The same memory, with no agent anywhere in the workflow.
A junior estimator on a roof at seven in the morning, scoping a rooftop replacement. Nothing here is drafted or generated. He asks four questions, gets three answers and one honest refusal, and makes every call himself. This is what the memory is worth before a single agent exists.
The value in this workflow is how long a person on site waits for an answer from the office, and how often a call gets made on a guess. We would measure both on your own field work. The diagnosis measured quoting and not this one, so we do not put a number on it.
Bexley Foods · Store 104
Cody is doing exactly what he did last year. The form is the only thing that changed.
The walk itself is the most valuable hour in the whole quoting workflow and nothing about it should be automated. A person has to stand on the roof and look at the clearances. What was wrong before was not the walk, it was that half of what he noticed lived in a notebook and a camera roll and never reached the bid file.
The fields are not generic. They are the things that turned out to move the price on jobs like this one, which is something the diagnosis worked out by reading twelve months of bids against what those jobs actually cost.
- Who is here
- Cody Sturgill, second-year estimator. On his own.
- What is running
- A form and a search box. No agent, no drafting, no generation.
- Where it saves
- Straight into the site record for Store 104, inside Holcombe's own tenant.
7:41Nothing has been priced and nothing has been drafted. Cody is writing down what he can see, in a form that puts it where the bid will need it later.
Answers cite where they came from
Every answer names the entry it came from, so Cody can tell a company position from a guess.
The second answer is the one worth looking at twice. It is not a number, it is a rule with its reasoning attached and the conditions that change it. That is what lets a second-year estimator apply a senior estimator's judgment without either of them being in the same place, and it only works because someone sat down and wrote the rule out properly.
The first answer is a site record. Holcombe has been to Store 118, and what that job actually cost is now something the company knows. It used to be something Doug remembered.
- Elapsed
- Under two minutes, standing on the roof.
- Before this
- Cody would have called Doug. Doug would have been on another roof, and the answer would have arrived that afternoon or not at all.
- Who decided
- Cody. The memory answered three questions. It did not price anything and it did not tell him what to do.
7:52Three questions, three answers, each one citing the entry behind it. None of this required an agent. It required the answers to exist in one place, with someone's name against them.
Answers cite where they came from
· Lawton LAW-204, post-job note: adapters used, added a day.
· Store 121, post-job note: curbs reused, no issue reported.
Added to the gap list. Cody's call on this job is recorded against Store 104 either way.
Nothing in the memory covers this question, and the answer says so.
A tool that answers everything is a tool nobody can trust, because the user cannot tell which answers are real and which were filled in. This one offers the two related notes it does have, and labels them as not being a company position.
Cody still has to decide, on the roof, this morning. He does that the way he always has. What is different is that the question does not evaporate: it is now on Doug's list with a number against it, and the next estimator who asks will get an answer.
- The yellow
- Marks the one moment the system hands to a person. It appears here and nowhere else in this demo.
- What happens next
- Doug rules on it in the weekly block. It comes back as an approved entry, the same path W-12 took in the answer desk demo.
- Gap count
- Eleven becomes twelve today, and eleven again when Doug signs. The list is meant to move.
7:56It did not guess. That is the behaviour every one of these workflows depends on, and it is built into how the memory works.
Saved · Store 104
What he leaves behind is a complete scope, not a memory of a roof.
His reasoning on the curbs is recorded next to the decision itself. When this job closes and the adapters either were or were not the right call, there is something to compare the outcome against, and that is how G-14 eventually gets answered properly rather than by opinion.
| Scope as walked | 3 RTU replacement, 7.5 ton, crane placement, exterior stair access, case covers over aisle 4. |
| Pricing rules that apply | DW-7 at 12 percent. Store trades 6am to 11pm, so night work is likely and the adder holds. |
| Comparable job | BEX-118, closed at 23.1 percent against 23.6 estimated. |
| Open question | G-14, curb reuse. Cody priced adapters and wrote down why. |
| What happens next | When Bexley requests the bid, this record is what gets priced. Nobody has to remember the roof. |
8:04Thirty-eight minutes on the roof and the scope is complete, sourced and sitting where the bid will need it. The quote drafter demo is what happens to this record later, and it is worth noticing that this step had to go right first.
The only software here is a search box and a form.
- What the model does
- Finds the right entry and quotes it with its source. It does not price, draft, decide or recommend. When nothing covers the question it says so. That behaviour is written into the instructions file as a rule.
- Where the knowledge lives
- The same memory the five agent demos read from. Nothing on this page needed a separate build, which is the whole argument: the core was already there.
- What gets written back
- The scope, the photos, Cody's reasoning on the curbs, and the raised gap. Doug's eventual ruling on G-14 returns as an approved entry that every later job reads.
- Buildable today
- A SharePoint list with a mobile form, and search over the memory scoped to the tenant. Both are inside Microsoft 365 Business Premium, which Holcombe already pays for.
If Holcombe never built a single agent, this workflow would still be better than it was, and the reason is the same reason the agents work. The knowledge was written down properly, with its reasoning, its exceptions and somebody's name against it.
App interfaces on this page are simplified illustrations, not vendor screenshots.
The same thing elsewhere
At a distributor this is a rep at a customer site, checking what that account has bought, what the current tier gets them and what the lead time actually is, before promising anything. At a professional firm it is a partner in a scoping meeting, answering what the firm has done on work like this and what it charged, without going back to the office first. The workflow changes and the memory underneath it does not.