Can AI Design a Solar Array? What Design Software Does and Doesn't Do

Yes, for a first draft. Aurora AI builds a 3D roof model from an address in under ten seconds, and AutoDesigner can propose a system size and panel layout from a customer's electricity use. A qualified designer must still verify roof geometry, setbacks, shading, stringing, electrical limits and local rules before anything is sold or submitted.

The useful way to think about it: AI removes the drawing work, not the responsibility. A tool that places panels in seconds gives you something to check, and checking is where small installers lose money or safety. A wrong shade assumption costs you production complaints a year later; a wrong string calculation can mean an inverter that never starts on a hot afternoon or one that is overdriven on a frosty morning.

Follow me on Instagram@sagnikteaches

In a small installation firm that risk sits with whoever signs off designs, often the owner. AI changes the job rather than shrinking it: less time clicking panels into place, more time checking a layout the software produced against survey photos, datasheets and the rule that applies on this street. For what each design tool costs once you count the proposals you actually send, see AI solar design software compared by cost per proposal.

Connect on LinkedInSagnik Bhattacharya

What the AI part actually produces

Vendors use "AI" for several different jobs, and mixing them up is how people over-trust a screen. In Aurora's product, for example, there are at least three separate pieces, and each deserves a different level of scrutiny.

Subscribe on YouTube@sagnikteaches

• Roof modelling. Aurora's own page says Aurora AI produces a 3D roof model in under 10 seconds, and that an address and one month's utility bill are all it needs to generate a home solar design. This is a computer-vision job: it reads aerial imagery and infers planes, pitches and edges. Aurora also says it has been run over 3 million times, which tells you it is common, not that any single model is right.
• Layout and sizing. Aurora's help centre and API documentation describe AutoDesigner as having two modes: an energy mode that places panels to hit a yearly kWh target or a percentage of the customer's annual consumption, and a Max Fit mode that fills the roof. It works only with the panels and inverters you select, the orientations you allow and a minimum TSRF (total solar resource fraction, the share of the ideal sunshine a spot actually receives after tilt, direction and shade). This is optimisation, not vision. It is only as good as the roof model and the constraints you feed it.
• Yield and shading estimates. Production figures come from calculation engines that use the model, the imagery and a shading analysis. They inherit every error upstream.

OpenSolar offers similar auto-design features. Its core app is free; since April 2026, per-project API access and accounting connectors have been paid extras, priced by region rather than as one published figure. Prices and plan names move, so check the vendor page for your account before you quote a number to anyone. Whichever tool you use, the principle is the same: the output is a hypothesis about a roof you have not yet stood on.

Five things a roof model cannot know from the sky

Imagery is taken on a particular day, from above, at a fixed resolution. Anything that is not visible from above, or has changed since, is invisible to the model. Here are the five gaps that cause most trouble.

Roof condition and structure

The model can tell you the plane is 7 metres wide. It cannot tell you the covering is brittle, the rafters are undersized for the added load or the battens are rotten. Those are surveyor checks, and the layout depends on them. If the roof needs work first, the layout may move or be dropped altogether.

Obstructions smaller than a pixel or newer than the photo

Soil vent pipes, flue terminals, roof windows, satellite dishes and extension roofs are the usual misses. A vent stack half a metre from where the model thinks the third panel sits is enough to change a row.

Edge and access allowances

Many authorities require clear paths or setbacks from ridges, edges and hips for firefighter or maintenance access. The rules differ by authority and change over time. AI tools can apply setbacks you configure; they cannot know which version applies on this street. Aurora's homepage, for example, describes its output as a "code-compliant panel layout", which in practice means compliant with the settings your account holds, not a guarantee about the rule the inspector will apply. The designer owns that lookup, and it should be written on the job record with the date it was checked.

Shading beyond the model

Aurora says its shading comes from LIDAR and high-definition imagery. That is a strong starting point. It does not know the neighbour is about to remove, or plant, a row of trees, or that a new extension is going up next door. A site shade reading or fresh photos are cheap insurance.

