Cost, security, limitations, and when AI is the wrong answer. We answer them here the same way we'd answer them across the table. If a plain answer scares you off, we just saved us both a meeting.
Here are our real ranges. The AI Operations Audit is $2,500, fixed: one day on site, a detailed written report, and a one-on-one review meeting. Quick-win pilots run $10,000 to $40,000 and take 6 to 8 weeks. Core systems run $25,000 to $100,000 over one to two quarters. Strategic programs run $150,000 to $300,000+ per year and are only worth doing after quick wins have proven the approach in your business.
What moves the number: the condition of your data, how many systems we connect, and how hard your success metric is to measure. Full breakdown on the services page.
Yes, and anyone who tells you otherwise is hiding them. A deployed agent has ongoing costs: hosting, AI model usage, keeping the regulatory library current, monitoring, and support. We quote these as a flat monthly amount alongside the build price, so you see the full picture before you sign. Standalone regulatory Q&A is a straight subscription. One-time projects like the audit have no recurring fee at all.
AI models charge by the token, which is roughly three-quarters of a word. Every question your team asks, and every answer that comes back, consumes tokens, like a utility meter for AI. For a typical operational deployment the monthly total is modest and predictable, and we'd rather you never think about it: we include model usage as a flat number in the recurring fee instead of passing through a variable bill. If your usage grows enough to matter, that's usually good news, and we'll talk about it before it changes anything.
Fixed, almost always. A defined scope, a success metric agreed in writing, and a price that doesn't move when the work gets hard. We think billing by the hour rewards slowness, and we'd rather be rewarded for finishing.
The exception is short advisory work. If you need a few hours of straight answers before a vendor decision or a board meeting, we're not going to force a project around it. Ask, and we'll keep it simple.
Across the use cases we audit for, quick wins typically return 5 to 15 times the investment within a year. Think $10k to $25k in, $50k to $250k back, because they target measurable leaks: margin erosion, quote turnaround, fees you shouldn't be paying. Larger programs return bigger absolute dollars at lower multiples (3 to 6x) over longer periods.
Two honest caveats. These are ranges from real engagements, not guarantees. And the audit exists precisely to tell you whether your numbers support them before you spend real money.
$2,500 doesn't fully cover our day plus the written report. It's priced so the decision is easy. The "catch" is transparent: some audits turn into pilot projects, and that's how we make money. But the report is yours either way, it names what we don't recommend, and you can hand it to any vendor. Including not us.
More often than the industry admits. AI is the wrong answer when your data doesn't exist or is wrong (AI can't analyze records nobody kept), when the process itself is broken (automating a bad process gets you bad results faster), when the decision is rare and high-stakes enough that a human should own it end to end, or when a $500 spreadsheet fix solves the problem. Every audit report has a section titled "What we did not recommend, and why." It's usually a third of the list.
Usually not before we start. Data preparation is part of most engagements, and "messy but existing" is workable. What we can't fix is data that was never captured. If nobody recorded actual labor hours per job, no AI can tell you why jobs lose money. The audit's first step is a data snapshot for exactly this reason: it tells you what's usable today and what needs to be captured going forward.
No, and we don't sell it that way. Most of the value in industrial AI doesn't come from replacing the operator. It comes from removing the manual search, interpretation, coordination, and review work piled around the operator. The estimator quotes from evidence instead of memory. The new supervisor gets the answer without pulling the veteran off a job. The typical outcome is capacity: the same team handling more work, ramping new hires faster, and losing less knowledge when someone retires. Companies buy Junipix because their problem is too few capable people, not too many.
Three ways, by design. First, our agents are domain-constrained: they answer from your data and authoritative sources, not from an open model's imagination. Second, every material answer carries citations to the regulation section, policy paragraph, or operational record it came from, so it can be checked. Third, low confidence routes to humans: where the system can't confidently answer, it flags the item for review instead of guessing. An answer that can't be traced is a liability, and we build like it.
The usual failure isn't dramatic. It's a demo that impressed everyone and then quietly stopped being used, because it answered questions nobody actually asks, or nobody trusted where the answers came from. We prevent it with a success metric agreed before work starts, deployment inside the workflow people already use, citations on every answer, and a weekly check-in with one empowered person on your side. If a pilot misses its metric, that goes in the closing report in plain numbers.
Of course we did. We sell AI. Building this site by hand would be like a bakery buying its bread at the grocery store. We used our own tools and workflow to draft it, then Tony and Evan argued over every word the old-fashioned way. If something on here reads wrong, blame us, not the machine. It was supervised the whole time, which is exactly how we tell you to run AI in your shop too.
For general questions, you should. It's excellent and nearly free. But ChatGPT hasn't seen your job history, doesn't know your policies, and can't tell you which of your quotes lost money. Worse for compliance work, it doesn't know where you operate. It will cheerfully cite California rules to a Texan, or New York requirements to a company in Florida, with no citation to check. You don't want that anywhere near your operations.
