It is Wednesday evening and the finance pack goes to the chair on Friday. The dashboard says September revenue was £412,300. An hour ago your assistant ran its own query and said £431,900. Both answers came back in the same calm paragraph. You have two days and no way to tell which one you would put your name to.
Neither answer said where it came from, what it counted or as of when. Until it does, a number from an assistant is a claim, and a finance pack is a list of claims the board will later be held to.
All figures in this piece are illustrative. No client's data appears in it.
What goes wrong when an assistant meets your data
Anthropic described this failure in June 2026, when it wrote up how its own teams answer analytics questions with Claude. Pointing an assistant at a warehouse, it says, "can create a false sense of precision". Without skills, which are written procedures the assistant reads before it works, its accuracy on Anthropic's own evaluations "didn't exceed 21%". With them the figures were "consistently above 95% in aggregate and regularly around 99% in certain domains".
Those are Anthropic's tests on Anthropic's data. We have not measured the same on yours, and nothing here promises it. What carries over is the method: tell the assistant how to work before it answers, and write that down where the whole team can use it.
Five habits that make an answer checkable
A skill is a plain text file that an assistant reads before it answers. We use three, described below. None is published yet, so here is the short form of the first, the number check: five habits, in the order they happen.
- Check the agreed definition first. Before writing a query, the assistant looks for the company's own definition of the measure: revenue, net revenue retention, active customer. If there is none, it asks which one you mean instead of choosing quietly.
- Say what one row is. It states whether a row is a customer, an invoice or a customer-month, and tests that, because joining tables of different grain is how totals double.
- Apply the standing filters. Test accounts, cancelled contracts and anything else finance has agreed always apply.
- Check its own work. Before it answers, it looks for joins that multiplied rows, a denominator that does not match your question, and a result far from the last figure you gave it.
- Sign the answer. Every answer ends with where the number came from, the date of the latest data in it and who owns it.
To use them today, paste this into the standing instructions of the assistant you already use, whether that is Claude, ChatGPT or Gemini:
Before answering any question about company data, use the company's agreed definition of each measure, and ask me which definition I mean if none exists. State what one row of each table represents. Apply the standing filters finance has agreed. Check your own query for joins that multiply rows, a denominator that does not match my question and results far from the last figure I gave you. If you cannot reach the data, write the query for me to run and wait for the result. Never invent a figure. End every answer with the source, the date of the latest data and the owner.
What runs for which question
We use three skills in the examples that follow, and some questions need none. The number check makes a figure checkable. The chart rules apply one set of rules, so every chart in a pack looks the same. The combined run only sequences those two, so a request for a number and a chart does both in the right order.
| The ask | What runs | Why |
|---|---|---|
| "Why does the dashboard say £412,300 when my query says £431,900?" | The number check | The question is whether a number is right, and there is nothing to chart. It makes the assistant find the agreed definition, test the grain, check its own query and sign the answer. |
| "Show recognised revenue by plan for Monday's deck." | The combined run | The ask needs a checked number and a chart. It runs the number check first, then hands the answer, source line included, to the chart rules. |
| "Here is the bank's CSV of cash by month. Chart it for the chair." | The chart rules alone | The numbers were verified outside the warehouse, so there is nothing for the number check to do. |
| "What did the board decide about the September hiring plan?" | No skill: Ask Poseidon | The answer is in the record, not the warehouse. Ask Poseidon answers from confirmed records and cites them. |
The block above gives you the number check. The other two stay with us for now.
The CFO, on Wednesday evening
Back to the pack. The CFO asks:
"What was recognised revenue for September, and why does the dashboard say £412,300 when my own query says £431,900?"
Skill invoked: the number check. The question is about whether a number is right, so its checks are what is needed, and no chart is asked for.
An assistant working to the five habits does not pick a side. It finds finance's agreed definition, which counts revenue when it is recognised. The query had summed invoices issued in the month, and one of them is a £19,600 annual invoice for work not yet delivered. That accounts for the whole difference. The answer ends like this:
Recognised revenue, September 2026: £412,300. Source: finance warehouse, revenue recognition table, finance's agreed definition. Data to 30 September 2026. Owner: finance. Checked: one row per invoice line, test accounts excluded. The £431,900 figure is reproduced by summing invoices issued, which includes £19,600 not yet recognised.
On Tuesday that was an argument. On Wednesday it is a line she can defend.
Next she asks her assistant to submit the figure to Boardverse as a finance snapshot, which holds one month of cash, burn and revenue, with revenue meaning recognised revenue. Each figure is marked Actual or Forecast. The submission is held as Pending until a different named person confirms it, and the submitting identity can never confirm its own. Only confirmed figures count. A Pending one counts nowhere. The monthly profit and loss goes in as a financial pack, which is variance-linked: a material variance blocks submission until it is tied to a real board action or resolution, and an approved pack is immutable and supersede-only.
