Deep Dive · Verified Sources Only
I asked five different AIs to research this article. They handed me more than a dozen statistics that fell apart the moment I checked them against their own sources. I kept count, and I kept the receipts. I have run this kind of experiment before, and it never stops being educational.
Here is why that matters in Kingsport, Johnson City, and Bristol specifically. A Vanderbilt poll this spring found that 69% of Tennesseans are uncomfortable with AI evaluating symptoms and writing treatment plans without a doctor in the room. The same year, the US Chamber's survey put Tennessee small business AI use at 46%. We do not trust this technology, and we are already using it. Both of those things are true at once, and that tension is the whole story of AI right now, here and everywhere else.
So here is the deal for the next sixteen minutes. The good news in this post is measured, not promised. The bad news is documented, not imagined. The ugly news has case numbers. And scattered through all of it are the things the AI research tools invented along the way, because showing you what they make up is more useful than telling you to be careful.
Made Up No. 1
Three different AIs gave me the same statistic: 77% of small businesses have no written AI policy. One said it came from "a policy blog." One invented a research firm called Aufsite Research. One pinned it on the US Chamber of Commerce. I went to the Chamber's actual report. Their 77% is about something else entirely, businesses saying regulation would hurt their growth. Same fake number, three fake sources. When a statistic cannot keep its own story straight, that is the tell.
The Good: the gains are real, and they favor small shops
Strip away the marketing and the strongest research on AI says something specific and useful: it helps inexperienced people the most.
The most rigorous workplace AI study to date, published in the Quarterly Journal of Economics by economists at Stanford and MIT, tracked 5,179 customer support agents after their company rolled out an AI assistant. Productivity rose 14% on average, but new and lower-skilled workers improved 34%, while the veterans barely moved. A new hire with AI performed like someone six months into the job. In plain terms: AI does not make your best person better. It makes your newest person useful faster. If you cannot afford to hire experience, that is the single most relevant finding in all of AI research.
It repeats across the literature. A randomized experiment published in Science gave ChatGPT to 453 professionals on real writing tasks: 40% faster, 18% higher quality, and the gap between strong and weak writers shrank. Three randomized trials at Microsoft, Accenture, and a Fortune 100 firm found developers with an AI coding assistant completed 26% more tasks, with junior developers gaining most.
And it works at the smallest possible scale. A peer-reviewed case study followed a one-person Slovak e-commerce shop that installed an AI chatbot: average customer response time fell from 118 minutes to 64, the bot absorbed about three quarters of routine questions, and satisfaction went up. That is a business smaller than most of the ones reading this.
Made Up No. 2 (Plot Twist)
That Slovak study reported a 14.54% satisfaction improvement. Numbers that precise are usually a hallucination wearing a lab coat, so I flagged it for the kill list and went hunting for the original. Found it. Real journal, real paper, real decimal. Verification cuts both ways: sometimes you catch the machine lying, and sometimes you owe it an apology.
Now the honest part of the good news. Who is actually using this stuff depends entirely on who is counting. The US Chamber says 58% of small businesses use generative AI, up from 23% just two years before. The Census Bureau, which only counts firms using AI in actual business operations, puts it near 20%. JPMorgan Chase, looking at what businesses pay for, found 17.7%. None of those numbers is wrong. Most "AI adoption" is somebody in the office using the free version of ChatGPT. That counts for something, but know which number you are.
One more honesty checkpoint, because hype is a disease and I have written before about the gap between AI's promises and its books: surveys of executives still find most report no measurable productivity impact from AI so far, and economists who pool hundreds of studies find the worker-level gains have not yet shown up in the aggregate economy. Both things are true. The gains are real, task-level, and concentrated where the research says they are. They are not magic, and they are not economy-wide. Anyone who tells you otherwise is selling.
Back Home
Tennessee employers match the calm version of this story. The University of Tennessee's Boyd Center surveyed business leaders statewide this winter: 63% use AI without cutting a single job, 10% reduced headcount, 23% have not started. And Tennessee small business adoption trails the national average, 46% to 58%. I read that gap as runway, not failure. The research above says the biggest gains go to whoever closes their skill gaps first, and around here, almost nobody has gone first yet. If you want the practical version of where to start, I wrote it.
