WILD×AI Survey Sprint · the deliverable
Everything in this document traces to something someone at Stonepile said out loud, on a date, on a recording. Where a number is not established, it says so.
You came into this asking which AI to buy.
What follows answers a different question, and we think it is the one underneath. Two things are true about this business at the same time. It runs on judgment that four people carry in their heads and nobody has written down. And it is about to be asked to hand that judgment to systems that cannot read minds.
Part one is what a competent operational audit would find: how the work runs, where it stalls, what that costs you in hours and money. It is real, the numbers are yours, and it is the smaller half.
Part two is what your business can say about itself. What is true here that has never had to be said, because among the people who work here it never needed saying. This is the part no vendor has asked you about, and it is the part that decides whether anything in part one is worth building.
Part three is the crossing between them. Where you actually stand, in what order things can happen, and why the order is not a preference.
Part one
Walked end to end with every person who touches it.
Stonepile runs two processes that carry almost everything: a job from inquiry to scheduled work, and a job from the site to an invoice.
Inquiry to scheduled work. A call or a website form reaches Dana, who logs it and books a site visit. Marcus visits and builds the estimate in the spreadsheet workbook. The estimate goes out as a PDF and comes back accepted by email or phone. Then Dana types the whole thing into the service system again, line by line, because nothing moves between the two on its own. Ray assigns crews on Monday, on the whiteboard.
That last step is worth pausing on, and we come back to it twice.
"Monday morning I build the week on the whiteboard. There's no list. I just know who works where and who I can't put on the same job."
Ray Kowalczyk, owner · session 1 · 00:08:40
On-site work to invoice. The crew completes what was scheduled. Extra work gets identified on site, photographed, texted to Ray, and agreed verbally. Parts used get written on a paper job card, when the day allows. The card comes back one to four days later, Dana raises the invoice from the card and the service system, and variations make it onto the bill if the card recorded them.
"Extra work on site, I text Ray a photo and we agree it. Whether it gets on the invoice depends who's in the office that afternoon."
Tomás Rivera, field lead · session 3 · 00:15:27
Both processes work. Neither is broken in a way anyone would call an emergency, which is precisely why they have been like this for years.
Six things hurt, and your people named all six without being led there.
Every won estimate is typed twice. Built in the spreadsheet, re-keyed into the service system after it is won. Nothing automatic connects them.
The week's schedule exists in one head. Crew assignment encodes who works well where, who cannot be paired, and what each site needs. Ray described a week of his absence as coping rather than working.
Work agreed on site is not reliably billed. Agreed by photo and text, recorded on a card if the day allows, invoiced if the card survives the trip back. More than a third does not.
Missed calls during busy periods are never followed up. Eleven a week, reported by your own phone system. No callback process, and no record of who rang.
Half the estimating work rebuilds quotes that already exist. Marcus quotes repeat work from scratch because the previous version is in a folder somewhere.
Parts used on site are captured inconsistently. Written on the card when the day allows, reconstructed later or absorbed.
Four of these carry numbers you gave us. Every rate below was stated by Ray when we asked, never assumed from a table.
| What | How we got there | Per year |
|---|---|---|
| Unanswered inbound calls during peak periods | 11/wk measured · 15% conversion · $310 average call | $26,598 |
| On-site variations agreed and never invoiced | 6/month · 40% unbilled · $840 average | $24,192 |
| Re-keying won estimates into the service system | 9 hrs/wk at $38/hr, both stated | $17,784 |
| Rebuilding estimates that already exist | 5.5 hrs/wk at $52/hr, both stated | $14,872 |
| Parts going out unbilled | Mechanism confirmed. Size never established. | not calculated |
Across four items. Each one shows its arithmetic and a range in the evidence file, so you can see exactly how soft or hard it is. The missed-call figure rests on a measured call volume and Ray's own conservative estimate of what converts. The variation figure is the softest of the four: the volume and average value are Dana's, and the unbilled proportion is her judgment rather than anything counted. Two months of deliberate counting would settle it.
