How we use AI in our products, and where a person decides
How we use AI in our products, and where a person decides
Device voiceIntroduction
Your phone or computer is reading this aloud in its own voice, so it sounds different.
Here's how we use AI, in one line: it does the slow reading, ranking and drafting, and a person makes every decision that matters. Nothing goes into your books, to an investor, or into someone's career without a human click. This post explains what the AI does in each product, what it never does, and why we drew the line there.
Who this is for: founders, finance leads and HR leads who want to know exactly what the "AI" in 1936 AI does before trusting it with their company's work.
Why we use AI at all
Most back-office work isn't hard. It's slow. Reading 400 CVs. Matching 300 bank lines to bills and invoices. Pulling the right paragraph from a data room when an investor asks a question at 11pm.
That's work where software can genuinely help, because it's mostly reading and sorting. Deciding is different. Deciding who to hire, what to book, or what to tell an investor needs judgement and accountability. So we split the work along that line.
What the AI does in each product
| Product | What the AI does | What a person does |
|---|---|---|
| Masthead | Reads every CV and gives it a score with reasons. Runs first screening calls and summarises them | Decides who moves forward, who to meet and who to turn down |
| Margin | Suggests a match for each bank line from your bills, payment runs and receipts, and says how sure it is. Reads bills you attach and drafts the entry | Approves each match and entry before it reaches the books |
| Galley | Drafts answers to investor questions from the documents in your room, showing where the answer came from | Edits and sends the answer, or doesn't |
A few principles sit under all three.
It shows its working
A suggestion is only useful if you can check it quickly. So Masthead shows why a candidate scored the way they did. Margin shows why it thinks a bank line matches a bill: the same amount, the same vendor, a rule you set. Galley shows which document an answer was drawn from.
If you can't see why, you can't trust it. If you can see why, checking takes seconds.
It says when it isn't sure
Margin's bank matches come with how sure it is. A 98% match on an exact amount from a known vendor is a different thing from a 60% guess. When the software isn't confident, it says so, and leaves the line for you rather than guessing quietly.
A person approves anything that matters
This is the rule we won't bend. In Margin, proposed entries reach the books only after a person approves them. In Masthead, the score is a suggestion; you decide who moves on. In Galley, nothing goes to an investor without you sending it.
We learned this the hard way, running our own company on these tools first. You can read about that in what broke when we ran our company on our own software.
What the AI never does
- It never posts to your books on its own. Every entry waits for approval.
- It never sends anything to an investor or a candidate on its own. Drafts stay drafts until a person sends them.
- It never rejects a candidate. It ranks and explains. People decide.
- It never makes up a number. This one matters most, so it gets its own section.
Numbers are worked out in code, not by AI
Every figure in a report, whether it's your P&L, cash in the bank, runway, headcount, attrition or an investor's reading time, is worked out in code from your actual records. The same records always give the same answer, and you can trace each number back to the entries behind it.
AI is good at reading and writing. It isn't a calculator you should trust with your balance sheet. So we don't ask it to be one. Where AI helps with numbers, it's in finding things ("show me bills from this vendor") or explaining them in words. The figure itself always comes from the books.
Where your data goes
AI features need to read your documents to help with them. We treat that seriously, and the detail belongs on our privacy and security pages rather than in a blog post, because the specifics, such as which providers we use and their terms, can change. Please read those pages for the current position.
As a general principle, we send the AI only what a task needs, keep each company's data separate, and log AI suggestions so you can see what was suggested and who approved it.
Why this is slower, and why that's fine
Letting the software act on its own would save a few clicks. It would also mean that, one day, something wrong would reach your books, your investors or a candidate, and you wouldn't know until later.
A click to approve takes a second. A wrong entry in a filed GST return takes weeks to fix. A candidate rejected by a machine without a person looking is a bad way to treat someone. We'd rather be a little slower and completely trustworthy.
How to get the most from it
- Read the reasons, not just the score. The "why" is where the value is.
- Set rules for the routine. In Margin, a rule for recurring lines, such as bank charges, makes future suggestions sharper.
- Approve in batches. Check the high-confidence suggestions quickly, then spend your time on the uncertain ones.
- Tell us when it's wrong. Every correction helps us improve the product for you and for everyone else.
Questions people ask
Does the AI in 1936 AI's products make decisions on its own?
No. It reads, ranks, suggests and drafts. A person approves anything that matters: who moves forward in hiring, which entries reach the books, and what goes to an investor. If the AI isn't sure, it says so and shows its reasons, so the person deciding can check quickly.
Does AI calculate the numbers in my reports?
No. Numbers in reports, such as your P&L, cash position, headcount or reading time, are worked out in code from your actual records. AI can help explain or find things, but it never invents or estimates a figure that appears in a report.
Can AI reject a job candidate in Masthead?
Masthead reads CVs and screening calls and gives each candidate a score with reasons. It suggests who to advance. A person on your team decides who moves on and who doesn't. The score is a starting point for your judgement, not a replacement for it.
Is my company's data used to train AI models?
This depends on our providers' terms and our own policy, which are set out on our privacy and security pages. Please read those pages for the current position rather than relying on a blog post, since provider terms can change over time.