Her assistant does the preparing. Someone else does the confirming, and the record shows who. The analysis does not sit in one person's assistant account either, which is the problem we described in When the model moves in-house, the memory has to be yours.
The chair also wants cash by month charted for the pack, and the bank has sent a CSV.
"Here is the bank's CSV of cash by month. Chart it for the chair."
Skill invoked: the chart rules alone. The figures came from the bank, so there is nothing in the warehouse for the number check to verify. The rules pick the chart type from the question, a line for change over time, and apply the same labels and palette as every other chart in the pack. The chart's caption should name the bank's file as the source, so nobody mistakes it for a warehouse figure.
The chair's next question is "against budget?" Budget against actual, released on 21 October, answers it. A Founder uploads the board-approved budget in Settings. There is no typing a budget into the app, and an assistant cannot submit one. A different Founder confirms it as a whole: every month and line becomes official together, or none does. Financial Review then shows the variance between actual and budget for each line, month and year to date. A line is outside tolerance when it reaches both the percentage and the amount a Founder sets in Settings, and the default is 10% with no amount floor.
| Line (illustrative) | Actual | Budget | Variance | Reads as |
|---|---|---|---|---|
| Revenue | £412,300 | £400,000 | +£12,300, +3.1% | Favourable, within tolerance |
| Monthly burn | £96,000 | £85,000 | +£11,000, +12.9% | Adverse, outside tolerance |
Revenue above budget is favourable, and burn above budget is adverse. The same variance appears as one card on the Dashboard, in Ask Poseidon's answers and in the board pack, for anyone whose role may open Financial Review.
The CEO, on Monday morning
The CEO has two questions before the leadership meeting, usually asked as one: are we on track, and what did the board agree last time? They need different tools.
The first is a number and a chart.
"Show recognised revenue by plan for Monday's deck."
Skill invoked: the combined run. The ask needs a checked number and a chart, so it runs the number check first, with the five habits exactly as above. It then hands the answer to the chart rules, and the chart's caption is the source line from the first step, word for word. Two skills run, in that order, and the CEO typed one sentence.
The second is a fact about the record.
"What did the board decide about the September hiring plan, and has anything been recorded since?"
Skill invoked: none. The answer is in Boardverse, not the warehouse, so it goes to Ask Poseidon, the assistant inside the platform. Ask Poseidon reads every module the signed-in person may open, answers from confirmed records only, and shows the source of each answer by module and record. Here the answer points to the resolution it came from. Why the record matters more than the model explains why that citation matters.
The weekly briefing works the same way. It reads the top records in each module and saves the result, so Monday's page shows last week's briefing instead of generating a new one on every visit. A new one is made once a week, on request, or when the connected model changes. A Founder has to connect the workspace's AI model first.
The Decisions module records decisions as they are made. An assistant can propose one through the connection, and it waits as Pending like any other submission.
Decision Memory, released on 21 October, adds a step before a decision is proposed or decided: the board sees earlier decisions on the same question and how they turned out. On the check date the owner answers "Did it work?" in one tap. Overdue checks show on the Dashboard and in the board book, and Ask Poseidon reads decisions with their outcomes.
The board observer, the night before
The pack arrives the evening before the meeting. An observer has no vote and a limited view, and is expected to read it all the same. On page 14 is a line: recognised revenue, September, £412,300.
Skill invoked: none. The observer is not asking an assistant to produce a number. They are reading one that was checked and signed before it reached them, and the signature is what they check.
Three questions are enough for any figure in it. Where did it come from? As of when? Who owns it, and who confirmed it? A figure that arrives with those three answers beside it can be questioned on its substance. A figure that does not can only be questioned on whether it exists.
What an observer sees in Boardverse depends on the role. Ask Poseidon queries only the modules the signed-in person may open, and an observer does not see deals, for example. A figure still marked Pending is a proposal, not a result, and it counts in no total.
If an assistant's answer reaches an observer without a source line, the reply is the first question again. That costs one email, and it is the same email whether a person or a model wrote the answer.
Setting it up this week
- Write down the five or six measures the board sees every quarter. One line each: what counts, what is excluded, who owns it. That page is the agreed definition the first habit looks for.
- Paste the five habits into the standing instructions of the assistant your finance team already uses.
- Ask it ten questions whose answers you already know. Where it differs, find out why before you trust it on a number you do not know.
- If you want it to submit figures to the record, connect it. Boardverse takes Claude, ChatGPT, Gemini CLI or Google Antigravity today, and other assistants through MCP.
- Run the ten questions again whenever a definition changes. Anthropic reports its offline accuracy drifting from about 95% at launch to about 65% over a month, and says that without active maintenance skills are "wrong within weeks".
The five habits do not give your assistant access to your data. Where there is no connection, it writes the query and you run it, which is slower and safer. They have not been measured against your company's questions. And they do not replace the person who confirms the figure, because the confirming step belongs to a different person.
Boardverse is where the answers go once they are right. Check out Boardverse, or start with the five-answer diagnostic.