The Bad: the costs are real, and they find the unverified
The failure numbers are not subtle. S&P Global found 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before, scrapping nearly half their pilot projects on average. RAND, interviewing veteran data scientists, concluded more than 80% of AI projects fail, roughly double the rate of ordinary IT projects. You have probably also seen the famous MIT finding that 95% of generative AI pilots produced no measurable financial return. That one is real but worth a footnote nobody else gives it: it came from a small, conference-recruited sample, and the bar it measured was "extracting millions," not "produced any value." Treat it as a weather report, not a thermometer reading.
Made Up No. 3 (Including by Me)
While researching those failure rates, one AI handed me an authoritative four-part breakdown of exactly why projects fail, with decimals: 33.8% abandoned, 28.4% no value, 18.1% could not justify cost. Attributed to RAND. RAND never wrote it. Somebody on the internet invented the precision, and three separate AIs repeated it back to me as research. In the same pile, an AI quoted "Tennessee State CTO Jerry Jones," and I flagged that one as an obvious fabrication, because Jerry Jones owns the Dallas Cowboys. Then I checked. Tennessee's actual Chief Technology Officer is a real man actually named Jerry Jones. The machine was right and the fact-checker was wrong. Verification cuts everybody, including the person doing the verifying, which is the strongest argument I know for doing it anyway.
Because here is where the bad news stops being abstract. I have written before about why AI invents things, but the consequences now have docket numbers. Stanford researchers tested the expensive, professional-grade legal AI tools, the ones marketed as hallucination-resistant, and found them wrong on 17% to 33% of queries. A researcher in Paris maintains a database of court decisions involving AI-fabricated material. It passed 1,600 cases, and about 90% of those decisions were written in a single year.
And the people getting burned are not who you would guess. A Stanford analysis of the US cases found that 90% of the law firms sanctioned over AI-fabricated filings were solo practices or small firms. Not the giants with armies of associates. People with no second set of eyes. If that pattern holds for lawyers, who are professionally trained to verify, it holds for every small business sending AI-drafted contracts, quotes, and client emails out the door unread.
The corporate version of the same lesson comes in two sizes. Small: a Canadian tribunal ordered Air Canada to pay $812 after its website chatbot gave a customer wrong fare advice, rejecting the airline's argument that the bot was somehow a separate entity. You own what your chatbot says. Large: Deloitte Australia refunded part of a $440,000 government contract after its 237-page report turned out to contain AI-invented citations and a quote from a court judgment that did not exist.
It even gets your time. A randomized trial of experienced software developers found they were 19% slower with AI tools while believing they had been 20% faster. The study has its caveats, but the perception gap is the durable lesson. "Feels faster" is not "is faster." Measure.
Zoom out and the quality cost shows up everywhere AI content floods in: an economic study of the book market found new releases nearly doubled while average quality fell, and watchdog audits found the major chatbots now repeat false claims about a third of the time, double the year before. I have covered the slop economy and what it is doing to trust, so I will leave those there.
Back Home
The biggest appellate punishment yet for fake citations happened in our own federal circuit. In March, the Sixth Circuit fined two Tennessee attorneys $15,000 each, plus the other side's full legal fees and double costs, over two dozen fabricated or misrepresented citations in appeals out of the Eastern District of Tennessee. Here is the part worth memorizing: the court did not bother to determine whether AI wrote the fakes. The duty to verify applies, in the court's words, however the citations were generated. A federal court two hours from here just stated this post's thesis under oath. And the root cause is the same here as everywhere: only about a third of small business AI users ever got any training, while three quarters of their bosses believe none is needed.
The Ugly: past wasteful sits criminal, and it is documented
Made Up No. 4 (The One That Wasn't)
The statistic at the center of this section, $893 million in AI-enabled fraud, was the one I was most certain was fabricated, because the FBI had never tracked AI fraud as its own category. I held it out of this post for two weeks. Then I pulled the actual 2025 Internet Crime Report PDF off the FBI's own site. It is real, page-numbered, the first AI section in the report's 25-year history. And the FBI itself calls the number an undercount. The scariest stat in this post is the one I tried hardest to kill. From here on, this article stops being funny on purpose.