The electrical side of the house

Where the consumer unit or main panel sits, whether it has a spare way, the cable run to the inverter and the state of the earthing are all survey items. A beautiful roof layout is useless if the inverter has to go 30 metres from the board. A structured survey captures all five gaps in one visit; a solar site survey checklist with photos sorted into a report shows a format that feeds straight into the design check.

A worked small roof, step by step

This is an illustration with made-up but plausible figures, not a real project. Say a semi-detached house has one south-facing roof plane, 7.0 m wide along the eaves and 3.6 m long from eaves to ridge. The AI tool produces a model and offers a layout. You are the designer checking it.

Step 1: reconcile the usable area

The tool shows the whole plane. You take off a 0.45 m edge setback at the eaves, ridge and each side (the figure here is an assumption; use whichever allowance applies to the job and write down where you got it). That leaves a usable rectangle of 7.0 - 0.9 = 6.1 m wide and 3.6 - 0.9 = 2.7 m long. A soil vent pipe sits in the lower left of that rectangle, and the model does not show it.

Step 2: test the layouts the tool did not offer

Take a typical 400 W module measuring 1.134 m by 1.722 m, with the small gaps between modules ignored for simplicity. In portrait, the 2.7 m height fits one row (1.722 m), and the 6.1 m width fits 5 modules (5 x 1.134 = 5.67 m): 5 panels, 2.0 kW. In landscape, the 2.7 m height fits two rows (2 x 1.134 = 2.27 m) and the 6.1 m width fits 3 modules (3 x 1.722 = 5.17 m): 6 panels, 2.4 kW. If the vent pipe blocks one landscape module, the count drops to 5 unless you reposition. The tool may pick one orientation; you compare both and keep the one that survives the obstruction.

Step 3: check shading against the worst hours

Say the tool's annual shading estimate is a few percent, driven by a tree to the south-west. You look at the survey photo and see that the tree is taller than the model assumed and shades the lower-left modules on winter afternoons. If those modules sit on a string with the others, one shaded module can drag the whole string down. That is the point where you decide between module-level electronics (microinverters or optimisers) and a plain string, or move the layout.

Step 4: check the string against the inverter

Say the datasheet gives the module an open-circuit voltage (Voc) of 37.0 V at 25 C with a temperature coefficient of -0.26% per degree, and a maximum power voltage (Vmp) of 31.0 V with roughly -0.30% per degree. The site's record low is -10 C, so cold Voc = 37.0 x (1 + 0.0026 x 35) = 37.0 x 1.091, which is about 40.4 V per module. Six modules in series give about 242 V, comfortably under a 500 V inverter input limit. For the hot end, assume a cell temperature of 70 C: Vmp = 31.0 x (1 - 0.003 x 45), about 26.8 V. Three modules would give about 80 V, right at an inverter whose minimum MPPT voltage is 80 V, so a three-module string risks dropping out on hot days. Six modules in one string give about 161 V and sit well inside the window. All these figures are examples; the real ones come from your datasheets, the inverter manufacturer's sizing tool and the temperatures your authority or standard tells you to use.

Step 5: write down the result

The signed-off record for this roof reads something like: 6 landscape modules (2.4 kW), vent pipe avoided by shifting the array 0.3 m right, edge setbacks confirmed against the current local rule on a stated date, one string of 6 within inverter limits, shading re-checked against survey photos. Anyone opening the file later can see which parts the tool drafted and which the designer confirmed.

Using ChatGPT or Claude alongside the design software

General assistants are useful around the design, not inside it. They are good at turning your survey notes into a clean checklist, explaining a datasheet line in plain words, and catching arithmetic slips when you give them figures to check. They are poor at inventing the numbers themselves. Paste in your datasheet values and ask for the working, never for "the answer". If you use a consumer plan, review its training settings; business plans such as ChatGPT Business (from $20 per user per month billed annually, two-seat minimum) and Claude Team don't train on your content by default.

