PRAPI Research · 2026-06-19
How founders actually run their companies on AI (and what they dropped)
We asked founders and operators a plain question: which AI has actually become part of how you run your company, what you use each one for, and which ones you tried and then dropped, and why. Eighty-nine people answered. The pattern was close to unanimous, and it cuts against the marketing. AI earns a permanent place in a real company when it removes routine load and sits close to the operator's own data and workflow. It gets cut, fast, when it tries to replace judgment or puts distance between the founder and the customer. Pavankumar Kamat of Panto AI said it cleanest: the AI that survives "shortens decision loops and preserves trust; anything that promised to replace judgment got cut fast." David Hunt of Versys Media drew the same line from the other side. The systems that stayed "augment judgment," he said, and the ones they dropped "tried to replace judgment outright." The kept stack is unglamorous: first-draft engines for internal text, orchestration that routes leads and removes human lag, meeting capture that turns talk into tasks, research that replaces fifteen open browser tabs. The discarded pile is remarkably consistent across very different businesses: generic copy generators that cost more to edit than to write from scratch, autonomous customer-facing bots that erode trust on contact, and bulk content engines that AI search quietly ignores. This report is the field guide founders wrote for each other: what to keep, what to kill, and the one filter that tells them apart.
12 contributors cited
The rule under every answer: augment judgment, never replace it
Across 75 usable answers, one sentence could stand in for most of them. Pavankumar Kamat, co-founder of Panto AI, put it as a practical rule: "adopt AI where it reduces routine cognitive load and keeps a human decision point; discard any tool that saves time by shifting ambiguity or risk onto the team. Today's durable wins are augmentation plus observability, not replacement."
David Hunt, COO of Versys Media, reached the same conclusion by sorting his own stack: "The tools that stayed are embedded into existing workflows, sit close to our data, and augment judgment. The ones we dropped tried to replace judgment outright." Warren Davies of BeyondCRM, a thirty-year CRM consultant and an open skeptic, framed the failure mode: founders rush in "adopting technology before understanding the problem it's solving. Start with your actual workflow gaps, then find the tool, not the other way around."
That is the spine of every answer below. Keep reading and you will see the same line redrawn in a dozen industries.
What stays: the unglamorous stack
The tools that survive do narrow, repetitive, document-heavy work and then step aside.
First drafts of internal and operational text. Not the finished product, the raw material. Ashley Cordova, VP of Zia Building Maintenance, turns a scope-of-work template that "used to take me an afternoon" into a usable draft "in 20 minutes." Fahed Bitar, project executive at S-Line Contractors, cut RFI response time "from 48 hours to 6" by putting Claude on contract-language review. Luke Jordan of iRepair Heating and Air uses AI to turn "deferred interest" and "revolving credit" into plain language homeowners read before signing. The hard line is universal: nothing reaches a customer unedited. Mendy Perlman of Web Hosting Services said it best. "AI writes my first draft. I rewrite every word that goes to a client. That is not inefficiency. That is the product."
Orchestration that removes human lag. The highest-value use is rarely "AI writes for us." Magee Clegg, founder of Cleartail Marketing, keeps SharpSpring not to generate copy but "to remove human lag: route a lead by geography, trigger the right follow-up, notify sales, and track whether that lead becomes revenue." David Hunt and Carlos Cortez of S9 Consulting both pointed to n8n-style orchestration as the "glue" that connects a CRM, forms, and messaging so an AI interaction "triggers the next operational step instead of becoming another disconnected chat transcript."
Capture that turns meetings into work, and research that replaces tabs. Founders keep transcription and summarizers (Fathom, Descript, Otter, Copilot) because, as one put it, the synopsis lands on the ticket and "meeting output becomes work, not noise." Hansjan Kamerling of Adaptify AI keeps Perplexity for one honest reason: it "replaced my habit of opening 15 browser tabs."
What gets dropped, and it is the same list everywhere
Five categories came up again and again, across trades, agencies, SaaS, real estate, and creative studios.
1. Generic copy generators. This was the single most-cited casualty. Jasper, Copy.ai, and Copysmith appear in answer after answer as tools that produced brand-weak output and "increased editorial overhead rather than reducing it." Caleb Johnstone, SEO director at Paperstack, dropped Jasper because he "rewrote it every single time, which defeated the point entirely." Samuel Huang of Tele Ads Agency: "Every headline sounded like a template from 2018."
2. Autonomous customer-facing bots. Founders in relationship businesses killed these fastest, because the failure is trust, not output quality. Ashley Cordova dropped an intake bot: "The moment someone felt like they hit a bot before a human, trust eroded, and in this business, trust is the entire product." Luke Jordan put it in concrete terms: when a furnace dies in January, "an AI hedge or delay in that moment costs you the job and the trust." Rusty Rich of Latitude Park keeps a strict human-in-the-loop policy after autonomous review responders "delivered robotic, tone-deaf answers to complex complaints."