The FBI's 2025 numbers: 22,364 complaints involving AI, $893 million in losses, inside a record year of over one million total complaints and $20.9 billion stolen. Americans over 60 lost $7.7 billion of that, up 37% in a single year. And the undercount is not a hedge, it is visible in the data: victims tagged AI in under 8% of investment fraud losses, a category AI-generated personas and scripts almost certainly saturate. The FBI only counts AI when the victim could tell. Increasingly, nobody can tell.
One case shows the ceiling. In early 2024, a finance employee at the engineering firm Arup joined a video conference with the company's CFO and several colleagues, got instructions, and wired $25.6 million in fifteen transfers. Every other person on that call was a deepfake, video and voice, built from public footage. Not a sloppy email. A meeting. The same year, scammers tried the identical play on the CEO of the ad giant WPP and failed, because one employee got suspicious and checked. Vigilance works. It is also now the entire defense.
At the household level the same technology runs the grandparent scam. The FTC's warning is blunt: a scammer needs a few seconds of audio from social media to clone a family member's voice, so do not trust the voice. Call the person back on the number you know. Better, agree on a code word. If the panicked voice on the phone cannot say it, hang up. It is free, it takes one dinner conversation, and it defeats a billion-dollar criminal industry.
The discrimination cases have moved from theory to settlements. The EEOC's first AI hiring case ended with a $365,000 payout after a company's software auto-rejected older applicants. A tenant-screening algorithm cost SafeRent $2.275 million for scoring Black and Hispanic applicants lower. Massachusetts extracted $2.5 million from a lender whose AI underwriting disadvantaged Black and Hispanic borrowers. A federal court has let a nationwide collective action proceed against Workday over alleged age discrimination by its screening AI, with the key ruling that the software vendor itself can be liable. If you use an AI tool to screen tenants, applicants, or borrowers, its bias is legally your bias.
The heaviest cases involve kids and chatbots, and I will keep this exact. A Florida mother sued Character.AI and Google after her 14-year-old son died by suicide following months of dependency on a companion chatbot; a federal judge ruled the bot is a product subject to liability law, not protected speech, and the companies reportedly settled this January along with related suits. A California family's suit against OpenAI alleges ChatGPT discussed suicide with their 16-year-old son over a thousand times before his death. These are lawsuits, and allegations are allegations. But forty-four state attorneys general did not wait for verdicts to formally warn the AI industry about children, and the FTC opened an inquiry into companion chatbots. If you or someone you love is struggling, the 988 Suicide and Crisis Lifeline is there by call or text, any hour.
And the information environment itself is under quiet attack. Fake "local news" websites, mass-produced and increasingly AI-generated, now outnumber America's surviving daily newspapers. One operation runs hollow local news sites in 355 towns with a single person and a bot. Researchers at Yale found people often trust these polished fakes more than real local outlets. The 2024 election's predicted deepfake apocalypse largely did not arrive, and the one infamous case, the AI-cloned Biden robocall in New Hampshire, ended with a $6 million FCC fine and, notably, a full criminal acquittal. The law here is wet cement.
Back Home
None of this is staying national. Tennessee's Attorney General led that 44-state warning to the AI industry, which makes our state the tip of the spear, not a bystander. An Elizabethton grandmother spent five months in jail after AI facial recognition software in North Dakota misidentified her in a bank fraud case; she was over 1,200 miles away the whole time, and by the time charges were dismissed she had lost her home and her car. Tennessee seniors reported $108 million in fraud losses last year, up 75%, with investment scams tripling. And in May, a Johnson City teenager was charged with dozens of exploitation counts for allegedly using an AI app to create explicit images of classmates; three of the victims are now suing xAI in federal court. Parents should know this category of harm exists, and that it exists here.
One absence worth noting: as of this writing, no fake AI-generated news site claims to cover the Tri-Cities. The national pattern says one eventually will. The time to know which local sources are real is before that day, not after.