Our agents are constrained to authoritative sources and your own documents, know your jurisdiction, return the exact citation, and log an audit trail. And they're tuned to you and your company specifically, in a way that never teaches the underlying model anything about your business. Ask ChatGPT and a Junipix agent the same PHMSA question and check the citations. That comparison is the whole answer.
A consultant leaves you a slide deck. We leave you working software. Junipix is a product company that implements: the audit and pilots exist to deploy reusable systems we've built and hardened across industrial customers, not to bill hours. That's also why we can quote fixed prices with success metrics. We've built most of it before.
If you have software engineers with time on their hands, genuinely maybe, and the audit report would give them a great roadmap. What usually tips companies our way: the hard part isn't the AI model. It's everything around it. Data pipelines from QuickBase and ERPs, retrieval that estimators actually trust, security isolation, and the regulatory library we've already built. Buying 80% of that and customizing the rest is faster and usually cheaper than a first in-house attempt. Some customers do both: we build the first system, their team learns from it and takes it over.
Run the numbers on three things: margin you can't explain (for most $20–300M operators, 1 to 2 points of leakage is $200k to $6M a year), quotes you didn't send because estimating was backed up, and what happens the month your most senior person retires. Doing nothing is a decision with a price too. It's just not on an invoice.
No. Your data is used to answer your questions, full stop. Here's why ours doesn't: we use commercial AI services under enterprise terms that contractually prohibit training on customer data, your documents and records stay in your isolated environment, and the "tuning" that makes an agent yours lives in configuration and retrieval on our side, never inside the model itself. Nothing learned from your deployment is shared with any other customer, and nothing about your business ends up in anyone's model.
The production environment is fenced off: not openly reachable from the internet, access restricted to an allow-list of approved networks, and the database in a private subnet with no public endpoint. Isolation is enforced at the database layer with row-level security, the same posture used by regulated SaaS in finance and healthcare, so an application bug can't leak data across customers. Everything is encrypted in transit and at rest, every authentication event is audit-logged, and edge protections absorb the standard volumetric attacks before traffic reaches the application.
We also connect to your systems with minimum-privilege, read-only access wherever possible. We deliberately operate without administrative rights on your systems of record, so even a worst case on our side can't corrupt your source data.
We run on major cloud infrastructure (AWS-class) that carries SOC 2, ISO 27001, and the other certifications your IT team will ask about, and our architecture follows the same control patterns those frameworks require: private networking, least-privilege access, encryption, tenant isolation, and audit logging. We're glad to sit down with your IT or security lead and walk through the whole posture.
And if your own company is being pushed toward stronger AI and data controls by customers or insurers, we can help you get there. We've built to these standards and can help you set policies, controls, and vendor requirements that hold up to review.
Yes. Most customers run on a private cloud instance because it's cheaper to operate and easier to keep current. But if your policies, your customers, or your contracts require it, we can deploy on servers at your site. It adds some cost and some ongoing maintenance coordination, and we'll walk you through that trade honestly before you choose.
You own your data. We're the system of analysis, never the system of record. Your source systems remain untouched and authoritative. If we part ways, your data is returned or destroyed per the agreement, and because we never held the master copy, there's no hostage situation.
Audit: one day on site, with the written report and a one-on-one review within two weeks. First pilot: typically 6 to 8 weeks from kickoff to measured result. Larger systems: one to two quarters. If a vendor promises transformation in a week, ask them what happens in week two.
Absolutely. Book a free 30-minute intro call. No deck, no pressure. Bring your ugliest operational problem and we'll tell you on the spot whether it's worth an audit, worth a pilot, or not an AI problem at all.
Three things: access to the relevant systems and data, one internal champion who can answer questions and make small decisions, and a weekly 30-minute check-in. Projects with those three things land on time. Projects without a champion drift, and we've learned to say that out loud before signing.
We review results against the success metric we agreed to at the start, in plain numbers, in writing. If it worked, we propose the logical next step: production rollout, an adjacent use case, or a retainer. If it missed, the report says so and why. Every step has to earn the next one. You're never locked into a program.
Yes. Think of it like hiring an owner's rep on a construction project: we sit on your side of the table while someone else builds. We review proposals before you sign, rescue projects that stalled or blew past budget, and provide ongoing oversight against a written success metric. Details on the services page.
The conflict of interest, stated plainly: we build these systems too, and sometimes our honest advice will be that your vendor won't get there. When that's our call, we show our work, and the plan is yours to take to anyone. We'd rather lose a build than shade the advice. That reputation is worth more than any single project.
The short answer is yes. We've worked with food service clients, including national brands you'd recognize (we don't name customers, so you'll have to trust us or ask on a call). The underlying problems repeat everywhere: inventory that doesn't match reality, quoting and pricing from gut feel, compliance questions answered by whoever's nearby, and know-how concentrated in a few tenured people. Those patterns show up in retail stores, restaurants, coffee shops, distribution, construction, aviation operations, and plenty of industries we haven't met yet.
A lot of what we do is also simpler than it sounds: we extend what your existing systems can do and make them easier to use. If your business runs on operational data and experienced people's judgment, the audit will find where AI pays off. If it won't, we'll tell you that too.