> ## Documentation Index
> Fetch the complete documentation index at: https://control-dev.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.
# Create a budget from scratch with an AI client
> Watch an AI client analyze historical actuals through MCP, pick a seeding method per account, and write a full-year budget into Control.
This example shows a budget built without a workbook: you tell a connected AI client the rules you want, the client
reads your historical actuals through MCP, chooses a seeding method for each account, and writes the resulting monthly
values into a new budget in Control.
The recording has narration, and the figures, entities, and account names in it are illustrative. Nothing in the
video is a target or a benchmark.
## What happens in the recording
1. The user asks a connected AI client to create a next-year P\&L budget with three rules: check historical data
account by account, decide per account whether December run rate or a three-month average is the better guess, and
clean one-offs from the data.
2. The client reads statement actuals through MCP and analyzes each account's history against those rules.
3. It creates the budget in Control and resolves the group accounts it will write to.
4. It writes the monthly values through the Control MCP write tool — a full year, seeded from actuals.
5. It reports its work back: the method chosen for each account, the one-offs it excluded and why, and the seasonal
patterns it deliberately kept rather than flattening — with every decision open for review.
## Try it in your workspace
1. [Connect an AI client with MCP](/product-docs/ai/mcp-server) and confirm the session has write access.
2. Tell the client the baseline rules you want: which months form the baseline, which methods are acceptable, and how
it should treat one-offs. The recording uses "December run rate or three-month average, whichever fits the account"
— yours can be growth assumptions, headcount plans, or anything else it can apply account by account.
3. Ask for the analysis first: the proposed method per account, the items it would exclude as one-offs, and the
patterns it would keep as seasonal — without writing anything.
4. Review the proposal, correct it, and then confirm the write.
5. Open the budget in Control to verify the values, then refine allocations cell by cell or ask follow-up questions
such as "How does this plan compare with this year's actuals?"
## One-offs and seasonality are judgment calls
The most consequential decisions in a seeded budget are not the arithmetic — they are which movements count as
one-offs to exclude and which are seasonal patterns to keep. A repair charge inside a rent account is a one-off; a
holiday-pay spike every June is not, and a budget that flattens it will report false variances all year.
Make the client show its reasoning for both lists before anything is written. Every excluded item should name the
account, the month, the amount, and why it is unusual, so you can veto exclusions that are actually recurring. Where
the history is too lumpy to call — sparse professional-services spend, an account that changed meaning mid-year — the
client should say so and ask, not pick silently.
Example prompts: "Before writing anything, list every one-off you plan to exclude, with account, month, amount, and
reasoning." "Which accounts did you treat as seasonal, and what pattern did you keep?" "Salaries are moving to a
headcount plan in Q3 — seed them from the run rate but flag them as needing manual revision."
A seeded budget is ordinary budget data: any reviewer can inspect, edit, and lock it in Control, and the seed is a
starting point rather than a finished plan.
## Troubleshooting
### A seasonal account came out flat
A flat run-rate line on an account with a real seasonal pattern usually means the client averaged the pattern away.
Ask it to re-read that account's monthly history, describe the pattern it sees, and re-seed the account keeping the
monthly shape. Accounts driven by payroll timing, annual renewals, or seasonal revenue are the usual suspects.
### An excluded one-off was actually recurring
Re-add it explicitly: give the client the account and the months, state that the item recurs, and ask it to restate
the affected months. The reverse also happens — a genuine one-off kept in the baseline inflates every month built on
it. This is why the exclusion list is worth reviewing line by line before the write.
### The written values do not match the stated method
Ask the client to reconcile a sample of accounts: recompute the stated method from the actuals it read and compare
with the values it wrote. A mismatch usually traces to a baseline month it silently skipped or an exclusion applied to
the wrong month. You can also open the budget in Control next to the actuals and check the same accounts directly.
### The budget landed on the wrong accounts
The seed writes to group accounts, so the client must resolve them from your chart rather than guess from labels. Ask
for the account identifier behind each written row, and reject matches made on similar names alone.
## Next steps
The companion workflow when the plan already exists in a workbook.
Compare the seeded plan with actuals on a reconciled basis.
Refine allocations, review, and lock the budget in Control.
Connect Claude Code, Claude Desktop, or another MCP client.