Where the Tri-Cities actually stands
Here is our position on the map, in numbers. Wages in the Johnson City metro run about 77 cents on the national dollar. Our single biggest job category, at 16.1% of employment, is office and administrative support, which happens to be the kind of work generative AI affects most directly. At the same time, Tennessee as a whole ranks among the lower AI-exposure states in federal analyses, because manufacturing and healthcare dominate our economy and those jobs are the hardest to automate with a chatbot. Both facts are true. The region is buffered. The region's most common job is not. If you employ people who spend their day on email, scheduling, bookkeeping, and documents, AI is not a someday question for you.
And the decision has already been made up the hill. Eastman is embedding AI into its commercial, manufacturing, and innovation workflows and hiring AI engineers right here in Kingsport. Ballad Health runs at least three named AI systems across its 29-county service area, catching medication errors, powering virtual care, and planning cancer treatments. Whatever a Main Street owner decides about AI, the two largest employers in the region clocked it in a while ago.
Now the part of this story nobody else will write, because it takes living here. Researchers who did fieldwork in rural Appalachian Tennessee published a peer-reviewed study on how our communities adopt technology, and their findings will not surprise anyone raised here: reluctant adoption, rooted in self-reliance and privacy. One participant put it in six words: private business should be private business. Every consultant who ever blew through this region read that as a problem to overcome. I read it as the correct instinct, early. Look back at everything above. The Bad section is a list of what happens to trusting adopters. The Ugly section is a list of what happens to trusting people. Appalachian wariness, pointed at AI, is not a deficit. It is verification culture with an accent, and it is precisely the skill this technology demands.
We have also seen this movie. In 1930, fewer than one in ten American farms had electricity. Thirty years later, nearly all of them did, and the economists who studied it found the counties that got power early compounded the advantage for decades. The delivery mechanism was not a corporation from off; it was co-ops, neighbors, trusted institutions. Right now, federal broadband money is wiring the last unserved corners of Sullivan, Washington, Carter, and Greene counties, with buildout targeted by 2028. Same movie. New utility. Same rule: early matters, and trust decides who goes early.
The trusted institutions are already moving, quietly. Tusculum hosted the region's first Appalachian AI Summit this spring. Northeast State runs named AI courses for working people. The Small Business Development Center in Johnson City has held free AI workshops, including with the Kingsport Chamber. The infrastructure for learning this is forming twenty minutes from your shop. The problem was never that nobody teaches it here. The problem is that nobody told you.
Even the state legislates the way the region adopts: skeptically, specifically, on its own terms. Tennessee passed the nation's first law against AI voice cloning and this spring banned AI systems from posing as mental health professionals, both nearly unanimously, while the big sweeping AI regulation bills died in committee. Narrow, concrete protections against named harms. That is not a state behind on AI. That is a state that read the room.
The verdict, from Bristol and beyond
Final tally on my five research assistants: more than a dozen statistics invented, misattributed, or quietly mangled, every one of them delivered with total confidence, several repeated by multiple AIs in unison. Also: hundreds of real, checkable facts I could not have gathered alone in a month. That is the honest verdict on this technology. Wrong often enough to need a human. Right often enough to be worth one.
Two years ago, Tennessee became the first state in America to make cloning a person's voice with AI a crime. It passed the House 93 to 0 and the Senate 30 to 0, and Governor Lee signed it saying it protected Tennesseans "from Beale Street to Broadway, to Bristol and beyond." Nobody has been charged under it yet. Around here, we believe in locking the door before something gets stolen.
That is the Tri-Cities' actual AI advantage, and it is the same as our oldest habit: we do not believe things just because a stranger, a machine, or five very confident chatbots said them. Keep that habit. Aim it at this technology. The good is real if you measure it, the bad is avoidable if you verify, and the ugly mostly preys on people who trusted a voice.
Start with the free thing tonight: pick a family code word. Then, if you want help separating what AI can actually do for your business from what it only claims to, that is the entire reason Tri-Cities AI Lab exists. Human-reviewed, sources attached, no magic promised.
Questions people actually ask
What percentage of AI projects fail?
Depends on the bar. S&P Global found 42% of companies abandoned most AI initiatives in 2025. RAND puts overall project failure above 80%. The famous MIT "95% fail" figure measured whether pilots produced major financial returns, on a small sample, so treat it as directional.
Is AI actually worth it for a small business?
The peer-reviewed evidence says yes for specific tasks: routine writing, first-line customer service, and helping newer employees perform like experienced ones. The same evidence says gains require human review, and most failures trace to skipping it.