3. Bulk AI content engines. Operators who chase volume get punished by the same search engines they are trying to win. David Caruso, who runs 60-plus brands from Thailand, dropped bulk AI-SEO content flatly: "Pumping out 500 AI articles felt productive and Google ignored every one of them." Rob Dietz of Dietz Group abandoned automated content spinning because "generative search engines easily identify duplicate, low-quality content and completely ignore it." Scott Kasun of ForeFront Web warned that letting AI generate full website builds, the trend "some call vibe coding," produces "messy code and massive technical debt."
4. Anything that touches the craft core. In businesses built on a signature output, AI gets walled off from the work itself. Adam Gorham, founder of a New England wedding cinematography studio, cut AI color grading, automated editing, and music licensing because they "couldn't read emotional rhythm. A first dance is a feeling, not a clip." His rule: "AI owns the admin. It never touches the story." A commercial photography studio abandoned ChatGPT for captions and proposals outright because the copy was "clearly AI-written, which isn't something we could present to clients."
5. Replacing a judgment call someone is accountable for. The most expensive failures came from tools trusted with a real decision. Fahed Bitar killed an AI estimator after it "hallucinated cost estimates mid-meeting on a live road construction bid. That one conversation nearly cost us the contract." Others dropped predictive schedulers that ignored permit delays and candidate-scoring vendors that "amplified proxies and gave false confidence."
The filter founders use to tell the two apart
Three tests recur.
A hard utility bar. Caleb Johnstone's rule is representative: "If an AI tool doesn't save us at least 5 hours a week, it's gone. That rule cut our stack from a dozen experiments down to 3 tools we actually use every day."
Problem first, tool second. Ruth Jennifer Cruz of Wolf King USA audited 12 B2B companies and found each "wasted over $100,000 on AI tools they barely touched. The root cause is simple. Companies buy solutions before they know the problem." Magee Clegg's advice is the same in fewer words: "Start with the revenue bottleneck, and only keep tools that measurably reduce it."
Human-in-the-loop as policy, not preference. The founders who scaled AI furthest are the ones who wrote the human checkpoint into the process, not the ones who removed it.
The signal worth watching: measurement is starting to break
A quieter theme ran underneath the tool talk: the way AI search reaches customers is breaking the dashboards founders rely on. Jennifer Bagley of CI Web Group put it sharply: "We now track revenue outcomes, not just clicks. When AI search sends a customer directly to a booking with zero website visit, your analytics show nothing happened." David Caruso now uses Perplexity to check "whether our products actually show up when buyers ask AI before they ever touch Google." Rob Dietz is restructuring content with schema so it can be cited at all. The founders paying attention have stopped asking how to make AI write for them, and started asking whether AI can find them.
What this means if you are choosing AI right now
The operators in this report are not anti-AI. Several run AI companies. Their consensus is a buying discipline, not a verdict. Keep the systems that draft, route, capture, and research close to your own data. Rewrite anything a customer will read. Kill anything that stands between you and the customer, or that quietly takes over a judgment call you are accountable for. The tools that lasted all passed the same test: they gave the founder time back without taking the relationship, the brand voice, or the decision with them.
Contributors
AI survives in our stack when it shortens decision loops and preserves trust; anything that promised to replace judgment got cut fast. Today's durable wins are augmentation plus observability, not replacement.
The tools that stayed are embedded into existing workflows, sit close to our data, and augment judgment. The ones we dropped tried to replace judgment outright.
The pattern I see with founders rushing into AI mirrors bad CRM implementations: adopting technology before understanding the problem it is solving. Start with your actual workflow gaps, then find the tool, not the other way around.
AI writes my first draft. I rewrite every word that goes to a client. That is not inefficiency. That is the product.
The moment someone felt like they hit a bot before a human, trust eroded before we even had a conversation. In this business, trust is the entire product.
RFI response time dropped from 48 hours to 6 after we put Claude on contract language review. But an AI chatbot hallucinated cost estimates mid-meeting on a live bid, and that one conversation nearly cost us the contract. AI earns its place on repetitive, document-heavy work. Judgment calls still belong to humans.
The highest-value use is not AI writes marketing for us. It is using automation to remove human lag. Do not buy AI because it feels advanced; start with the revenue bottleneck and only keep tools that measurably reduce it.
If an AI tool does not save us at least 5 hours a week, it is gone. That rule cut our stack from a dozen experiments down to 3 tools we actually use every day.
