What AI Looks Like on a Dirty Construction Site

Part II of "AI Is Shortening the Road Into Small Business"

What this article covers

How AI can support the first weeks of a real operating business.
A post-construction cleaning job where photos, staffing, chemicals, and deadlines collide.
What already works in hiring, purchasing, scheduling, quality control, and daily reporting.
Where ordinary business software ends and the new AI layer begins.
Why more affordable attempts may create more companies, purchases, contracts, and jobs.

Full article

In Part I, I wrote about the shorter path from an idea to a first serious business test. Here I want to follow the next step: the first real job, where a plan meets the floor, the workers, and the customer.

Imagine a multistory office building at the end of construction.

The client wants the final cleaning completed floor by floor so the building can pass acceptance. On paper there is a schedule, a team, equipment, chemicals, a deadline. Then you walk the site.

There is dried putty on one surface, paint drips on another, construction adhesive on a tile floor. If somebody attacks that adhesive with the wrong scraper, the floor is scratched before the client has even moved in. Meanwhile you notice defects left by the construction crews, report them, and those crews come back into the same rooms your people are trying to finish. Cleaning has started. Construction has not quite stopped. Everybody is late.

That is a normal large job.

Now picture how I would want AI to help with that first survey. Photos and video from a phone or camera could become the basis of a preliminary work order: the likely contamination, the surface, a possible construction defect. Two people here, one there; a machine for the fourth floor, a suitable chemical for the adhesive, protection for the finished wall. Quantities and time would need an estimate too.

A supervisor checks it, corrects it, sends the team.

Two hours later the original plan is already wrong. Five people are standing in an area where two could finish. On another floor, a stain is heavier than it looked and one person needs help. A system following progress, labor hours, open defects, and available equipment could suggest moving three people upstairs. The supervisor would check whether that move makes sense on the actual site.

This is where AI becomes useful in a small business. Not on a Sunday afternoon when somebody is lying on a sofa and says, "Hey AI, I feel like owning a company. Build the whole thing for me."

It does not work like that now. Although perhaps one day it will.

AI is useful when the person in the room understands the work. He knows why one scraper damages tile and another does not. He knows that a photograph can confuse glue with sealant. He knows five people in one room can look busy and still waste half a shift.

That is the kind of help I want: a way to use knowledge of the work across the whole job, including the places I cannot watch at every moment.

The operating plan starts before the first worker arrives

Every small business has its own version of that building.

A tailor needs measurements, fabric, fittings, production time, delivery promises. A woodworking shop needs drawings, material yield, machine time, finishing, installation. A building-material company needs stock, supplier terms, transportation, pricing, credit. A cleaning company needs scope, staffing, chemicals, equipment, inspection, invoicing.

Once the owner decides to launch, AI can connect work that used to live in separate notebooks, spreadsheets, messages, and people's heads.

The useful sequence is simple:

Capture the real situation through documents, photos, video, calls, estimates, and transactions.
Turn it into a proposed plan, checklist, comparison, or exception.
Let the person responsible approve or correct it.
Measure what actually happened and feed the correction back into the next plan.

The last step matters most. Without it, AI keeps producing polished versions of the owner's old assumptions.

Cleaning shows what is mature and what is new

This is why I keep returning to cleaning. You can go to the site and look at the result. Was the floor finished? Did the worker arrive? Did the chemical remove the adhesive without damaging the surface? If the labor estimate was wrong, you can investigate which part of the job took longer.

Much of the necessary technology already exists, although it was not called GenAI when it was built. Products such as CleanGuru and Aspire provide bid estimation, scheduling, geofenced time tracking, inspections, job costing, invoices, reporting, consumables, and route optimization.

That is the mature layer.

The new layer lets an owner talk to the operating system in normal language and feed it less structured information. Photos from a site survey. A voice note from a supervisor. A customer complaint. A stack of proposals. An inspection video. The AI can draft a bid from the survey, explain why a job missed its labor budget, create training instructions in the worker's language, group repeated complaints, or identify that one building consumes far more chemicals per square foot than similar sites.