A prompt for checking your own arithmetic

I am checking a solar string. Use only the figures I give you; if one is missing, say so instead of assuming.
Module: Voc 37.0 V at 25 C, Voc coefficient -0.26 %/C, Vmp 31.0 V, Vmp coefficient -0.30 %/C.
Site design temperatures: lowest -10 C, highest cell temperature 70 C.
Inverter: max DC input 500 V, MPPT window 80-450 V.
Modules per string: 6.
Show the cold Voc per module and per string, the hot Vmp per string, and say whether each is inside the inverter limits. Then list what I should confirm on the manufacturer's sizing tool.

A reasonable output states cold Voc near 40.4 V per module and 242 V per string, hot Vmp near 26.8 V per module and 161 V per string, and "inside limits". What you would fix: if the assistant quietly assumes a different coefficient or a different cell temperature, reject the answer and make it use yours. Treat any reply that does not show its working as unusable.

Four illustrative ways an auto-layout goes wrong

None of these is an account of a real job. Each follows directly from one of the gaps above, and each is the kind of slip a sign-off step should catch before the customer does.

The dormer that wasn't there

Imagery was two years old. The homeowner had since added a small dormer, which removed 1.5 m of usable height from the lower half of the plane. The proposal promised 8 modules; the survey found room for 6. The fix was a call to the customer and a revised price. The process change: always compare the model with a current photo before the price leaves the building.

Two planes, one string

The auto-layout put three modules on an east plane and three on a west plane in a single string to hit the target size. The voltages looked fine on paper, but with different orientations the current mismatch cost production across the day. The designer split the planes across separate inverter inputs. Rule of thumb: check that every string has one orientation and pitch unless the equipment is designed for mixed input.

Setbacks left at the default

A tool template carried a setback value from another job. The array looked neat and was non-compliant with the current local rule, which the designer only spotted at the permit-prep stage. Reconfirm the setback on every job instead of trusting the saved template.

Shade that moved with the seasons

A proposal quoted annual output from a model whose tree line was drawn at summer height in the imagery. The customer's meter showed an unhappy December. Give an honest range, and record the winter shade picture beside the estimate.

Where the free tools fit

The pages that rank for this topic are mostly vendor and calculator pages: they show what the software can do. They rarely say what the designer owns. Free or low-cost tools such as OpenSolar are a fair way to get proposal sketches out quickly, and a paid platform adds more modelling and contract features. The choice should follow your volume and your sign-off process rather than the length of the feature list. If you quote fifteen homes a month, a fast layout tool plus a strict check step pays for itself. If you quote two, the check step matters far more than the tool. For installers weighing a full stack, the comparison of CRM, design and tracking tools covers how design software connects to the rest.

A sign-off checklist for the designer

Paste this into your project template and require a name and date against each line before a proposal is sent or a permit pack is built.

1. Imagery date and currency. Is the model newer than any known building work? Compare against a current photo or site visit.
2. Roof planes and pitch. Measured or confirmed on survey, not only read from the model.
3. Obstructions. Vents, flues, windows, dishes, chimneys and parapets marked and clear of the array.
4. Setbacks and access paths. Current local rule identified, value recorded, date checked.
5. Structure and covering. Surveyor's note on rafters, fixings and roof condition.
6. Shading. Annual figure plus a note of the worst hours; site reading or photo taken.
7. String design. Cold-end voltage and hot-end voltage inside inverter limits, using the manufacturer's tool; one orientation per string unless the equipment allows otherwise.
8. Cable and protection. Run lengths, voltage drop, isolators and protective devices specified by the person qualified to do so.
9. Point of connection. Main panel capacity and spare way confirmed.
10. Yield statement. Production is given as an estimate with assumptions listed, never a guarantee.
11. Drafted-by record. Which parts the tool produced and which the designer changed.