I audited 12 B2B companies and found they each wasted over $100,000 on AI tools they barely touched. Companies buy solutions before they know the problem.
- David Caruso, Founder at BuyFactory.direct
I run 60-plus brands across 700-odd websites, so AI is not a gimmick for me, it is staff. But I dropped the bulk AI-SEO content generators. Pumping out 500 AI articles felt productive and Google ignored every one of them. Quality still wins.
AI owns the admin. It never touches the story. Automated editing could not read emotional rhythm, and a first dance is a feeling, not a clip.
We now track revenue outcomes, not just clicks. When AI search sends a customer directly to a booking decision with zero website visit, your analytics show nothing happened.
In fashion the winning model isn't the most general one, it's whichever respects the physical truth of the garment: how fabric drapes, whether a logo survives. Flux Krea is exceptional on fabric and materials; we dropped Veo and Seedance because they don't drape soft fabrics convincingly. The tools that stuck do one thing faithfully.
The first AI tool I bought for my team I killed inside a week, a resume filter that kept knocking out strong developers for a non-linear career, a bootcamp instead of a CS degree, a two-year gap. Talent spotting is exactly the kind of judgment you can't hand to a model. The one I'd never give up is far less glamorous: cash flow planning, three hours of spreadsheet dread compressed into a five-minute gut check.
My rule is simple: if an AI tool touches client work, it either saves measurable time or reduces liability exposure, otherwise it's a toy we can't afford. We dropped an AI accessibility scanner that promised auto-detection, the false-positive rate was too high, and in our world a wrong call ends up in a demand letter. Humans still sign every audit.
I treat a general assistant as a first-draft engine, product copy, the awkward supplier email I keep putting off, the first cut of a campaign brief. I never ship its output raw, but starting from a draft I edit rather than a blank page is where it pays back. For a supplement brand I will not let a machine make a health-sounding statement unsupervised.
I run a primary-care practice, so my use of AI is shaped hard by HIPAA. Nothing touching patient data goes near a general consumer tool, full stop. What stuck is on the administrative side, away from the chart: first passes at internal documents and plain-language website copy I then rewrite in my own voice. It saves me the blank-page tax.
As a solo founder, AI is basically my entire team. Claude and ChatGPT are my pair-programmer and bilingual copywriter, and a Whisper model powers Confileo's own transcription tool. The one thing I dropped: using AI to auto-generate hundreds of SEO landing pages, they came out thin and templated and Google flat-out refused to index them. AI is unbeatable for leverage, but a terrible autopilot for quality.
I use AI to automate repetitive operational tasks, not to replace human decision-making, classifying project requirements, surfacing data-quality issues, routing tasks. These workflows cut administrative processing time about 60 to 65 percent. We dropped standalone chatbots that didn't connect to our systems, they produced answers but lacked the operational context to be reliable.
- Ashish Dsa, CTO & Co-founder at Arbor
I run an AI company, so I'm a heavier user than most, but the honest list is shorter than you'd think. What stuck: a coding agent for the first pass, an LLM for drafting and triage where a human does the final pass, and AI transcription on every meeting. What I dropped: most standalone AI-for-X point tools, they get eaten the moment the model we already pay for can do the same thing in-house.
- Kruno Sulić, Founder & SaaS Product Builder at Cliprise
The tools that became part of how I run the company remove repetitive work without creating a second job in review. What stayed: tools that get me from blank page to first usable version faster. What didn't: tools that promised autonomous business operations, the common failure was confident but shallow output that looked finished while hiding errors underneath.
- Roman Surikov, Founder at Ronas IT | Software Development Company
AI became useful only after we stopped treating it as a separate innovation project and put it inside existing workflows. My rule now: AI can prepare, compress, compare, and suggest. It can't approve, promise, or take responsibility. The biggest productivity gain comes from using AI before expert work, not instead of it.
Running solo, every tool has to earn its place: ZimmWriter to draft, Claude as the editor that fact-checks and tightens, Make.com for the busywork. I dropped Page Optimizer Pro once I noticed Google now rewards genuine information gain, content that adds something new, over pages that just hit keyword-density targets.
The AI that actually runs my company is the unglamorous kind, coding assistants and the repetitive content work. The part most operators skip is what I deliberately keep AI away from. No autonomous agent goes near patient data or anything clinical, because once you hand an agent broad permissions and a browser, prompt injection stops being theoretical.
Claude is the tool most embedded in how we run the company. Our weekly content performance analysis across over 300 pages used to take six to eight hours of pulling and cross-referencing; with Claude it takes under one, and content output roughly doubled without adding headcount. What we dropped: automating customer follow-up emails, quality fell immediately because every customer situation needs contextual judgment a template cannot fake.