Data sources
A management report is usually built from three places: deals and the sales funnel in the CRM; revenue, stock and debts in 1C, the accounting system most Russian companies use; money on accounts at the bank. The agent takes the numbers from each source itself, with no exports to spreadsheets.
Sources are connected through the MCP protocol. Anthropic opened it on 25 November 2024 as one standard way to connect AI to data and services. Each system is wrapped as an MCP server, and the agent sees it as a set of actions, such as “revenue for yesterday” or “debts older than a month”.
The bank is connected through its API, that is, a programming interface for data exchange, or the agent uses the statements that already go into 1C. The second way is simpler: fewer access rights and one point of reconciliation.
Five metrics are enough for the first report: revenue, money on accounts, customer debts, new deals and stalled deals. Add the rest when the manager starts asking questions the report does not answer.
The main work here is not in the model but in the data. If deals in the CRM are closed a week late, the report will show a sales drop that does not exist. Before building, find out who enters numbers into each system and when.
Report and deviations
This agent does not need RAG, that is, search through a document base: the answer is in the accounting numbers, not in texts. Instructions and tools are enough; read more on the page about an agent without RAG.
In its guide to agents, Anthropic advises starting with a simple solution and adding complexity only when needed. For a report this means: a program does the calculations by set formulas, and the model writes the text and explains the deviations.
The model does not calculate by itself. A language model writes text well but can make arithmetic mistakes, so the agent takes sums, percentages and the comparison with the plan from a ready calculation.
The report format is set in advance: metric, actual for yesterday, plan, deviation in percent and one line on the cause. A sample line: “Revenue is 12% below plan: two large shipments moved to Monday”.
The manager sets the deviation threshold, and the agent writes only about what went past it. A report without a threshold turns into a list of all numbers, and people stop reading it.
The report goes where the manager will read it: Telegram, email or a dashboard. The agent answers a follow-up question, such as “why did the margin fall at the Lenin Street store”, from the same sources.
After a month, look at which lines the manager asked about and acted on. Lines nobody reads are removed from the report.
Read-only access
A reporting agent does not need to change anything in the books, so it gets read-only rights: a separate 1C user without the right to post documents, a CRM key without write access, bank statements without access to payments.
The MCP specification requires the same order. A tool in it means code execution, and the application must get the user’s explicit consent before calling a tool. The protocol itself does not enforce these rules: whoever connects the agent does.
Read-only rights remove the main risk. A model error will spoil the text of the report, but not an entry in 1C. The request log shows which data the agent took and when.
Access keys are stored in the builder platform or on the server, not in the text of the instructions. Rights are reviewed once a quarter: if the agent got a new source, check that write access did not come with it.
The CRM holds clients’ names and phone numbers, and that is personal data. Under 152-FZ, Russia’s personal data law, data of Russian citizens is stored in databases in Russia, and before any transfer abroad the operator notifies Roskomnadzor, the data protection regulator. So the agent gets summary numbers, not client cards; our storage rules are on the page about data security.
Build it or order it
A summary from one spreadsheet or one CRM can be built yourself in a builder platform: a scheduled script takes the numbers, the model writes the text, a bot sends it to Telegram. This is the “do it yourself in a week” level.
A report across three systems with a plan and deviations needs a developer. You need to connect 1C and the bank, grant rights and check that the agent’s numbers match the accounting books. How the agent works with 1C is shown on the page on integration with 1C.
Do not skip reconciliation with the accounting books. For the first two weeks the agent’s report is compared with the manual one, and only after they match is the manual report dropped.
We build such an agent as a pilot: from $1,499, 14 days, with a metric in the contract, such as “a report by 9 a.m. every day, numbers match the accounting books”. The ready service with a dashboard and questions to the data is described on the page of the manager’s dashboard.
After launch, plans, accounts and CRM fields change, and the report needs updates. So count the support cost from the start: ours costs from $480 a month, and you can cancel in any month.