The fifth item is deliberately left open. We could have applied a plausible percentage to your parts spend and produced a number. It would have sat on this page looking exactly like the four above it, and it would have been a guess wearing their clothes. Tomás confirms parts are recorded inconsistently and the office confirms the gap is sometimes reconstructed and sometimes absorbed. Nobody has counted, and no system holds the answer, because the missing records are precisely what would have to be counted.
Five figures your team confirmed during the engagement, written down now rather than reconstructed later. Loaded cost of $38 an hour for the office manager and $52 for the lead estimator. An average on-site variation at $840 and an average completed service call at $310. Eleven missed inbound calls a week, which is the only measured number in the engagement and comes from your phone system rather than anybody's memory.
These exist so that a year from now the question "did it work" has an answer rather than an opinion.
A full Terrain Map carries a dated twelve-month read on what AI is doing to your sector specifically, sourced and honest about going stale. This example does not include one, because a forecast about a real industry has to be researched and sourced rather than illustrated, and inventing one for a demonstration would be the exact thing this document is built to argue against.
Eighty-three thousand dollars is real, and it is the floor of this engagement rather than the point of it. It is the door. Every item above is the kind of thing a competent operations consultant would find, and if that were all we had, you should hire the cheapest one.
Part two is why you should not.
Part two
Everything here is evidenced the same way as part one. Someone said it, on a date, and it is quoted. The same standard applies, on material that is simply harder to see.
"We're at the point where a bad month hurts. There's no fat. If two big jobs slipped I'd be having a difficult conversation with the bank."
Ray Kowalczyk, owner · session 1 · 00:58:14
Stonepile is healthy and tight. There is no reserve of time or money to absorb a setback. That single fact outranks every preference about what should happen first. A twelve-week program with a payoff at the end is not available to this business, however good it would be, and any consultant who proposes one has not listened. It is why the first phase in part three returns capacity before it builds capability.
Asked what the business is actually for, Ray said: "we do good work and we turn up. That's it. That's the whole thing." Then, immediately: "Nobody's ever asked me that. Twenty-two people and nobody's asked me that."
That is a real identity and it is doing enormous work. It is why the crews stay, why the difficult buildings still get served, and why the pricing sits where it does. It has never been stated in a form anyone else could use.
"Two of the lads have been here eleven years. Their kids know my kids. That's not nothing when you're deciding how hard to push a schedule."
Tomás Rivera, field lead · session 3 · 01:02:40
Any change presented here as efficiency will be heard as a threat to people Ray considers obligations, and heard correctly. The honest framing is capacity returned to named people, because that is what it is.
"There's four outfits our size in the valley and we all know each other's crews. Two of them sold to the same buyer in the last three years. I get the letter about twice a year and I put it in the drawer."
Ray Kowalczyk, owner · session 4 · 01:24:10
This changes what the capture work in part three is worth. A business whose knowledge lives in two heads is fragile, and it is also close to unsellable on terms Ray would accept, because there is nothing to hand over except people and people do not transfer. Writing down what this business knows is the difference between an exit and a wind-down. That is a stronger reason to do it than any hourly saving in part one.
Ray is fifty-eight. His daughter does the books two days a week and is not taking the business on. Marcus, who holds twenty-six years of pricing judgment, is two years from retiring and volunteered this himself, unprompted, at the end of a session:
"Twenty-six years. There's no book. If I get hit by a bus you've got a problem."
Marcus Ojeda, lead estimator · session 2 · 00:44:51
Four capabilities rest on those two people: the week's crew assignment, pricing beyond the standard items, which buildings carry surprises, and the judgment about which work to decline. None of it is written down.
We raise this once, plainly, and then leave it with you. It is also the finding that reframes the whole engagement. Everything in part three that captures what people know is not an operational improvement. It is the beginning of the only succession plan this business currently has.
