[{"data":1,"prerenderedAt":377},["ShallowReactive",2],{"navigation":3,"/blog/023-the-3-levels-of-ai-adoption":102,"/blog/023-the-3-levels-of-ai-adoption-surround":373},[4],{"title":5,"path":6,"stem":7,"children":8,"page":101},"Blog","/blog","blog",[9,13,17,21,25,29,33,37,41,45,49,53,57,61,65,69,73,77,81,85,89,93,97],{"title":10,"path":11,"stem":12},"Why Most Small Businesses Don’t Need Digital Transformation — They Need Clarity","/blog/001-no-digital-transformation","blog/001-no-digital-transformation",{"title":14,"path":15,"stem":16},"Building the Right Thing First: A Senior Architect’s Approach to Small Projects","/blog/002-building-the-right-thing-first","blog/002-building-the-right-thing-first",{"title":18,"path":19,"stem":20},"Practical AI for Real Businesses (Specific Recommendations That Actually Work)","/blog/003-practical-ai-for-real-businesses","blog/003-practical-ai-for-real-businesses",{"title":22,"path":23,"stem":24},"Practical Technology Guidance for Small Businesses in North Texas","/blog/004-practical-technology-guidance","blog/004-practical-technology-guidance",{"title":26,"path":27,"stem":28},"Modern SEO — Optimizing for Humans and AI","/blog/005-modern-seo-optimizing-for-humans-and-ai","blog/005-modern-seo-optimizing-for-humans-and-ai",{"title":30,"path":31,"stem":32},"The Hidden Cost of “Cheap” Software for Small Businesses — When Saving Money Becomes Expensive","/blog/006-the-hidden-cost-of-cheap-software","blog/006-the-hidden-cost-of-cheap-software",{"title":34,"path":35,"stem":36},"Modernization Without Mayhem — Replacing Legacy Systems Incrementally","/blog/007-modernization-without-mayhem","blog/007-modernization-without-mayhem",{"title":38,"path":39,"stem":40},"Small Studio, Serious Work — How Boutique Teams Deliver Government-Grade Software","/blog/008-small-studio-serious-work","blog/008-small-studio-serious-work",{"title":42,"path":43,"stem":44},"When a Small Business Should (and Shouldn’t) Build Custom Software","/blog/009-when-to-build-custom-software","blog/009-when-to-build-custom-software",{"title":46,"path":47,"stem":48},"Common AI Uses for Small Business — Tools That Actually Work","/blog/010-common-ai-uses-small-business","blog/010-common-ai-uses-small-business",{"title":50,"path":51,"stem":52},"What Government Software Projects Fail At — And How to Avoid Repeating Those Mistakes","/blog/011-government-software-failures","blog/011-government-software-failures",{"title":54,"path":55,"stem":56},"The High Cost of Cool — Why Boring Tech Wins","/blog/012-high-cost-of-cool","blog/012-high-cost-of-cool",{"title":58,"path":59,"stem":60},"Decomposing Legacy Monoliths — A Practical Strategy","/blog/013-decomposing-legacy-monoliths","blog/013-decomposing-legacy-monoliths",{"title":62,"path":63,"stem":64},"The State of AI Augmentation — Doing More With The Team You Have","/blog/014-state-of-ai-augmentation","blog/014-state-of-ai-augmentation",{"title":66,"path":67,"stem":68},"Co-Intelligence: Working With AI Instead of Just Using It","/blog/015-entering-era-of-cointelligence","blog/015-entering-era-of-cointelligence",{"title":70,"path":71,"stem":72},"Beyond Shooting From the Hip: Designing a Content Strategy That Scales","/blog/016-scaling-digital-content-strategy","blog/016-scaling-digital-content-strategy",{"title":74,"path":75,"stem":76},"Fewer, Better Local Relationships — How to Choose Community Partners That Actually Help Your Business","/blog/017-local-resources-community-impact","blog/017-local-resources-community-impact",{"title":78,"path":79,"stem":80},"How Small Technical Teams Turn Delivery Problems Into Shared Capability","/blog/018-level-up-technical-teams","blog/018-level-up-technical-teams",{"title":82,"path":83,"stem":84},"Document the Parts of Your System People Need to Operate and Hand Off","/blog/019-modern-software-documentation","blog/019-modern-software-documentation",{"title":86,"path":87,"stem":88},"Your Software Has No Owner's Manual -- That's Why Your Team Is Afraid to Touch It","/blog/020-software-has-no-owners-manual","blog/020-software-has-no-owners-manual",{"title":90,"path":91,"stem":92},"The Hidden Cost of \"We'll Figure It Out Later\" Architecture — You Already Have Architecture. It's Just Accidental.","/blog/021-hidden-cost-of-well-figure-it-out-later-architecture","blog/021-hidden-cost-of-well-figure-it-out-later-architecture",{"title":94,"path":95,"stem":96},"Skilling Up vs Calling an Expert — Teach Repeat Work, Buy Judgment for High-Risk Work","/blog/022-skilling-up-vs-calling-an-expert","blog/022-skilling-up-vs-calling-an-expert",{"title":98,"path":99,"stem":100},"The 3 Levels of AI Adoption — And Where You