Skip to content
be-human.ai
← Back to the journal

Tools & integrations

MCP Servers for Business Agents: Which Tools to Expose First

Which company tools to expose through MCP servers first, which read scopes to allow, and which write scopes to keep closed until approvals exist.

Be Human typographic cover with the title: MCP Servers for Business Agents: Which Tools to Expose First
Conceptual illustration created with AI · Be Human

For most companies, the first MCP servers should expose read-only views of the systems where the business already keeps its working record: customer and job status, the order or ticket queue, and the shared policy documents staff use every day. Write access stays closed until a named person has approved each action type, a success criterion can be measured, and every call is logged. This article explains how to choose those first tools and which scopes to leave closed.

Key takeaways

  • Start with one or two read-only servers tied to a decision the business already makes every week.
  • Expose lookups and drafts first. Keep payments, deletions, bulk updates, permission changes and outbound messages out of reach.
  • Scope each tool to one object type, one environment and one named owner.
  • Treat tool descriptions, retrieved documents and returned records as untrusted input, since any of them can contain instructions.
  • Open write access only after evaluation results, an approval step and a rollback path are in place.

What an MCP server does in business terms

The Model Context Protocol is a standard way for an AI agent to discover and call tools that a server makes available. In practice, an MCP server sits in front of one of your systems and answers a short list of named requests, such as "look up the status of this order" or "list open tickets for this account." The agent does not get the system itself. It gets only the operations you chose to publish, with the input and output shapes you defined. The protocol is described in the [Model Context Protocol introduction](https://modelcontextprotocol.io/docs/getting-started/intro), and the business question behind it is covered in What is MCP, and why does it matter for business agents?.

This matters because the server is where your permission decisions become real. Whatever the server does not expose, the agent cannot reach, no matter how capable the model is. That is why the choice of tools is a management decision, not only a technical one.

Which company systems to expose first

The right first systems share four traits: the question is asked often, the answer lives in one place, a named person owns that data, and a wrong answer is annoying but reversible. Customer and job status, order tracking, support ticket history, published price lists and internal policy documents usually meet these tests. Ledgers, payroll, bank accounts, HR files and legal agreements usually do not, at least not in the first phase.

Use this guide to sort candidate systems:

  • Expose first if staff ask the same lookup question several times a day and the answer already sits in one system of record.
  • Expose first if the data has a stable identifier, such as an order number, a ticket ID or a customer reference, that the tool can use as its key.
  • Hold back if the system moves money, creates legal commitments or stores health, employment or identity documents.
  • Hold back if nobody can say who approves changes to the data or who answers when it is wrong.
  • Hold back if the only way to find a record is by free-text search across fields that were never designed for lookup.

For a deeper check on whether a system is ready for any agent access at all, see Does Your Software Have an API? A Quick Check Before AI.

Which read scopes to allow, and how to narrow them

A read scope is not all or nothing. Narrow it along four lines: object type, fields, time range and team. An order status tool, for example, should return the order state, the promised date and the tracking reference. It should not return the customer's payment details, internal margin notes or the full history of every edit made by staff. Build the field list as an allowlist, so that new fields added to the source system stay hidden until someone reviews them.

Also cap results. A tool that returns every matching record can leak more than intended and can overwhelm the model with data it does not need. Return a page of results, a clear count of what was omitted, and an instruction to narrow the query. The [OWASP Authorization Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) describes the general principle: check access on every request, not only when the connection is first set up.

Read scopes deserve the same review cycle as write scopes. The data a server can read is often more sensitive than the actions it can take, and it accumulates risk over time as more fields get added for convenience.

Which write scopes to leave closed at the start

In the first phase, keep closed any tool that changes a record that someone else relies on without review. That includes deleting records, issuing refunds or payments, posting accounting entries, changing prices, editing user access, sending bulk messages to customers and confirming bookings in a shared calendar. Each of these can be undone only with effort, and some cannot be undone at all.