How many small businesses in Tennessee use AI?
About 46% per the US Chamber's 2025 state survey, below the 58% national figure. By the Census Bureau's stricter operational definition, roughly one in five firms nationally. A statewide UT survey found 63% of Tennessee employers use AI without reducing staff.
How big is AI scam fraud?
The FBI's 2025 Internet Crime Report tracked AI for the first time: 22,364 complaints and $893 million in losses, which the FBI itself calls an undercount. Tennessee seniors alone reported $108 million in fraud losses in 2025.
How do I protect my family from AI voice cloning scams?
Agree on a family code word. Scammers can clone a voice from a few seconds of social media audio, so if a panicked call demands money, ask for the word and call the person back on the number you already have. The FTC's guidance is the same: do not trust the voice.
The receipts
Primary sources, in order of appearance. If a number in this post is not on this list, it should not be in this post, and I want to hear about it.
- Vanderbilt Poll, Center for the Study of Democratic Institutions, May 2026 (n=1,203)
- US Chamber of Commerce, Empowering Small Business, 4th ed., Aug 2025 (advocacy survey, 50-state breakdown)
- Brynjolfsson, Li, Raymond, "Generative AI at Work," Quarterly Journal of Economics, 2025
- Noy and Zhang, Science 381:187-192, 2023
- Cui, Demirer et al., GitHub Copilot RCTs, MIT Economics working paper
- Durica et al., Information (MDPI) 16:1078, 2025
- US Census Bureau, Business Trends and Outlook Survey, May 2026; SBA Office of Advocacy, Sept 2025
- JPMorgan Chase Institute, "Understanding AI Use by Small Businesses," Dec 2025
- UT Boyd Center, Tennessee Business Leaders Survey, Winter 2026
- S&P Global Market Intelligence via CIO Dive, 2025; RAND Corporation, 2024; MIT Project NANDA, 2025
- Magesh et al., Journal of Empirical Legal Studies 22:216, 2025 (Stanford RegLab)
- Charlotin AI Hallucination Cases Database; Pfefferkorn, Stanford CIS, Oct 2025
- Whiting v. City of Athens, 6th Cir., Mar 13, 2026; Mata v. Avianca, S.D.N.Y. 2023
- Moffatt v. Air Canada, 2024 BCCRT 149; Fortune on Deloitte Australia, Oct 2025
- METR developer RCT, July 2025, with METR's own Feb 2026 caveat
- Reimers and Waldfogel, NBER WP 34777, 2026; NewsGuard AI tracking, 2024-2025
- FBI IC3 2025 Annual Report and Elder Fraud data, April 2026 (ic3.gov)
- CNN and WEF on Arup, 2024; The Guardian on WPP, May 2024; FTC consumer alerts on voice cloning
- EEOC v. iTutorGroup, 2023; Louis v. SafeRent, D. Mass. 2024; Massachusetts AG and Earnest, July 2025; Mobley v. Workday, N.D. Cal. 2025
- Garcia v. Character Technologies, M.D. Fla. (settlement as reported by CNN/CNBC, Jan 2026); Raine v. OpenAI complaint, Aug 2025; NAAG 44-AG letter led by TN AG, Aug 2025; FTC inquiry, Sept 2025
- FCC Forfeiture Order 24-104; NHPR on the Kramer acquittal, June 2025; Harvard Ash Center, 2025; Yale ISPS, Sept 2025
- WVLT and CNN on the Elizabethton case, Mar 2026; WSMV on TN elder fraud, May 2026; WJHL and WCYB, May 2026
- BLS OEWS Johnson City, May 2024; US Treasury AI labor report, 2024
- Eastman Q1 2026 prepared remarks; Ballad Health newsroom and Becker's Hospital Review
- Hamby et al., International Journal of Communication 12, 2018; Lewis and Severnini, J. Development Economics, 2020
- TNECD BEAD awards, Mar 2026; Tusculum University; Northeast State; Tennessee SBDC
- ELVIS Act, signed Mar 21, 2024 (tn.gov); SB 1580, 2026; The Center Square on enforcement, Dec 2025
Ready to put AI to work in your business?