New cleaning products are already selling AI-drafted proposals, pricing assistance, scheduling, and operational advice. What we do not have yet is good independent evidence separating the value of GenAI from the value of basic digital discipline.

That distinction matters. If the company never measured labor hours by job, adding a chatbot does not fix the missing data. If inspections were inconsistent, an AI summary of inconsistent inspections remains inconsistent.

For the system to explain why a floor took too long, the company first needs a record of the work on that floor. That is where I would start.

Hiring already has an AI layer

Before sending a team to that building, someone has to hire it. The owner writes a job description, collects resumes, schedules calls, takes notes, compares candidates, then tries to remember who said what. That can take a large part of the first weeks.

AI can screen resumes against explicit requirements, draft role-specific questions, transcribe an interview, summarize the evidence, and identify what still needs to be asked. LinkedIn already offers AI-led audio and video screening that generates questions, transcripts, summaries, and ratings.

I would use that as a first filter and note taker.

The U.S. Equal Employment Opportunity Commission is clear that automated resume screening and video-interview evaluation remain employment decisions. Antidiscrimination law still applies. Accommodations still apply. The fact that software produced the recommendation does not move responsibility away from the employer.

Real-time question suggestions during a human interview also exist, but this is still an early category. There are few independent studies showing that a live copilot improves hiring decisions for small businesses. It may help you remember a follow-up. It may also distract you from the person sitting in front of you.

For the cleaning job, I would still want to hear how the person thinks through the work. The transcript can help me return to an answer; it cannot have that conversation for me.

I would not let a score decide who gets the job.

Purchasing and negotiation need a back table

The equipment and chemicals for that job bring another set of decisions. Even comparing suppliers takes time when ten of them quote the same item in ten different ways.

One includes delivery. One changes the package size. One offers a lower unit price but requires a larger minimum order. Another gives thirty days of credit. The cheapest line on the screen may be the most expensive choice after freight, waste, storage, and payment terms.

AI can normalize the quotes, compare specifications, flag missing terms, calculate total landed cost, and prepare questions. Commercial platforms already offer AI search, product comparison, supplier intelligence, and document matching.

Negotiation works the same way. Before the call, the model can organize your BATNA, deadlines, acceptable range, supplier constraints, and the points you should not trade away too early. After the call, it can analyze the transcript and compare the agreement with what was promised.

That preparation gives me something useful to take into the negotiation. During the conversation, I still need to understand what the other side is offering and which conditions I am accepting.

There is not yet strong field evidence that AI reliably improves the final commercial terms for small companies. I would use it for preparation, calculations, and keeping track of the terms, with the final agreement left to the person responsible for the purchase.

Daily management becomes a measurement loop

Once a business is running, the question changes. You are no longer asking whether the idea sounds good. You are asking what happened today.

How many leads came in? Which quote converted? What was the average check? How many labor hours did the job consume? Where was rework required? Which material ran out? Which customer is waiting? What happened to gross margin?

Accounting products already analyze profit-and-loss statements, balance sheets, trends, and anomalies through conversational tools. A small owner can ask why labor cost rose this month or which customer segment is paying slower.

The explanation still depends on what was recorded. If hours were entered against the wrong job, the owner needs to catch that before acting on the comparison.

AI should point to an exception and propose a question. The owner checks the cause on the floor, in the warehouse, in the call recording, in the bank account. That daily loop is more valuable than a large presentation created at the end of the quarter.

More attempts do not mean more successful companies

The same process has a wider consequence that interests me. If it becomes easier to prepare a business and organize its first jobs, more people may get far enough to try.

As I discussed in Part I, Gusto's founder survey points mainly to a faster, less expensive path for people who already wanted to start a business. It does not establish how many additional firms AI caused to exist.

That is still a large effect. A person with an intention can reach a serious test sooner.

I can now consider an agentic-automation business in logistics, woodworking, building-material sales, cleaning. I can build an interactive children's product that once required several separate creative roles. I can work on cleaning software because I understand the process and AI reduces the cost of building the first version.

More attempts will also produce more weak ideas, more closed projects, more polished nonsense. Fine. The important question is whether a weak idea can be stopped earlier and a good one can reach a customer before the owner runs out of time, money, or nerve.

