Your Existing System, Safe for AI to Use
AI access to existing systems: scoped, logged, and nothing rewritten. Your team asks the system questions in plain language, or your customers’ AI assistants work with the product, through a separate layer that only reads, strips personal data before any model sees it, and logs every query. A senior engineer maps your data first, in a free audit. Then we build the layer at a fixed price, beside the system you already run.
- 0 names reach the model
- 2× every figure computed
- Every AI query logged
- A free 30-minute call
- A free audit of your data and who can reach it
- A read-only layer beside your system
- 3-month warranty
If This Sounds Familiar
The data is already there, in a system that works. Getting an answer out of it still goes through one person, or through a tool nobody approved.
“Every new number is a request to our developer.”
Reports are a one-time pull, or a project to automate, so most questions never get asked.
“People are already pasting our data into AI chats.”
There’s no safe way in, so people use an unsafe one, with customer data inside.
“Customers ask whether our product works with their AI assistants.”
The system has no controlled, logged way for an outside assistant to read what it should, and nothing else.
“The board wants to know what AI is doing for us.”
The most concrete answer is your own system answering questions, with numbers you can check.
What People Try, and Where It Breaks
What changed: models can now read a company’s data and answer in plain language well enough to rely on, if they get clean, well-described data and nothing they shouldn’t see. The hard part moved from the model to the layer underneath it.
| What people try | Right when | Where it breaks |
|---|---|---|
| Export to spreadsheets and ask a chatbot | A one-off question on data with nothing personal in it. | Copies of customer data end up in tools nobody approved, and no answer can be traced back to its source. |
| Buy a BI tool | The questions are known in advance, and someone owns the dashboards. | Every new question still needs someone who knows the schema, and the tool doesn’t know your business rules. |
| Connect an AI tool straight to the database | A sandbox with synthetic data, to see what’s possible. | The model can read everything, personal data included, and a prompt can’t enforce read-only. |
| Modernize the system first | The system can’t carry the business anyway, and a modernization is already due. | AI access waits behind the whole project, when a separate read-only layer could ship first and show what the data really holds. |
How It Works
We map the data before we connect anything. A free 30-minute call settles what you want to ask and who needs the answers; the audit shows what the data holds and what it means; the build puts a read-only layer beside the system. You can stop after any step.
-
A 30-minute call
Free, with a senior engineer: the questions you want answered, who should be able to ask them, and whether an audit is the right next step. No data access, no preparation.
-
The audit
A senior engineer reads the system and its data: what exists, what each field really means, where two sources disagree, what’s personal, and who may reach what. You get a written report and a plan for the layer. Free, for now; the report is yours either way.
-
The build
A fixed-price project beside your system, with nothing in it rewritten: a regular copy of the data with personal details stripped, typed queries as the only reader, dashboards, an assistant with your company’s context, and reports written in plain language. Every figure is checked against the source before you see it.
-
After release
A 3-month warranty covers bugs. Model spend runs under a monthly cap you set, and every AI query stays on the record. You own the code and the infrastructure; we can stay on for new reports and questions, or hand over to your team.
What You Walk Away With
Dashboards on the full data set, an assistant that knows your business rules, and reports you can expand by asking, from a phone if that’s where you are.
The model reads a copy with names and identifiers stripped out, a check stops the import if one slips through, and nothing it produces is used for decisions about individual people.
Each answer shows the queries behind its figures, and every AI call is logged with what it asked, what came back and what it cost.
Every field the layer reads is described, and the places where the system disagrees with itself are written down as questions, not buried in a number.
What It Looked Like in Practice
AI access to an in-house system
names reach the model · personal data stripped on the way inOperations on mobile since 2012, with little insight beyond a summary view
- names or identifiers in what the model reads 0
- every figure checked against the source 2×
- assistant query on the record, with its sources Every
TripHero
rebuild deliveryGoal of growing from 6 to 30 hotels in two quarters
- production platform, fixed budget <12 weeks
- audit passed SOC 2
- staff time saved 2+ hrs/day
Evolve Edits
infrastructure cost · per monthDecade-old portal moving terabytes a month through one on-premise machine
- infrastructure cost · per month $1,700 → $300 −80%
- downtime incidents per quarter 125 → 12 −90%
- every client uploads at full speed, however many at once No cap
Who Does the Work
Senior engineers who’ve practiced the same written methodology for over a decade. The engineer who runs your audit stays involved in the work that follows, and you talk to the principals, not a sales team.
Safe to Let AI Near Your Data
- We sign an NDA before any access.
- You grant our access, scope it, read-only where the work allows, and can revoke it at any time.
- The AI layer only reads, enforced by the database rather than a prompt, and personal data is stripped before any model sees it.
- Our AI coding tools run only under business terms that exclude training on your code, and only with your approval.
- We carry professional liability and cyber insurance.
- The code, infrastructure and accounts we build or change stay yours.
Questions Buyers Actually Ask
Does the model see our customers’ personal data?
No. Personal data is stripped on the way into the copy the model reads, so it works with tokens, never names or identifiers. A check on every import stops it if one slips through, and nothing the layer produces is used for decisions about individual people.
Can the AI change anything in our system?
No. The import reads your system without writing to it, and the assistant queries the copy through a database role that can only read. That’s enforced by the database itself, not by instructions to the model.
How do we know the numbers are right?
Every number comes from a checked query, not from the model’s arithmetic. Before you see a figure, we compute it twice, once by the product and once by hand against the source, and every answer shows the queries behind it.
Do we have to modernize the system first?
No. The layer sits beside the system you already run and changes nothing in it. What the audit and the build find about your data is often the best map for a modernization later, if you want one.
How much does it cost?
The call and the audit are free, for now. The layer is built as a fixed-price project with a 3-month warranty. Model usage runs under a monthly spend cap you set, and nothing obliges you to hire us after the audit.
Start With a 30-Minute Call — about the data your system already holds
A 30-minute call with a senior engineer about what you want to ask your system and who needs the answers, then the honest next step: the audit, approaches to weigh, or “not a fit,” said out loud.
Notes on Building High-Quality Software
A short founder’s note and a digest of what we’ve published, sent only when there’s something worth the inbox.