Why another platform? Why not just LLMs?

Every demo, someone asks it. Sometimes politely, sometimes with the raised eyebrow of a CFO who just watched ChatGPT build a three-statement model in 40 seconds: why do we need another platform? Why not just use an LLM?

Fair question. An LLM is the most impressive analyst you've ever hired. It's also the only one who might give you a different answer every time you ask, can't tell you where a number came from, and forgets everything the moment your colleague opens a new chat.

We use LLMs all day, and Valzo runs on them. But three problems don't go away with better prompts: finance is a team sport, business logic drifts, and finance needs receipts.

You prompt. It answers. You check it, fix two numbers, prompt again. Three rounds later you have something you trust, and you email it to whoever asked. That loop works beautifully. For one person.

Now put three people on the same model. Their logins don't share your skills, your markdown files or the instructions you spent a week tuning. Same question, three answers, and no idea which one reached the board.

B2B work is multiplayer by nature. That's true of a single invoice moving through approvals, and painfully true of a strategic call. Take a pricing decision:

• Sales:“We're not competitive. Cut the price.”
• Marketing:“Sure, and watch ROAS fall off a cliff.”
• CEO: has already done the mental math on revenue and EBITDA for the board. Mostly on the drive in. Mostly optimistically.
• CFO:the one who actually has to run the numbers, and make everyone understand the discount isn't free.

That's two or three rounds of back and forth with four people's inputs. In a chat window, it's four private conversations and a spreadsheet called FINAL_v7_actually_final.xlsx.

Then comes the real test. Two months later, someone asks why you did it. Good luck rebuilding the reasoning from four chat histories, one of which belongs to a sales VP who has since moved on.

Finance decisions need a shared model, shared assumptions and a record of who changed what, and why. A chat window gives you none of the three.

“But LLMs have shared projects now.” They do, and they're great. A shared project shares instructions. It doesn't share a governed model of your business: one definition of contribution margin, one set of allocation rules, one approved version of the forecast, and a log of who moved which assumption. Sharing the prompt isn't the same as sharing the numbers.

The counterargument: give the LLM skills, context files and a big enough context window, and it figures it out. Honestly? Mostly, yes. Bet on the trend and models only get better at this.

But a skill is still a suggestion. The model still decides what to use and how to use it. And in finance, 90% right is 10% wrong, which is a polite way of saying wrong. You need deterministic code for the math, and LLM reasoning for everything around it.

Picture a $15M coffee company selling through four channels:

1. 1Marketplaces and delivery apps: Amazon, Uber Eats
2. 2Its own website
3. 3Retailers: Walmart, Tesco
4. 4Its own cafés and affiliate cafés

Some costs belong cleanly to a channel. Others, like the roastery, the warehouse and the finance team, have to be allocated. Fine. Write the skill, set the rules, done.

Then month 3 arrives with a new channel to evaluate. Or a new delivery method. Or a new AI tool tuning ads and customer profiles on just two of the channels.

• Who updates the skill?Someone has to notice it's out of date first.
• Who catches the drift?Across 100 ledger accounts and five data sources (Amazon orders, payroll, the ERP, marketing spend, funnel data), a model that quietly changes its mind looks exactly like a model that's right.
• Who reviews it?Add a human reviewer and you've traded slow output for fast output plus slow review. Skip the review and you're betting the management deck on the model having a good day.

And “we'll just build a semantic model” holds up until you launch a new product, or your AI marketing expands to a third channel. Then someone rebuilds it. Usually at 11 pm, the night before the board meeting.

A marketing team can live with an answer that's roughly right. Your auditor, your lender and your board can't. Finance numbers get challenged, and when they do, “the model said so” is not a defense.

• Same question, different answer. Ask an LLM to recompute Q2 gross margin next month and you may get a different number, with equal confidence. Finance needs the same inputs to give the same output, every time.
• No lineage. A good finance number traces back through the formula, to the GL account, to the invoice. A chat answer traces back to a chat answer.
• The data doesn't fit.A $15M company's ledger holds hundreds of thousands of transactions across years and systems. You can't paste that into a prompt, and “I summarized it first” is exactly where errors go to hide.
• Access matters.Payroll, customer contracts and margins shouldn't live in whoever's chat history uploaded them last. Finance needs permissions, not attachments.

None of this is a knock on the models. A calculator doesn't need to be creative, and a poet doesn't need an audit trail. Finance needs both, in the right places.

Put the problems together and you get the case for a platform. Not instead of AI. Around it.

• AI does the heavy lifting:reading the ledgers, mapping new accounts, proposing how a new channel's costs should be allocated, drafting the analysis. The fast 90%.
• One shared model fixes the multiplayer problem: sales, marketing, the CEO and the CFO work on the same numbers and the same assumptions, and the reasoning is still there two months later.
• A harness fixes the drift problem: add a channel or a product and the AI proposes the new mappings, a human approves them, and the rules update for everyone at once. Checks catch drift, not a tired analyst at 11 pm.
• A deterministic engine fixes the receipts problem: same inputs, same answer, every time. Every number traces to the transaction behind it, and only the right people can see it.

LLMs get you to 90% fast. A finance harness gets you the last 10%, which in finance is the only part anyone remembers.

That's why we built Valzo.