Letting AI draft layouts without losing control

Start by deciding where AI output is allowed to go. A reasonable rule for a small firm: AI layouts may go to the customer as a sketch marked "subject to survey". They may not go into a signed contract or a permit pack until the checklist above is complete. That one rule prevents most of the damage.

Next, time the check. A small roof like the one above should take a competent designer 15 to 30 minutes against the checklist; a complicated multi-plane roof takes longer. If checking takes longer than designing from scratch, either the tool is wrong for your jobs or the checklist is vague. Fix the tool choice or sharpen the checklist, not the corner you cut.

Third, keep a small error log. Every time the survey disagrees with the model, note the type (obstruction, pitch, shade, structure) and the size of the difference. After thirty jobs you will know which errors recur on the roofs you work on and can brief the survey step to look for them first. This is also the evidence that tells you whether a more expensive tool is worth it.

Finally, handle customer wording honestly. A sentence such as "This layout was drafted with design software and will be confirmed by a qualified designer after a site survey; the final number of panels and the annual production may change" costs nothing and avoids an argument later. If your contract mentions AI use, how to mention AI use in client contracts and proposals has wording to adapt.

When the tool and the survey disagree

Disagreements are normal and not a sign the tool is useless. Believe the survey. Update the model if the tool lets you, or redraw by hand, and write the cause in the error log. If a pattern emerges, for example the tool overstates pitch on shallow roofs, add a specific line to the checklist. Tell the customer early when the change affects price or output; a revised quote sent a day after the survey reads as care, while one sent after installation starts reads as a problem.

When a roof is awkward, with several planes, hips or heavy shading, consider a human design service or a plan-set service from your software vendor instead of forcing the auto-layout. Aurora, for instance, lists expert design and plan-set services alongside its software. That is a fair use of the money when the alternative is hours of your designer's time.

The same discipline carries into paperwork. Once the layout is signed off, drafting the permit application with AI works from the checked record, not from the tool's first output, and a clear customer proposal with plain-language FAQs can then state what was verified and what remains an estimate.

Questions installers ask about AI solar layouts

Can I send an AI-generated solar layout straight to the customer?

Send it as a proposal sketch only. The roof model comes from imagery, so measurements, obstructions and shading can be off. Have a designer or surveyor confirm the layout against a site visit or photos before you quote a fixed price or sign a contract.

Does AI replace the site survey?

No. Imagery cannot show roof condition, rafter size, attic access, the position of the consumer unit or a fire-access requirement that the local authority applies. Use the AI model to prepare the survey, then record what the survey finds against each assumption.

Can ChatGPT or Claude check my string sizing?

They can talk you through the method and spot arithmetic slips if you paste the datasheet figures. They should not be the final check. Run the numbers in the inverter manufacturer's own sizing tool or your design software, and keep that output on file.

Which part of an AI solar design is most often wrong?

Shading and roof edges. Trees grow, imagery ages, and small obstructions such as vents, satellite dishes and chimneys are easy to miss or mis-place. Compare the model with fresh photos and a shade reading on site.

Further reads

• Solar AI Chatbot for Installers: Qualify Leads by Bill, Roof and Shade — Qualify enquiries before a design is drawn at all.
• Solar Lead Generation: Cost per Lead, Cost per Survey and AI Follow-Up — Lead costs and follow-up speed for installers.
• Solar Installation Quality Checklist: Sort Photo Evidence With AI — Photo evidence that the installed array matches the signed-off design.
• Solar Battery Quote Template: Options Explained With AI, No Backup Promises — Explain battery options in a quote without backup promises.
• AI for Solar Installation: Five Jobs a Small Installer Can Automate — Other jobs in a small installer worth automating first.
• Solar Installer Marketing: Local Pages and Case Studies Written With AI — Marketing pages that bring in the roofs you design.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Aurora Solar's Aurora AI page and homepage; Aurora help centre and API documentation on AutoDesigner; OpenSolar's published pricing notes; inverter and module datasheet conventions.