The Census Bureau's September 11, 2026 release makes the distinction concrete. It reported 531,728 seasonally adjusted business applications for August and projected 28,501 employer startups from that application cohort within four quarters. The second number is a forecast of future payroll businesses, not a count of firms that opened in August. An application is not an operating company.

The funnel remains severe. For one U.S. cohort of establishments born in March 2013, 79.6% were operating after one year, 50.6% after five years, 34.7% after ten.

AI does not remove that filter. It may help more people reach it with better preparation and less sunk cost.

More attempts create work around them

People are afraid AI will remove jobs. Some of that concern is reasonable. Routine work will change, and some tasks will require fewer hours.

I think we are underestimating the opposite mechanism.

A new cleaning contract buys chemicals, equipment, uniforms, transportation, insurance, accounting, payroll software. A woodworking shop buys material, blades, finishing products, packaging, delivery. A children's story product needs distribution, payment processing, testing, customer support, maybe new artists and editors once demand exists. A solo business may start with one person and still hire contractors.

The economic effect is not limited to payroll inside the new firm. It includes the purchases and professional work around every serious attempt. Financing too. The Federal Reserve's 2026 report on its 2025 survey found that 60% of responding employer firms had sought financing in the preceding 12 months. Operating expenses and expansion were the two leading reasons.

Early working papers are beginning to find more business entry and smaller founding teams in areas exposed to GenAI. That evidence is still young. We should call it a signal.

My bet is simple: lowering the cost of competent experimentation will create more real demand around the attempts that make it through. Some of those businesses will hire. Others will remain small and buy services from somebody else. I want to understand how much of that demand can grow alongside the tasks that become automated.

Banking will be part of the next compression

Digital banking already shortens some account applications and reviews, while lenders use cash-flow data and machine learning in underwriting. GenAI is entering document extraction and customer service. I expect these pieces to connect more closely with business formation and accounting, but that is my forecast. The existing services do not show that banks have redesigned products for AI-enabled founders.

The Federal Reserve's 2026 report also shows the cost of speed. Among applicants for loans, lines of credit, or merchant cash advances, the share applying to online fintech lenders rose from 17% in 2020 to 29% in 2025. Among borrowers, 60% using online lenders reported higher-than-expected costs, compared with 37% at small banks and 32% at large banks. A shorter wait can help, but the financing still has to fit the work it funds.

The owner still has to know the work

For my own projects, I keep coming back to the same starting point: a business where I understand the work.

If you make suits, you should understand fabric, fit, customers, delivery. If you sell building materials, you should know what fails on a site and why contractors complain. If you clean buildings, you should know what removes adhesive without destroying tile.

AI can help with the financial plan, tax questions, bookkeeping setup, credit options, marketing, hiring, purchasing, schedules. It can act like a universal first-pass specialist.

When the team leaves that building, the customer receives an actual floor, with the result of every decision made along the way. I want the software to help the person responsible see those decisions earlier, while there is still time to change them.

Practical takeaways

Automate measurable processes first: scope, schedule, labor time, inspection, cost, invoice.
Use photos and video as inputs, then require a person to confirm classifications and work orders.
Treat AI hiring tools as screening and note-taking support, never as an automatic judge.
Give procurement tools actual quotes, specifications, delivery terms, and payment conditions.
Review exceptions daily instead of waiting for a monthly explanation.
Decide in advance what evidence means go, revise, or stop.
Keep legal responsibility, customer promises, safety, and quality with a named person.

In brief

AI becomes valuable after launch when it connects real operating data with accountable decisions.
Cleaning already has mature software for bidding, scheduling, timekeeping, inspections, and job costing.
GenAI adds a conversational and unstructured-data layer, but independent ROI evidence remains limited.
More affordable attempts do not guarantee more successful firms. They allow faster testing and earlier stopping.
The wider economic effect may include demand for suppliers, software, contractors, finance, and later hiring.
Domain knowledge remains the base. AI helps turn that knowledge into a repeatable business.

Denis Ostapenko · ORCID 0009-0006-2630-7280