North Energy, a residential solar company, kept everything it knew in 73 files on a shared drive. Warranty terms. A different permitting contact for every municipality the company builds in. Sales scripts. Finance program details. Training decks. Some of the files were scanned PDFs that no text search could read.
That setup charged the company twice for the same question. In the office, someone who needed a warranty term or a permitting contact either dug through folders or pulled a supervisor off their own work. In the field, a rep sitting at a customer's kitchen table with a question about a finance program's terms had nothing to reach for. The training binder doesn't come to the appointment.
What got built
North Energy's team worked with Level 99 to build an AI support agent that lives inside the company's internal tools and works from a phone. A staff member types a question in plain language and gets an answer pulled from the company's own documents, with the source files named underneath every answer.
The build, in about a week:
- A pipeline that re-syncs from the company's shared drive every night, picking up new and edited documents and leaving the rest alone.
- Text recognition applied to the scanned PDFs, so their contents became searchable for the first time.
- 2,570 searchable passages indexed from the 73 source documents.
- A branded chat widget dropped into existing internal pages. No new app, no new login.
What changed for two teams
For ops, the questions that used to interrupt a supervisor now go to the chat box first. The knowledge base maintains itself: someone edits a document, and the agent knows about the change by the next morning. Nobody uploads anything to a separate system.
For sales, the training library travels to the appointment. A rep can pull the exact objection-handling language or a finance program's real terms mid-conversation instead of promising to follow up later. Those follow-ups are exactly the kind of gap the sixth-call problem describes: a promise to call back is where a deal starts to die. New reps also train against the same source material the top closers already use, instead of a binder nobody updates.
Every answer cites the file it came from, so a staff member can open the actual policy and check it. Nobody has to take the agent's word for it.
The math, shown so you can check it
- Lookups dropped from 2 to 10 minutes down to under 30 seconds. Call it up to 95% faster. Digging through folders was the fast option; waiting on a free supervisor was the slow one.
- That adds up per person. One employee making three lookups a day at roughly five minutes each gets back about 15 minutes a day, over an hour a week. Multiply that by headcount and it's real time.
- Escalations used to cost two people, not one. A question that reaches a supervisor burns both the asker's time and the supervisor's. If even half of daily questions used to escalate, cutting that in half frees supervisor hours for the work only a supervisor can do.
- Upkeep is zero. The knowledge base re-syncs on its own every night. Nobody maintains it by hand.
What North Energy hasn't measured, and what this post won't claim, is close rate. Nobody tracked whether a rep who answers a finance question on the spot closes more deals than one who promises to follow up. The honest version: every "let me get back to you" that becomes an answer at the table is one fewer follow-up call to make, and follow-up calls are where deals stall. That's a structural reason to expect a lift, not a measured one.
Why this isn't just a solar-company problem
North Energy's documents happened to be warranty terms and permitting contacts. The shape of the problem shows up anywhere a company's real answers live scattered across files nobody can search fast: service manuals, compliance checklists, pricing sheets, old email threads that never made it into a wiki. The fix isn't a smarter filing system. It's an agent that reads what's already there and answers in plain language, with the source attached so the answer can be checked.
That's the same logic behind the break-even math on AI versus hiring: the cost of a slow answer is time, and time has a price whether it's a customer on hold or an employee waiting on a supervisor. An internal support agent is the same tool pointed at a company's own staff instead of its callers. Level 99 builds both kinds, and they run on the same plans starting at LVL 10.
The full case study has the complete build details and the rest of the math.
FAQ
Did lookup time actually drop, or is that an estimate?
It's measured against the company's own prior process: 2 to 10 minutes of folder digging or waiting on a supervisor, versus under 30 seconds through the agent. The percentage (up to 95%) is arithmetic on those two numbers, not a separate claim.
Does this agent replace the supervisors who used to answer questions?
No. It takes the first-line questions that have a documented answer, so supervisors stop getting interrupted for things already written down somewhere. Anything outside the documents still goes to a person.
What happens if a document is wrong or out of date?
The agent only knows what's in the source files, and it names the file for every answer. If a document is wrong, the fix is editing that document, the same as before. The agent re-syncs nightly, so a correction shows up the next morning without anyone touching the chat system itself.
Why not just train staff to search the shared drive better?
Some of the problem was searchability: scanned PDFs that no text search could read until text recognition was applied. The bigger problem was structural. A rep standing in a customer's kitchen doesn't have a laptop open to a shared drive, and a supervisor interrupted twenty times a day isn't a training failure. The agent fixes the access problem, not a skills gap.