Every organization runs on things everyone senior knows and nobody has written down. Among people who already know, it never needs saying. A machine does not know, and it does not leave the space empty. It fills it with how such a situation usually goes and hands that back as finished work.
We found three. They are the highest-value material in this document.
"We don't chase new construction under about two grand. Everyone here knows that. I've never written it down because everyone here knows it."
Ray Kowalczyk, owner · session 2 · 01:04:22
Nothing sensible. Dana described what happened last month: "The new girl quoted a small new-build job because nobody told her not to. Ray had to ring them back and walk it away."
One walked-back quote is a phone call. That same rule missing from a system that drafts quotes at volume is a steady stream of work you have decided not to want, going out over your name.
A significant part of your pricing and scheduling accuracy is knowing which specific buildings carry surprises. Asbestos in the plant room. Access that costs half a day. A facilities manager who will not sign anything on a Friday. Two people hold this, and it transmits by new people finding out the hard way.
Already documented, because Marcus ran the experiment himself: "I put a job into one of the AI things to see what it'd come back with. It priced it like a textbook. No allowance for the fact that half these buildings have asbestos in the plant room."
That is not a story about a bad tool. The tool did exactly what it is built to do: in the absence of anything specific about Stonepile, it supplied the statistical average of how such a job usually goes, and it sounded confident. Marcus caught it because he has twenty-six years. The next person to try it may not, and the two people who could catch it are the two furthest along in their careers.
Ray's answer, and the fact that in twenty-two people nobody had asked.
Him. Every decision that needs it routes through one person, which is the same finding as the schedule living in his head, arriving from the other side.
Nothing can be delegated to a person or to a machine without it, because there is no way to check whether a decision was right. This is the gate under every other gate.
Two things are authoring this business that it did not decide and has not examined.
The job numbering came from Ray's previous employer. Dana, who uses it daily: "Nobody here has ever worked there. It doesn't match how we actually sequence anything." Probably not worth keeping, and the cheapest thing on this list to change. It matters more than it looks, because the numbering is what any future system would key on, and an inherited scheme that does not match the work becomes permanent the moment it is built into something.
Flat-rate labor plus materials, because that is how the trade does it. Ray: "That's just how it's done. I've never sat down and worked out whether it's right for us." Worth keeping? Unknown, and that is the finding. This is not a recommendation to change your pricing. It is the observation that the most consequential number in the business was inherited rather than decided, and that nobody here has ever run the test.
These sit apart from the section above on purpose, and the difference is worth holding. An unwritten rule of yours needs capturing. An inherited one needs testing, and possibly refusing. Capturing an inherited pattern into an AI system is the most expensive mistake available in this work, because it makes something nobody chose permanent and fast.
Partly, and better than most. The strongest signal in this engagement is one nobody would expect us to praise: Marcus tried a general assistant on a real job, on his own initiative, and immediately saw the answer was generic in a specific and dangerous way. The capacity to judge AI output already exists in this business. It exists in one person, and he is leaving in two years.
You have genuine judgment about what you will not do. Work you turn away, jobs you price differently, customers you manage carefully. Many businesses this size have no refusals at all, only capacity limits. What is absent is any form that judgment can travel in.
This is worth saying plainly because the industry gets it backwards. Governance is the layer most AI programs over-invest in, and governance alone produces a governed system nobody adopts. You have the opposite problem, which is a better one: the judgment is real and it is unwritten.
"We bought the scheduling module two years ago. Trained everyone for a day. Dana uses it, nobody else touched it, so I got the lads a simpler app."
Ray Kowalczyk, owner · session 2 · 00:58:30
A capability you are paying for is dormant. Rather than find out why the training did not take, the business bought a second tool. Any recommendation in this document inherits that pattern. An initiative that does not name who operates it, and what is different this time, will land exactly where the scheduling module landed.
