
Run the same sum for your own pilot, then add the costs people forget. Somebody has to own the workflow, prompts drift as your products change, and models get retired. Budget 15 to 20 percent of the build cost per year for maintenance on anything you built at Level 3. Alongside time saved, track the share of outputs a person had to fix and the share handed off entirely. Both should fall month by month, and if they don't, the workflow needs work before it needs scale.
When to bring in help, and what kind
You can handle Level 1 yourself, and most owners can manage a few simple Level 2 automations. Bring in an engineer when AI has to touch your live database, customer accounts or payments, or when your automations have grown into a tangle that only one person half understands. That is the gap a forward deployed engineer fills, with one senior person working inside your systems, building the integration and handing over code you own.
For scale, my AI and automation projects start at $5,000, a review of an existing system starts at $2,500, and a monthly retainer starts at $1,500. There is no deposit. Fixed-price work is paid in four milestones, 25% after each quarter of the work is finished and reviewed, as explained in how milestone payments work. Whoever you hire, ask to see a working preview within two weeks. If the first deliverable is a strategy deck, you are paying for the wrong thing.
Questions people ask about adding AI to a business
How do I add AI to my business with no technical skills?
Start at Level 1. Buy the business plan of one assistant, such as ChatGPT Business, Claude Team, Microsoft 365 Copilot or Gemini in Google Workspace. Use it on one documented task for a month, for example drafting replies to enquiries or first drafts of proposals. Then try one simple no-code automation. You only need technical help once AI has to read or change your live systems.
How much does it cost to add AI to a small business?
Level 1 costs about $20 to $30 per user a month. Level 2 automations start at about $20 a month plus model usage. A custom Level 3 build starts at $5,000 with me, while agencies usually quote $20,000 to $80,000, and a busy assistant then costs roughly $150 to $450 a month in model usage. Budget 15 to 20 percent of the build cost per year for maintenance.
Can I add AI to my existing software without rebuilding it?
In most cases, yes. If your software has an API or a database you control, a small backend service can call a model, read only the data it needs, and write results back through the same routes your app already uses. Rebuild only when the old system has no safe way in, or when it is due for replacement anyway.
What is the best first AI project for a small business?
Pick a job that happens many times a week, is mostly text, takes a person seconds to check, and costs little when it goes wrong. Inbox triage with drafted replies is the most common winner I see, and turning invoices or forms into structured data comes a close second.
Is it safe to put customer data into AI tools?
It can be, on the right plan. Business tiers of the major assistants do not train on your data by default and come with a data processing agreement, while free personal accounts may. Check where each provider stores prompts, limit what the AI can see to what the task needs, and keep a written rule for staff on what never goes into a chat window.
What is the 30% rule in AI?
It is an informal rule of thumb, not a standard, and people quote it in different ways. The most common version says AI can take the routine bulk of a task while people keep the last 30 percent that needs judgment and accountability. I use a simpler form of it on every project, where AI drafts and a person decides.
Will adding AI to my business mean cutting staff?
For most small businesses, no. In the NFIB survey, 98 percent of small business owners using AI said it had not changed their headcount. In practice AI takes the repetitive part of a job, so the same team handles more customers or gets time back for work that needs a person.
Want AI working in your business, not just in a demo
Do you have a job in mind but nobody to build it? Did a pilot stall and you can't tell why? Send me the task you would most like off your team's plate. I'll tell you whether AI fits it, which level makes sense and roughly what it would cost. See what hiring me looks like, or start a project today.
You can browse everything I build under services, and read how I approach AI and automation in particular. The case studies and my career page hold the proof. When you are ready, get in touch. I read and answer every message myself.