Connect ChatGPT (or any AI assistant) to your company data: what it takes
Plenty of companies already pay for an AI assistant. People use it to draft, summarise and think a problem through. Then somebody asks it a question about the business, and it has nothing to work from. It has never seen any of the systems your business runs on. It knows what somebody pastes in, and nothing else.
This guide is for that moment. It covers why pasting exports stops working, what a real connection needs, and when building one yourself makes sense.
Pasting exports works, until it doesn’t
The usual start is an export. Somebody downloads a report, pastes it into the chat and asks a question. For one question about one report, that is fine. It stops working for four reasons.
- It is out of date once it is saved. The answer describes the day of the export, and nobody writes that date down.
- It is one system’s view. “Sales are up, so why is cash down?” needs figures from several systems. Each export was taken on a different day and cut a different way.
- It ignores who is asking. The assistant uses whatever it is given. A salary file pasted in to answer one question now sits in somebody’s chat history.
- It cannot be checked later. When the board asks where $1.2m came from, the trail is a chat transcript and a file on somebody’s laptop.
None of this is the assistant’s fault. It reasons well over what it is handed. The work is in the handing.
Why two systems give two answers
Two systems can both be right and still disagree. One books a sale when the deal is signed. Another records revenue when the invoice goes out. Ask each for March and you may get $412,000 from one and $386,000 from the other. Both are reading their own records correctly.
An assistant handed both figures will pick one, or split the difference, unless something tells it which department owns the word. The same thing happens with headcount: why HR says 41 and payroll says 43 walks through one case.
What connecting actually needs
Read-only access to each system
Every system has its own sign-in and its own way of granting access. A sound connection asks for the narrowest read-only access the vendor offers, and is tested with a real read before anybody relies on it. Read-only matters. An assistant that can write to your records can also get them wrong.
A boundary for each person
The simplest setup puts one set of credentials behind the assistant. Then everybody who asks sees the same thing, the new starter and the finance director alike. A proper connection checks, on every question, what the person asking may see, and leaves the rest out of the answer.
Figures you can trace
Every figure in an answer should point to the record it came from. That pointer has to be checked by software outside the model. A model asked for a source will always produce one, whether or not the record exists. What “traceable” actually means explains the check.
Agreed definitions
When two systems disagree, the answer should show both figures and say why they differ. That needs each term defined once, with the department that owns it, and used the same way for every question.
Somebody to keep it working
Vendors change their APIs, retire permissions and shorten how long a sign-in lasts. A connection that worked last month can start returning nothing without raising an error. Somebody has to watch for that for as long as it runs.
What about the assistant’s own connectors?
Assistants increasingly offer connectors of their own, and for a question about one app they may be enough. Before you rely on them for business figures, check three things. Does each person see only what they are allowed to? When two systems disagree, are the figures reconciled, or just placed side by side? And can you prove, months later, which record a figure came from?
When building it yourself makes sense
Building makes sense when the question lives in one system, a data team already owns that data, and everybody asking may see all of it. It also makes sense when your data already sits in one warehouse with reviewed definitions. A reporting tool with a question box may then be all you need.
It stops making sense when a question needs several systems, a figure two departments count differently, and more than one person’s permissions. That is a team’s ongoing work, for every system and every change a vendor makes. Build or buy sets out what each path involves.
If you would rather not build it, see how Synaptic Intelligence does it: read-only connections to the systems you choose, answers limited to what each person may see, and every figure linked to its record. Try it for seven days. No credit card required.