CentSight vs. general AI
Your team can already use AI. That’s not the same as your firm having one.
General assistants are genuinely capable, and plenty of good firms use them. The gap shows up the moment the work spans a roster of clients and a partner has to sign the output.
The same question
“Why did this client’s net income drop last month?”
A general AI assistant
No ledger access“Net income could be down for several reasons — higher cost of goods sold, one-off expenses, seasonality, or a timing difference between revenue recognition and collections. Upload the P&L and I can take a closer look.”
- Correct, and not useful — it lists what could be true
- Needs an export that is stale the moment you make it
- Next client means starting the context over
CentSight
Connected to the books“Net income fell $4,120, but the business didn’t get worse. A $4,495 annual software renewal was expensed in full in June; only $375 of it belongs to the month. Spread correctly, net margin is 11.4%, not 8.1%.”
- Figures computed against the ledger, not written by a model
- Every number traces to the transactions behind it
- Runs the same way across every client on your roster
Messy data is where the difference shows. A renewal expensed in the wrong period doesn’t surface in a summary — it surfaces when something reads every line and knows what an accrual is.
Credit where it's due
Where a general assistant is genuinely the right tool.
If your team is getting real value out of ChatGPT or Claude, it isn't because they're being fooled. There is work these tools are simply good at, and none of it changes because CentSight is in the building.
Drafting and rewriting
Turning rough notes into a client-ready email, a memo or the narrative around a board pack. It is faster than most people are, and the output is yours to edit.
Thinking a problem through
Pressure-testing an approach, listing what could explain a variance, working out how to structure an engagement. Judgment work, where directionally useful is enough.
Explaining and translating
Talking a founder through what a covenant actually means, or turning accounting language into something a non-finance client follows.
The one-off technical job
The spreadsheet formula, the regex, the script you'd otherwise spend an afternoon remembering how to write.
The line is narrower than “general AI is bad at finance.” It is what happens when the answer has to be a number, about a named client, that someone at the firm puts their name to.
Head to head
Where it breaks down, line by line.
Your team is probably already using ChatGPT or Claude on client work. This is exactly where that breaks down — and where CentSight was built not to.
Math and calculations
General AI
Generated as text by the model — fluent, and sometimes wrong.
CentSight
Computed in code against the ledger, every time. Never model-generated.
Being talked into a different answer
General AI
Can be re-prompted or pushed into changing an answer that was right.
CentSight
Numbers are computed, not negotiated. Rephrasing the question doesn't move the figure.
Multi-client separation
General AI
One chat window, one memory. Keeping clients apart is on the person typing.
CentSight
Every client company gets its own workspace: separate data, separate connections, separate history.
Live financial data
General AI
None natively. It needs an export, which is stale the moment it's made.
CentSight
A direct read-only connection to QuickBooks and bank data, current at the moment you ask.
Credentials and data handling
General AI
The firm ends up pasting in exports, or handing over account access to get real data in.
CentSight
CentSight holds the connections itself. Credentials never pass to a model provider.
Traceability
General AI
No path back to a source transaction. You get an answer, not a receipt.
CentSight
Every figure carries the account, period and entries it was computed from — something you can hand a client.
Proactive monitoring
General AI
Passive. It answers when asked, about whatever was uploaded that day.
CentSight
Runs scheduled reviews against connected books and flags anomalies and miscategorized transactions on its own.
Client-facing readiness
General AI
Not built for controlled client exposure. It's all or nothing.