Two things make this more hopeful than it sounds. Ray describes the pattern himself, without defending it. And he is ready to hand things over and has no mechanism for it, which is a good combination: it means the first move should give him something back rather than ask him for more.
"We don't really look back at a year. We just do the next week."
Dana Whitfield, office manager · session 4 · 00:58:12
No retrospective at any interval. No review of which jobs made money, which estimates were wrong and by how much, which customers cost more than they pay. Your phone system's missed-call report has been arriving for years and nobody reads it, which is the whole pattern in miniature.
One more thing belongs in this section, and it is about us rather than you. Ray has been pitched three times this quarter: "All the same deck with a different logo. I can't tell which of it is real and which is someone's demo." That is why every number in part one shows its arithmetic. Showing the working is not a stylistic preference here. It is the condition of being believed at all.
Written before the recommendations, on purpose. These are the places we would tell you not to put AI, and why.
Pricing judgment on non-standard work. Marcus's value is knowing which jobs will go wrong and pricing that in. That judgment should be captured and made available to a human, and never automated. The failure cost lands on margin and lands quietly, which is the worst combination available.
Deciding which crew goes to which site. This encodes who works well together, who a given customer will tolerate, and what a site needs on a given day. Support the decision with better information. Do not make it. The cost of a wrong pairing is a bad day on a customer's premises.
The customer conversation when something has gone wrong. Not a candidate at any point on this roadmap. It is the relationship, it is where a mechanical contractor's reputation actually lives, and it belongs to the people who own it.
Part three
Where you actually stand, in what order things can happen, and why the order is not a preference.
Four gaps sit between an organization and AI that works the way it does. They are sequential, and each one only opens if the one before it holds. This is where the evidence puts Stonepile.
Gap one · articulation
Not yet, in terms anything could act on. There is a real answer underneath and it is a good one, but it lives in one head and has never been asked for. Three concrete operating rules sit in the same condition.
What this gates: everything. A business that cannot say what it is for cannot say what it should refuse, cannot check whether a machine's answer is right, and cannot delegate to a person or a system with any confidence. Firmed, this ground would bear the rest: your refusals become checkable, the pricing judgment becomes teachable, and every gap below becomes worth attempting. None of it needs a tool, which is why it can start this week.
Gap two · carry-in
No. One general assistant, in occasional use by one person, running on nothing but the question typed into it. It has no access to how you price, what you refuse, or which buildings are difficult.
What this gates: any use of AI in estimating or quoting. Until something of this business travels into the tool, its output is the industry average with your name on it, and the only defense is that one person happens to be experienced enough to catch it. Firmed, this ground would bear an estimating assistant worth having, because it would know what you refuse and which buildings are difficult.
Gap three · practice
No, and it would be premature to try. Use is one person, occasionally, in a chat box. Nothing is saved, nothing is shared, each session starts from nothing.
What this gates: any durable return from AI at all. This is the gap most vendors sell into. Selling it before the two above are closed is how a business ends up with tools nobody uses, and that has already happened here once. Firmed, this ground would bear daily use that survives a bad week, because your context would live somewhere rather than being retyped each time.
Gap four · loop closure
No. No retrospective, no evaluation of what a tool got right or wrong, no route by which a pattern in the work reaches a decision. The missed-call report arriving unread for years is the same absence in a simpler form.
What this gates: whether any of this compounds. This is the gap almost nobody closes, and it is the difference between buying software and getting better at what you do. Firmed, this ground would bear improvement you did not have to buy: you would get better at pricing from your own completed jobs, year over year.
Most organizations we meet are at gap one and have been sold gap three.
Stonepile is one of them. Three vendors have pitched this business in a quarter, and every one of them was selling into the third gap while the first was open.
That is not a criticism of the vendors. It is what the market sells. It is also why the order below cannot be rearranged for convenience.
Seven initiatives. Each one traces to findings above, carries what it is worth or an honest gap, and says which of the four gaps it moves. Three of them move no gap at all and are here because they return capacity, which is a legitimate reason and a different one.