A useful middle ground is the draft tool. The agent prepares the change, such as a reply, a proposed reschedule or a draft invoice, and writes it to a queue where a named person approves or rejects it. The distinction between reading, drafting and sending is explored further in Read, draft and send: three different AI permissions.

A draft tool gives the agent a way to be useful without giving it the last word.
Illustrative scenario: a field services company wants an agent to answer questions about appointments. The first server exposes read-only job status and technician availability. A draft tool writes proposed reschedules to a review queue. Confirming a reschedule in the shared calendar stays closed until reviewed drafts have consistently respected the constraints the dispatcher cares about, such as travel time and crew skills.

Designing tools the model can use safely

Each tool needs a precise name, a plain description of when to use it and a strict input schema. Vague tools produce vague calls. A tool called "search" that accepts any text is harder to govern than "find_order_by_number" with one required identifier. The same applies to outputs: return consistent fields and explicit status values, so that a wrong call is visible rather than silently plausible. For guidance on writing these contracts, see AI tool contracts: design inputs and outputs that work.

Treat everything the tools return as data, not as instructions. A customer email, a supplier note or a document in a shared folder can contain text that tries to redirect the agent, such as a request to ignore its rules or to export other records. The [OWASP prompt injection prevention guidance](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) recommends separating trusted instructions from untrusted content and never letting retrieved text change tool permissions. In practice, the server should enforce scope itself, so that a persuasive paragraph cannot widen what the agent is allowed to do.

Identity, logging and a rollout you can run

Give each MCP server its own identity, using a service account or a narrowly scoped OAuth grant, and keep the grant limited to the read or draft scopes you approved. Avoid sharing one administrator credential across several servers. Record every call with the time, the tool name, the arguments, the record identifiers returned and the person or process that triggered it. This log is what lets you answer, later, what the agent saw and why it acted.

Roll out in stages rather than all at once:

  • Name one business owner for each system and the exact scopes that owner approves.
  • Build the read-only server, test it against records with known answers, and confirm that the field allowlist holds.
  • Run the agent in parallel with the existing process for a defined trial period and compare its answers with those of staff.
  • Add a draft tool only after read results are reliable, and route every draft to a named approver.
  • Review scopes on a fixed schedule and after any change to the source system, since vendors and internal teams both change what data exists.

Reviewing scopes after launch

Scopes drift. Someone adds a field to be helpful, a new report needs a broader query, or a draft tool gets reused for sending. A monthly or quarterly review of the tool list, the scope of each tool and the call logs catches this drift early. Keep the review short: which tools were called, which were never used, which returned errors, and whether any scope now exceeds what its owner approved.

In Be Human's managed model, the first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. Operations then cover monitoring, recurring evaluations, integration maintenance, permission review and incident response, and any new write scope is added as a recorded change only after evaluations pass and the owner signs off. The agent keeps its identity, memory and channels under your control throughout.

If you are not yet sure which of your systems should sit behind the first MCP server, the free Be Human AI diagnostic is a practical way to find the first task worth handing to an agent.

Frequently asked questions

Is an MCP server the same as an API? An MCP server is a standard way for an agent to discover and call tools, and it usually sits in front of one or more existing APIs or databases. The API still enforces its own rules, but the MCP server decides which operations the agent can request and how results are shaped. That extra layer is where most business scoping decisions belong.

Should the first MCP server be read-only? In most cases, yes. A read-only server lets you test accuracy, catch field exposure problems and build confidence with staff before any action changes a record. Draft tools can follow once read results are reliable, and direct write access should wait for evaluations and a clear approval boundary.

Can one MCP server serve several agents? It can, but that is usually a mistake in the early phase. Each agent should have the scopes its role requires, and sharing one server across agents with different responsibilities makes it harder to see which agent made which call. Separate identities and logs keep accountability clear.

What happens when a tool returns wrong or incomplete data? The agent should report the gap rather than fill it in, and the call should appear in the log with enough detail to trace the cause. Good tool design returns explicit status values and missing-field notices, so that the failure is visible. Repeated errors on the same tool are a signal to review the source system or the field mapping, not to widen the agent's permissions.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →