Before You Buy Any AI, Count How Many of Your Solar Plan Sets Come Back Redlined by the AHJ

One in three plan sets comes back redlined. That is your baseline.

Pull your permit log for the last twelve months and count two things: how many plan sets you submitted to an AHJ, and how many of those came back with redline comments before anybody stamped them. In most residential solar and storage shops doing 150 to 400 installs a year, the second number lands between a quarter and a third of the first. Nobody writes it on the whiteboard. It sits there as two or three extra weeks per job and a customer emailing to ask why the panels on the roof are still not on.

That number is where your first AI project should start. Not because permitting is interesting, but because it is the one process in a solar business that is repetitive, documented and dated on both ends. Every submittal has a date. Every rejection has a comment attached. You can prove a change moved it or prove it did not, inside one quarter.

Proving an AI paid for itself in a solar business means picking one repeatable process, writing down what it costs today in hours, in rejection rate and in calendar days, and then comparing those same three numbers ninety days after you switch it on. If you cannot state the before number today, you will not be able to argue about the after number in October.

What the permitting desk actually does between the signature and the sticker

Here is the honest version in a 6-crew shop. The energy consultant closes on Thursday night. The site surveyor gets there the following week, shoots the roof with Scanifly, photographs the main service panel label, the busbar rating, the meter and the attic framing. The designer builds the array in Aurora Solar or OpenSolar, sets the string plan, and exports a plan set: site plan, roof plan, single-line diagram, placard sheet, equipment cut sheets, and the structural letter if the AHJ wants a stamp.

Then the permitting coordinator takes over, and this is the part nobody has ever really systematised. She knows that this county wants a 36-inch ridge setback drawn even though the fire code allows the alternative, that the neighbouring city rejects any single-line diagram missing the rapid shutdown device by part number, and that one reviewer will not accept a load calculation under NEC 220.87 without twelve months of utility usage attached. That knowledge lives in her head and in a spreadsheet with a tab per jurisdiction. When she is out for a week, the rejection rate goes up and everyone notices.

The workaround everyone pretends is fine: submit, wait nine days, read the redlines, fix the drawing, resubmit, wait again. The same loop runs at the utility, and a bounced interconnection application is worse, because the queue position resets and the PTO date slips past the customer's first loan payment.

Why Deloitte's 84% is really a story about picking one process

Deloitte's State of AI in the Enterprise 2026 found that 84% of organisations investing in AI report positive returns. The interesting part is not the percentage; it is the pattern underneath it. The companies reporting a return did not roll a general purpose assistant out to everyone and hope. They took one messy, document-heavy process, put a human review step in front of the output, and proved either time saved or errors reduced before widening the scope to anything else.

For a solar installer, permit and interconnection packet preparation is that process. The 2026 tools are genuinely different from what you tried two years ago in one specific way: you can hand the assistant the entire packet plus your last forty redline letters from that same jurisdiction at once, and ask it to find everything in this drawing set that this reviewer has objected to before. Claude Cowork, which launched in January and went to web and mobile in July, and ChatGPT Work, which launched 9 July on GPT-5.6, both take a goal and work across your files rather than answering a question in a chat box. The output is a checklist of specific things to fix, with the prior rejection letter cited next to each one, which your coordinator reads in four minutes.

flowchart TD
  A["Contract signed, survey photos in Scanifly"] --> B["Designer builds plan set in Aurora"]
  B --> C["Assistant reads packet against this AHJ's last 40 redlines"]
  C --> D{"Anything this reviewer rejected before?"}
  D -->|Yes| E["Coordinator fixes it before submittal"]
  E --> C
  D -->|No| F["Submit to AHJ portal or SolarAPP+"]
  F --> G["Log first-pass result against that jurisdiction"]
  G --> C

The four numbers to write down before you switch anything on

Capture the baseline first, over the last twelve months, from records you already have. It takes an afternoon.

• First-pass approval rate by jurisdiction. Submittals approved with no corrections, divided by total submittals. Break it out per AHJ, because your average hides the two cities that are eating you alive.
• Calendar days from signed contract to permit issued, and separately, days from application to utility approval to energise. These are different animals and they fail for different reasons.
• Coordinator hours per resubmittal. Have her time five of them honestly. It is rarely under two hours once you count re-exporting the drawing, re-uploading, and the follow-up call.
• Post-contract cancellation rate, and the average age of a cancelled job. If your cancellations cluster at day 70, waiting is what is killing them, not price.