Phase one · first, and by itself
Nobody has to change how they work: a list appears and somebody rings it. It moves the fourth gap, because a signal your systems already generate stops being ignored and starts producing action, which is the cheapest possible version of closing a loop.
It goes first because it is the largest quantified return, the least effort, and visible to everyone within a week. In a business with no slack, the first thing has to pay before it costs.
Phase two · quick wins
This moves the third gap, and it is explicitly the test of whether this business can adopt anything at all. It is also the exact thing that failed two years ago, so repeating it without addressing why would be the most expensive mistake available here. It goes ahead only with a named operator and a two-week check.
It returns capacity and it is visible, and that is the whole case for it. Dana is the right operator because she is the one person who did adopt the last system, and the risk is not her: it is that this is the second thing bought to fix a workflow, and it needs to visibly work within a fortnight.
Phase three · the core
This carries no number and it outranks everything with one. It moves the first gap, which gates all three others. Both people named the problem themselves, which is the best possible starting position, and willingness is the scarce input here rather than time or money. That willingness is not permanent, and Marcus is closer to retirement every month.
We are sequencing this by consequence rather than by return, deliberately, and flagging it rather than burying it. If you do one thing from this document that is not the first move, do this one.
Your own prior work becomes context a system can reach, which is the first real carry-in this business would have. It pairs naturally with the capture work above, being the same conversation with a different purpose.
Recovers revenue already earned. Highest adoption risk on the list, and worth being blunt about why: it asks field crews to change what they do on a bad day, which is the exact condition under which the current process already fails. It only works if it is faster than the text message it replaces.
This exists to produce a number rather than to act on one, and the honest label for it is measurement.
Two loops, and neither needs any AI to start. That is the point of them.
Check the estimate against what the job actually cost. Every estimate carries an assumed labor figure and every completed job carries a real one. Nothing compares them. Side by side monthly, that tells you which kinds of work you price wrong and by how much, in your own numbers, with nobody asked to guess. It is the cheapest possible first loop, which is exactly why it should be first: build the habit of looking back on something that pays for itself, before anything harder is attempted.
Write down what the difficult buildings teach, where the next person will find it. Every job on a known-difficult building produces a lesson that currently stays in someone's head. Captured once per job, in a sentence, that record becomes the thing that makes an estimating assistant worth having later. In the meantime it is what a new hire reads instead of finding out the hard way.
Note the sequencing on that second one. It is a loop that produces the context that closes the second gap. The loops and the gaps are not separate programs. Done in the right order, closing one feeds the next.
A twenty-two person business that has already watched a rollout fail will read a large program as the same thing happening again. That is a correct reading, and the plan is shaped around it: one thing first, small, paying for itself, with a named operator and a date by which it either worked or did not.
The first move
Under a week of work, on a number your own phone system already reports. Eleven calls a week, roughly fifteen percent of which would have been real work, at $310 a job. Nobody changes how they work. Somebody gets a list and rings it.
Start here for three reasons. It pays before it costs, which is the only kind of first move available to a business with no slack. It is visible to everyone within a week, in a business that last watched a rollout fail. And it is the smallest possible instance of the thing this whole document is about: something your systems already noticed, that nobody was looking at, coming back and changing what happens.
The baselines are on the page. In ninety days you will be able to prove what changed rather than assert it.
Four sessions across three weeks with the people who do the work: the owner, the office manager, the lead estimator, the field lead. Twenty-eight evidence atoms, every one a quote with a name, a session and a timestamp. Fifteen findings on the operational side, fourteen on the other, four gap reads, seven initiatives, three places we told you not to put AI.
You keep this document and everything under it, including the evidence file that every claim above links back to.
What we would do next is not on this page, on purpose. A Terrain Map that ends by selling you the next thing is a brochure. This ends where it should: you can now say what your business knows, what it has never had to say, where it actually stands, and what to do first. What you do with that is yours.