CentSight
A ladder the firm controls: internal only today, client-facing if and when you decide.
| Dimension | A general AI assistant | CentSight |
|---|---|---|
| Math and calculations | Generated as text by the model — fluent, and sometimes wrong. | Computed in code against the ledger, every time. Never model-generated. |
| Being talked into a different answer | Can be re-prompted or pushed into changing an answer that was right. | Numbers are computed, not negotiated. Rephrasing the question doesn't move the figure. |
| Multi-client separation | One chat window, one memory. Keeping clients apart is on the person typing. | Every client company gets its own workspace: separate data, separate connections, separate history. |
| Live financial data | None natively. It needs an export, which is stale the moment it's made. | A direct read-only connection to QuickBooks and bank data, current at the moment you ask. |
| Credentials and data handling | The firm ends up pasting in exports, or handing over account access to get real data in. | CentSight holds the connections itself. Credentials never pass to a model provider. |
| Traceability | No path back to a source transaction. You get an answer, not a receipt. | Every figure carries the account, period and entries it was computed from — something you can hand a client. |
| Proactive monitoring | Passive. It answers when asked, about whatever was uploaded that day. | Runs scheduled reviews against connected books and flags anomalies and miscategorized transactions on its own. |
| Client-facing readiness | Not built for controlled client exposure. It's all or nothing. | A ladder the firm controls: internal only today, client-facing if and when you decide. |
The first row, shown
The math is computed, not written.
Ask why a client’s net income moved and the answer comes back as $4,120 because a query ran against the ledger — not because a model produced a figure that reads like one.
Here the drop traces to one entry: a $4,495 annual software renewal was expensed in full in June instead of spread across the year, so June carries $4,120 of cost that belongs to later months. Corrected margin is 11.4%, and the entries behind it are listed under the answer.
The sentence around the number is written by a model. The number is not.
Why did net income drop last month?
Net income fell $4,120 against last month, but the business didn’t get worse. An annual software renewal was expensed in full in June; only one month of it belongs there.
- Annual software renewal
- $4,495
- Belongs to June
- $375
- Prepaid, expensed early
- $4,120
- Net margin, corrected
- 8.1% → 11.4%
Source: 1 bill entry · Jun 4 · QuickBooks Online
The second row
It can't be argued into a different number.
A general assistant is agreeable by design. Push back on a figure, rephrase the question, tell it the number looks wrong, and it will often produce a different one — with the same confidence it had the first time.
CentSight can’t do that. The figure came out of the ledger, so the only thing that moves it is the ledger moving. If a client disputes the number, the disagreement is about the books, which is where it belongs.
The third row
A chat window has one memory. Your roster doesn't.
Multi-client work gets tangled in a general assistant for a structural reason, not a careless one: there is a single context, and a person has to hold the line between clients in their head. Here each client company is its own workspace, with its own connections, history and signals.
HW E-Commerce
Connected to QuickBooks, Chase, Stripe
Marketing Agency
Connected to QuickBooks, BofA, HubSpot
Pulse HQ
Connected to QuickBooks, Mercury, Stripe
Restaurant Group
Connected to QuickBooks, Wells Fargo, Slack
Separate data, separate history, separate connections. Nothing crosses between columns.
Connections
Credentials never reach a model provider.
Getting live data into a general assistant means someone exports it, or hands over account access to something that can fetch it. CentSight is the broker instead: it holds the read-only QuickBooks and bank connections itself, and keeps them working. Your clients' credentials stay on this side of the line.
- More on the way
Client-facing readiness
All or nothing is the wrong shape for client work.
Putting a general assistant in front of a client means handing over everything at once, or keeping it hidden entirely. Adoption here is a ladder the firm controls, and the first rung — internal use only — is a complete place to stop.
- 01
Where every firm starts
Internal only
Advisors use CentSight as their own tool — asking, reviewing, and deciding what's ready before anything reaches a client.
- 02
Once it has earned it
Read-only client access
Clients see the work an advisor creates in the tool — reports, analysis, answers — without touching it directly themselves.
- 03
If and when you want it
Full client access
Clients get their own access: a 24/7 financial analyst working for them, inside the workspace your firm still controls.
The firm decides which rung it is on, and when that changes. Nothing moves to a client without you putting it there.
The bigger question isn’t which AI — it’s whether you build it yourself.
Run your own side-by-side.
Bring the prompt you’d normally paste into a general assistant. We’ll run it against connected books and talk through what it looks like to embed CentSight into your firm.