Probably Are","/blog/023-the-3-levels-of-ai-adoption","blog/023-the-3-levels-of-ai-adoption",false,{"id":103,"title":98,"author":104,"body":108,"date":364,"description":365,"extension":366,"image":367,"meta":368,"minRead":369,"navigation":370,"path":99,"seo":371,"stem":100,"__hash__":372},"blog/blog/023-the-3-levels-of-ai-adoption.md",{"name":105,"avatar":106},"Bo Clifton",{"src":107,"alt":105},"bo-avatar.jpg",{"type":109,"value":110,"toc":352},"minimark",[111,115,118,121,124,127,130,133,138,141,144,147,150,153,156,159,162,165,169,172,175,178,181,184,187,190,193,197,200,203,206,209,225,228,231,235,238,241,244,247,250,254,257,260,280,283,286,290,293,296,299,302,305,308,311,314,317,320,323,326,330,333,336,339,342,345,349],[112,113,114],"p",{},"If people in your company use ChatGPT or Copilot every day, that tells you something. It doesn't tell you the business has adopted AI in any meaningful operational sense.",[112,116,117],{},"It tells you people have a new tool. That's not the same as changing how the work gets done.",[112,119,120],{},"A better question is simpler: where does the model sit in the work?",[112,122,123],{},"If it lives in personal tabs and copy-paste habits, you're at one level. If it sits inside a repeatable workflow with an owner, review step, and fallback path, you're at another. If it's wired into a product or operating system people depend on, that's another level again.",[112,125,126],{},"That framework is useful because it keeps you honest. Most SMBs aren't choosing between \"no AI\" and \"full AI.\" They're deciding which tasks are safe enough, repetitive enough, and valuable enough to put through a model.",[112,128,129],{},"You should also expect mixed levels inside the same company. Sales may be at Level 1. Support may have one solid Level 2 workflow. Your product may have one narrow Level 3 feature. That's normal.",[112,131,132],{},"And if people are pasting client notes, invoices, or source code into whatever tool they like, fix that before you talk about maturity. That's a data handling problem, not an AI strategy.",[134,135,137],"h2",{"id":136},"level-1-toy","Level 1: Toy",[112,139,140],{},"\"Toy\" can sound dismissive, but it doesn't have to be. Here it means one thing: the model is still outside your operating process.",[112,142,143],{},"People use it on their own. The work starts and ends with them. There's little shared process, little measurement, and usually no audit trail.",[112,145,146],{},"This level is good for learning.",[112,148,149],{},"A salesperson can use ChatGPT to draft a follow-up email after a call. A recruiter can summarize interview notes before writing feedback. A developer can use Copilot to speed up boilerplate. An office manager can rewrite a policy draft into plainer language.",[112,151,152],{},"That's real value. You shouldn't sneer at it. You should just name it correctly.",[112,154,155],{},"Level 1 works when the task is personal, the stakes are low, and a bad answer only costs a few minutes of cleanup.",[112,157,158],{},"It fails when personal use starts to masquerade as process. One employee gets useful results. Another gets junk. Someone pastes sensitive material into an unapproved tool. Nobody knows what was reviewed and what was trusted too quickly. The business gets convenience without control.",[112,160,161],{},"Your decision rule here is straightforward: keep a task at Level 1 only if it's occasional, low-risk, and easy for one person to check before it goes anywhere important.",[112,163,164],{},"Your stop condition is just as clear. If the same AI-assisted task happens every week across a team, or the output is being copied into tickets, CRM records, proposals, or customer messages, it shouldn't stay personal. Move it into a defined workflow or stop using AI for that task.",[134,166,168],{"id":167},"level-2-workflow","Level 2: Workflow",[112,170,171],{},"This is where most useful AI work belongs.",[112,173,174],{},"At Level 2, the model is one step inside a repeatable process. The workflow has a clear input, an expected output, a named owner, a review point, and a fallback path when the model is wrong or unavailable.",[112,176,177],{},"That sounds less exciting than \"AI transformation.