You have been thinking about AI. You have read the articles, maybe bought a license or two. But when it comes to your own company, your people, your workflows, the first step is not obvious.
So I stopped guessing and counted. Between January 2025 and mid-2026 I recorded and systematically mined real buyer conversations with mid-market operators, then deep-coded a subset call by call. What follows is not opinion. It is what the transcripts say, with the counts attached.
How the dataset was built
This is a first-party dataset, not a survey I bought or a trend I read. Here is exactly what it is, so you can weigh it.
- 84 transcripts and supplementary documents across 24 client and prospect companies, January 2025 to June 2026.
- 17 buyer calls deep-coded, with every quote grep-verified against its source transcript.
- 8 industries: construction and real estate, wealth management, legal, manufacturing and packaging, logistics, consumer goods, fashion, and professional services and solo operators.
- 85 catalogued insights drawn from those calls.
- Company profile: roughly 2 to 200 employees, founder- or family-led, across the US and Latin America.
- All companies are anonymized except where written permission exists.
What operators actually said (7 findings)
The denominator is the 17 deep-coded calls. Each count is how many of those 17 calls raised the same problem, unprompted.
| # | What operators actually said | Calls | Spread |
|---|---|---|---|
| 1 | Everything runs on spreadsheets or by hand; data is scattered across tools and people | 12 of 17 | every industry |
| 2 | "I am the bottleneck": the business lives in one person’s head | 8 of 17 | solo to mid-market |
| 3 | We do the work but cannot see margin or results to make decisions | 7 of 17 | construction, consumer goods, fashion, foundations |
| 4 | WhatsApp is the de facto system of record | 6 of 17 | LATAM construction, real estate, logistics |
| 5 | A wrong number or a missed item costs real money or a relationship | 6 of 17 | logistics, real estate, consumer goods, wealth |
| 6 | Manual document and spec generation eats skilled people’s time | 5 of 17 | fashion, legal, foundations, construction |
| 7 | Already tried AI; without structure it wastes time | 5 of 17 | solo, fashion, foundations |
One claim per line, each with its own count. The pattern underneath all seven: the technology is rarely what is stuck. The data is.
The quotes behind the numbers
Every quote below is copied verbatim from a source transcript and anonymized to a role. Spanish quotes stay in Spanish with a translation.
Finding 1. Scattered data, manual consolidation (12 of 17)
"We're running everything on spreadsheets and a patchwork of tools that don't talk to each other." consumer-goods founder
"abriendo la carpetita, sacando el papel, buscando la hoja" (opening the little folder, pulling out the paper, looking for the sheet).construction group
Finding 2. One person is the system (8 of 17)
"If I get hit by a bus, this company doesn't run. Everything's in my head or in a spreadsheet only I understand." consumer-goods founder
"Es una sola cabeza haciendo lo que pueden hacer ocho." (It's a single head doing what eight could do.).solo practitioner
Finding 3. Work happens, results stay invisible (7 of 17)
"We're growing revenue but I have no idea if we're growing profit." consumer-goods founder
"necesitamos información para tomar decisiones y hoy estamos [sin]." (we need information to make decisions and today we have none.).foundations owner
Finding 4. WhatsApp is the system of record (6 of 17)
"dejemos de dar órdenes por WhatsApp... si me lo pedís por WhatsApp, no vale." (let's stop giving orders over WhatsApp... if you ask me over WhatsApp, it doesn't count.).residential real estate developer
"nosotros usamos WhatsApp para todo... todavía es como muy artesanal." (we use WhatsApp for everything... it's still very manual.).construction group
Finding 5. A wrong number costs money or a relationship (6 of 17)
"tenemos mucha pérdida de plata en herramientas que no vuelven, insumos que no vuelven, que desaparecen." (we lose a lot of money on tools and supplies that don't come back, that disappear.).foundations subcontractor
Finding 6. Manual document generation eats skilled time (5 of 17)
"perdemos tiempo en la construcción de la ficha técnica porque es algo que una persona está generándola." (we lose time building the spec sheet because a person is generating it.).fashion brand
"the risk factor section... is repetitive." boutique law firm
The thesis, in their words (finding 7)
The most telling finding is the seventh. Buyers who have already tried AI on their own arrive at the same conclusion I do: the blocker is prompt quality and context, not the model. They say the thesis back to me before I pitch it.