Write those four on one page, date it, and put it in a drawer. Change nothing else for ninety days. If you also switch proposal tools or hire a second coordinator in that window, you have destroyed your own experiment.

Tuesday, 7:40 a.m., at the permitting desk

The coordinator opens a folder with six packets ready to go out. For each one, the assistant has already produced a one-page review: this county set is missing the ridge setback dimension on the roof plan, which that reviewer flagged on three of your last eleven jobs; this set lists an inverter on the placard sheet that does not match the single-line diagram, probably because the warehouse substituted a unit; this one is fine. It also drafted the interconnection application from the design file, waiting for her to check and sign.

She fixes two, rejects one of the assistant's suggestions because it misread the fire code amendment, and submits all six by 9:15. What used to be Tuesday and half of Wednesday is now ninety minutes. Nothing about her judgement was replaced; the reading was.

The arithmetic on 240 installs a year

Assumptions, all illustrative: 240 installs per year, average contract $24,780 (an 8.4 kW system at $2.95 per watt), first-pass approval currently 69%, resubmittal costs 2.4 coordinator hours at a fully burdened $34 an hour, contract-to-PTO averaging 104 days, and a line of credit at 11%.

The hours are the small win: 108 hours back, about $3,700. The real money is the twelve days. At $24,780 a job and 240 jobs, you turn over about $16,300 of contract value per calendar day, so work in progress falls from roughly $1.70m to $1.50m. That is about $196,000 of working capital you stop carrying, worth roughly $21,500 a year in interest at 11%. Add the jobs that do not cancel because they were not left sitting at day 70: one saved job at a 22% gross margin is another $5,450.

Where the coordinator still has to be the one who decides

Three places to keep a human firmly in the chair. First, anything that carries a professional stamp. A structural letter for a 1962 truss roof with a 14 kW array and a wall-mounted battery is an engineer's signature and an engineer's liability, and no assistant should be drafting the reasoning behind it. Second, code interpretation where the local amendment differs from the model code. Setback and rapid shutdown requirements vary by jurisdiction and the assistant will confidently apply the wrong version; catching that is exactly why she reads the one-pager instead of auto-submitting.

Third, the conversation with the plan reviewer. When a set gets stuck, what moves it is your coordinator calling the person she has spoken to forty times and asking what he actually wants to see. That relationship is an asset on your balance sheet even though your accountant will not list it. Automate the reading and the drafting. Do not automate the phone call.

One more limit: this speeds the paperwork, not the utility queue. If your utility takes 35 days to review interconnection applications regardless, a clean first submittal saves you a second 35-day round. That is all it can do.

Frequently asked questions

We already use SolarAPP+ in half our jurisdictions. Does this still matter?

Yes, and it matters more, because SolarAPP+ approves the eligible cookie-cutter jobs instantly and leaves your coordinator with only the hard ones: the panel upgrades, the ground mounts, the battery backup with a partial-home subpanel, the historic districts. Your average approval time will look great while your worst jobs get worse. Measure first-pass approval on the non-SolarAPP+ jobs separately or you will be fooling yourself.

How do I stop this from turning into a six-month project?

Scope it to one jurisdiction, the one with your worst rejection rate. Gather that AHJ's last forty correction letters into a folder, run the next twenty submittals through the review step, and compare. If it does not move that one city's first-pass rate in twenty jobs, it will not move your whole business, and you have spent three weeks finding out instead of two quarters.

Who owns this internally? I do not have anyone technical.

The permitting coordinator owns it, because she is the only person who can tell whether a flagged item is real. Give her two hours a week of protected time for the first month and let her keep a running note of every suggestion that was wrong. That note is what you use to decide whether to widen the scope, and it is also your defence against a vendor telling you it works everywhere.

Faster permits mean more customers reaching the install and PTO stages at once, and that shows up as inbound calls asking where things stand. CallSphere builds AI voice and chat agents that answer the phone and web chat, take the caller's address and job details, and book the site survey or service visit into your calendar around the clock, so the same coordinator who just got her Tuesdays back does not lose them again to the phone. It does not touch your plan sets or your interconnection applications; those stay where they belong.