\" Good. It's also where the value usually becomes real.",[112,179,180],{},"You should look for work that's frequent, boring, and structured enough to evaluate. Good candidates often include support triage, CRM note generation from recorded calls, invoice data extraction into a review queue, internal knowledge search, and draft responses for common service requests.",[112,182,183],{},"You don't need a giant platform to do this. In many SMBs, Level 2 starts with features already inside the systems you use, such as CRM or helpdesk AI in HubSpot, Salesforce, Zendesk, Intercom, or similar tools. In other cases, it starts with a simple automation layer in Zapier, Make, or Power Automate. The important part isn't the product name. The important part is that the workflow is explicit.",[112,185,186],{},"A human review queue should be boring and concrete.",[112,188,189],{},"If a support ticket comes in, the model can classify it, suggest a priority, draft a response, and fill a few fields. Then the ticket goes to an \"AI Review\" queue. An agent sees the original message, the draft, the suggested tags, and the source material the draft used. The agent can approve, edit, reject, or escalate. Rejected items get tagged so you can see why the model failed. If the AI step is down, the team handles the ticket the old way.",[112,191,192],{},"That's a workflow. A chatbot tab is not.",[134,194,196],{"id":195},"a-plain-end-to-end-example","A Plain End-to-End Example",[112,198,199],{},"Take a small service business that gets eighty support emails a day.",[112,201,202],{},"The inbox lives in a helpdesk system. Each new email triggers a workflow. The model reads the message, compares it against the approved knowledge base, suggests a category, drafts a reply, and flags anything that looks like billing, cancellation, refund, or legal risk.",[112,204,205],{},"Routine password-reset and scheduling questions land in an \"AI Draft Review\" queue. Agents approve or edit the draft before sending. Billing disputes, cancellation requests, refund demands, and anything that looks like a legal threat skip the AI reply step and go straight to a human.",[112,207,208],{},"You then measure four simple things for two weeks:",[210,211,212,216,219,222],"ul",{},[213,214,215],"li",{},"first-response time",[213,217,218],{},"percentage of drafts approved with light edits",[213,220,221],{},"reopen rate",[213,223,224],{},"escalations caused by bad drafts or bad categorization",[112,226,227],{},"If response time improves and reopen rates stay steady, the workflow may be earning its keep. If agents rewrite most drafts or customers keep coming back because the answers miss the point, stop and fix the process. If the process never gets clean enough, remove the AI step.",[112,229,230],{},"That's how you evaluate Level 2 plainly. You don't ask whether people \"liked the AI.\" You ask whether the workflow became faster or better without creating more rework, risk, or confusion.",[134,232,234],{"id":233},"where-level-2-breaks","Where Level 2 Breaks",[112,236,237],{},"Level 2 is strong, but it's not magic.",[112,239,240],{},"It usually fails for boring reasons. The process is messy. Nobody agrees on exceptions. The inputs are inconsistent. Review responsibility is vague. The model is working from stale documentation. Or the team never measures whether the output is actually helping.",[112,242,243],{},"You should also be careful about the shape of the work. If a task depends on unstated context, frequent judgment calls, or many edge cases, the model may create more cleanup than value. That's common in complicated quoting, sensitive HR matters, and support cases where a customer is already upset.",[112,245,246],{},"Your decision rule at Level 2 is this: use AI when the work is repeatable, the output can be checked quickly, and mistakes can be caught before anything final happens.",[112,248,249],{},"Your stop condition is this: if reviewers must substantially rewrite most outputs, if exceptions dominate the queue, if you can't define allowed data and ownership, or if there's no clean fallback path, stop. Keep the workflow manual until the process is better.",[134,251,253],{"id":252},"some-work-should-stay-out-of-scope","Some Work Should Stay Out of Scope",[112,255,256],{},"You should set a few hard boundaries early and keep them hard.",[112,258,259],{},"In most SMBs, AI shouldn't be allowed to do any of the following without explicit human approval:",[210,261,262,265,268,271,274,277],{},[213,263,264],{},"approve or release payments",[213,266,267],{},"send legal commitments or contract terms",[213,269,270],{},"change customer pricing, discounts, or credit limits",[213,272,273],{},"cancel accounts, issue refunds, or take other hard-to-reverse customer actions",[213,275,276],{},"make employment decisions",[213,278,279],{},"send regulatory or compliance statements in the company's name",[112,281,282],{},"AI can help prepare, summarize, route, or flag this work. It shouldn't take the final action on its own.",[112,284,285],{},"That boundary matters more than prompt quality. It's an operating rule, not a tooling preference.",[134,287,289],{"id":288},"level-3-system","Level 3: System",[112,291,292],{},"Level 3 isn't \"AI everywhere.