"hay mucha gente que usa la IA, pero pierde mucho tiempo... tienes que definir muy bien qué tú quieres." (lots of people use AI but waste time... you have to define very well what you want.).solo practitioner
"si vos pedís una estupidez, te devuelve una estupidez... encima vas a tener que leer la estupidez." (if you ask something dumb, it gives back something dumb... and you'll have to read the dumb thing.).foundations subcontractor
This is the whole argument for structure over tools. A model with no context is a fast way to produce confident nonsense. The work is defining what good looks like, once, so the output is right every time. That is what an AI Operating System is.
Industry-specific findings
The seven findings hold across industries. The specifics change. These are real, traceable examples from the corpus.
Construction. Data lives in email, spreadsheets, and job-site notebooks. The first win is cost visibility: knowing real-time spend, not finding out when the project is done. One company caught a six-figure cost overrun in month two because the system surfaced data they already had but could not see. The second win is document intelligence: crews asking questions about specs and compliance in real time instead of asking people.
Wealth management. A $14B wealth advisory firm runs about 90% of its business through email, with Outlook as the system of record. The first win is data consolidation: one place that knows what clients own, instead of asking advisors. The second is reporting: custom client narratives instead of templates.
Legal. The first win is research speed: answers to procedural questions in seconds instead of hours. The second is client communication: systematic responses to common questions instead of email-thread chaos. The constraint is privacy, so the AI layer stays closed and trained on nothing.
Manufacturing and packaging. A packaging manufacturer operating in four countries had product specs in design software, pricing in a legacy system, and customer relationships split between email and no structured CRM. The first win is product knowledge: sales teams asking what is in stock, compatible, and available in volume, and getting an answer.
The sequence matters. You cannot automate what you cannot see.
The cost reality
The cost question comes up in every conversation eventually. The engagements are fixed-fee, in three phases.
AI Ops Audit. Fixed fee, 2-4 weeks. A map of where your data lives, what is working, what is stuck, and a prioritized roadmap.
Foundation Build. Fixed fee, 4-16 weeks depending on complexity. Dashboards, automation, connected tools, trained people. You own the code and the data.
Ongoing AI Operations. Fixed-fee monthly retainer. This is fractional Head of AI: new tools evaluated, new processes designed, your team trained, and the system doing more every month.
A Chief AI Officer is expensive once you factor in salary and benefits. The retainer is a fraction of that, and it only starts after you have seen results in the first two phases. The full breakdown is in what AI consulting actually costs for a 20-200 employee company.
One external note, not a corpus finding. If you build custom tools, your company likely qualifies for the federal R&D tax credit (IRC Section 41), under which 65% of a consulting fee counts as Qualified Research Expenses. On a $50K build, that is roughly $4,550 in federal credit at the Year 3+ ASC rate. Add the Section 174A immediate-deduction value and the two together offset on the order of 18 to 23% of the build cost. Returns can be amended for prior years. This is general information, not tax advice. Verify with your CPA. The detail is in tax credits and grants for AI.
If you run a company with 20 to 200 employees and you have been thinking about AI without knowing where to start, the dataset points to one answer: start with the data, not the tool. The where to start with AI guide walks the first step, and the free assessment shows you where your company stands in about three minutes.
Or start directly: the AI Ops Audit maps where your data is and identifies your highest-ROI starting point. If you want the industry-specific version, construction, wealth management, and legal each have their own breakdown.
Ignacio Lopez
Fractional Head of AI, Work-Smart.ai · Coconut Grove, Miami. Fractional Head of AI for mid-market companies with 20 to 200 employees.
Connect on LinkedIn →