\" It's AI inside a system your business depends on.",[112,294,295],{},"For an SMB, that may be smaller than it sounds. It might be AI search inside your customer portal. It might be dispatch suggestions inside field-service software. It might be anomaly detection in an operations dashboard. It might be product-side answer suggestions that help users find the right article or next step.",[112,297,298],{},"For many SMBs, a narrow Level 3 feature is enough. You don't need to rebuild your business around AI to use this level well.",[112,300,301],{},"The point isn't scale for its own sake. The point is that the model is now part of a live application or operating surface, not an optional helper.",[112,303,304],{},"You should only move here when three things are true.",[112,306,307],{},"First, the underlying workflow already works. If Level 2 is sloppy, Level 3 will just make the sloppiness faster and harder to unwind.",[112,309,310],{},"Second, the volume or speed requirement actually justifies deeper integration. If a human can review every case without pain, you may not need Level 3 at all.",[112,312,313],{},"Third, you can contain failure. That means logging, rollback, monitoring, manual override, and a clear owner.",[112,315,316],{},"Level 3 also needs firmer entry criteria. If the business case depends on the model being correct almost all the time, but you can't measure that in production, don't ship it. If you can't explain what happens when the model is uncertain, don't ship it. If you can't turn the feature off without breaking customer operations, don't ship it.",[112,318,319],{},"In many cases, the right Level 3 role is advisory, not decisive. A model can suggest a route for a dispatch team. It can surface likely answers in a portal. It can rank likely matches in search. It shouldn't quietly change customer obligations, move money, or take irreversible actions because its confidence score looked respectable that day.",[112,321,322],{},"Your decision rule at Level 3 is simple: move here only when the workflow is already stable, the economics justify engineering effort, and failure can be detected and reversed.",[112,324,325],{},"Your stop condition is even simpler: if you can't monitor it, roll it back, or keep a person in front of irreversible actions, don't embed it in the system.",[134,327,329],{"id":328},"use-the-framework-honestly","Use the Framework Honestly",[112,331,332],{},"The most common mistake isn't starting too late. It's naming things too generously.",[112,334,335],{},"If five employees use ChatGPT well, that doesn't mean the company has operational AI. It means five employees found a useful tool. That can still be good. It just shouldn't be confused with workflow change or system design.",[112,337,338],{},"You'll usually make better decisions if you ask smaller questions.",[112,340,341],{},"Which repetitive task is worth tightening? Which queue needs review rules? Which data is allowed? What metric should improve? What would make you turn the AI step off?",[112,343,344],{},"Those are operator questions. They lead to better outcomes than maturity theater.",[134,346,348],{"id":347},"keep-one-standard","Keep One Standard",[112,350,351],{},"If an AI-assisted step can move money, change a customer's rights or obligations, or trigger an irreversible action, require explicit human approval and a recorded audit trail before it happens.",{"title":353,"searchDepth":354,"depth":354,"links":355},"",2,[356,357,358,359,360,361,362,363],{"id":136,"depth":354,"text":137},{"id":167,"depth":354,"text":168},{"id":195,"depth":354,"text":196},{"id":233,"depth":354,"text":234},{"id":252,"depth":354,"text":253},{"id":288,"depth":354,"text":289},{"id":328,"depth":354,"text":329},{"id":347,"depth":354,"text":348},"2026-04-19T00:00:00.000Z","Most companies are still using AI as a personal tool, not an operational capability; here is how to tell whether AI is still a toy, a workflow, or part of the system that runs your business.","md","https://images.pexels.com/photos/8204309/pexels-photo-8204309.jpeg?auto=compress&cs=tinysrgb&w=1260&h=750&dpr=1",{},8,true,{"title":98,"description":365},"Rr1NOPjDVJTX-kJ0u3Q6F4DZBqaT_yqM0SCeJ2bHusg",[374,376],{"title":94,"path":95,"stem":96,"description":375,"children":-1},"You should build internal skill for repeatable operational work, bring in specialists when mistakes are expensive, and use a hybrid handoff when you need both judgment and ownership.",null,1776639425626]