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

Tools & integrations

Model Context Protocol for Business: Which Tools to Expose First

MCP lets an AI agent use your company tools through one standard interface. Learn what exposing a tool means and which low-risk tools to expose first.

Be Human typographic cover with the title: Model Context Protocol for Business: Which Tools to Expose First
Conceptual illustration created with AI · Be Human

The Model Context Protocol (MCP) is an open standard that lets an AI agent discover and call tools offered by a server through one common format, instead of a custom connector for every pairing of model and system. For a business, exposing a tool means publishing one narrow, described action, such as looking up an order status or drafting a reply, under rules that you define. The first tools to expose should be read-only, tightly scoped and easy to check, because a mistake there is cheap to catch and cheap to undo.

Key takeaways

  • Exposing a tool through MCP means describing one action, its inputs, its outputs and its permissions, so an agent can call it without direct access to the underlying system.
  • Start with read-only tools that answer questions from a system of record, because a wrong answer costs far less than a wrong write.
  • Keep the protocol layer separate from approval rules, so a permission change does not require rebuilding the agent.
  • Choose each tool by checking reversibility, data sensitivity and how quickly a person would notice an error.
  • Define one observable success criterion per tool before connecting it, and review the tool list on a fixed schedule.

What exposing a company tool through MCP actually means

An MCP setup has three roles: a host (the application where the agent runs), a client that the host uses to talk to a server, and a server that offers capabilities. The server lists its tools. Each tool has a name, a plain description and an input schema that states what the agent may send. When the agent decides a tool is relevant, it calls it, and the server returns a result.

Exposing a tool therefore means writing a contract, not opening a door. Instead of giving an agent a login to your accounting software, you offer a function such as listing unpaid invoices for one customer, returning the invoice number, due date and total, and capping the result at a fixed number of rows. The server validates the input and returns only the fields the job needs. Anything outside that contract is out of reach by design.

One point is easy to miss. The protocol carries the request, but it does not decide who may do what. Authorization belongs to the server and to the systems behind it. The OWASP authorization guidance is clear that access should be checked on every request, not assumed because a request arrived from a familiar client. A standard connection makes integration easier, but it does not make a broad permission safe.

Why a standard protocol changes the work, and what it does not change

Before a shared standard, each agent needed its own connector to each system, and each connector had its own format, login flow and failure modes. A standard lets one tool server serve more than one client and lets a team replace the model or the host without rewriting every integration. That is the real gain: fewer bespoke links to maintain.

What it does not change is the risk. Every tool you expose adds a path from language to action, and every returned field adds something the model can repeat or act on. The protocol makes it easier to add paths, so the discipline has to come from how you choose and bound them. Teams that plan to connect everything and restrict access later usually find that the restriction never arrives in clean form.

Which tools should you expose first?

Ask four questions about each candidate system or action:

  • Does it only read, or does it write, send, delete or change access?
  • If the agent gets it wrong, would a person notice before a customer, supplier or ledger is affected?
  • Can the result be reversed quickly, with a procedure your team already knows?
  • Does the data include personal, financial or contractual details that need separate handling?

Then sort the candidates into four groups. The first group is read-only lookups against one system of record, such as order status, ticket history or a document index. The second group is drafts that a person reviews and sends, such as a reply to a supplier or a draft purchase request. The third group is bounded writes that are narrow, reversible and safe to retry, such as creating a task in one named project. The fourth group covers anything that moves money, signs documents, deletes records, exports data in bulk or changes permissions. Leave that fourth group out of the first engagement.

For a fuller treatment of the difference between these permission levels, see Read, draft and send: three different AI permissions.

A first set of three tools, step by step (illustrative scenario)

Illustrative scenario: a regional equipment distributor, used here only as a generic example, wants its sales coordinators to ask an agent about order status and delivery dates without opening the order system. The team wants answers taken from the system of record, not from memory or from old email threads.

A workable sequence looks like this:

  • Write the job in one sentence, for example: answer order status questions for a named customer using only the order system.
  • Define the success criterion you can observe, such as answering correctly against a fixed list of test questions the team wrote before any connection was made.
  • List the exact fields the answer needs. If the agent does not need the payment method, the tool does not return it.
  • Write the tool contract: a clear name, a description of when to use it, typed inputs, a maximum result size, and an error message that tells the agent what to do next, such as asking for the order number.
  • Create a dedicated service identity with read access to that one object type, separate from any employee account.
  • Test with questions that have known answers, plus requests the agent must refuse, such as a change to a delivery address.
  • Connect the tool in a test channel and log every call, including inputs, outputs and errors.
  • Review the log after a review period you set in advance, fix any contract that returned too much or failed silently, and only then decide whether to add the second tool.

The permissions behind this sequence can be recorded in a simple grid, as described in AI Agent Tool Permissions Matrix: A Least-Privilege Template.

What to keep out of the first set, and why untrusted text matters

Keep out any tool that can move money, sign documents, delete records, send messages to customers without review, export large datasets or alter user access. Those belong in later stages, behind explicit approval steps that name a person.

There is a second reason for caution. An agent reads text from emails, documents, tickets and web pages, and some of that text can contain instructions meant to steer it. The OWASP guidance on prompt injection treats this as a core risk for any system that lets a model act on external content. A sound design keeps privileges small, treats retrieved content as data rather than commands, and requires confirmation before any action that is hard to reverse. A read-only tool with a narrow return shape limits the damage even when a manipulated instruction gets through.

How to write tool contracts an agent can use correctly

A tool contract is partly documentation for the model. The name and description determine when the agent chooses a tool, so vague names such as get data lead to wrong calls. Use specific names, state what the tool does not do, and give each parameter an explicit format. Return compact, labeled fields rather than raw records, since every extra field is more text the model must weigh and may repeat. Anthropic's engineering guidance on writing tools for agents makes a similar point about descriptions and clear outputs.

Errors matter as much as results. A message such as not found leaves the agent guessing, while a message that says the customer identifier must be the account number, not the company name, lets it recover. For any write tool, require an idempotency key so that a retried request cannot create the same task or entry twice. For a deeper look at this problem, read Idempotency for AI agents: prevent duplicate actions.

How Be Human handles this in managed operations

Be Human's first engagement defines one useful role, the memory and tools it needs, the approval boundaries and one observable success criterion, so the first tool set follows the job rather than the catalog of available integrations. Managed operations then keep that set honest. They include monitoring of tool calls, recurring evaluations against the test questions, permission review, maintenance when a source system changes its interface, and incident response when a tool returns something it should not. The aim is that each tool stays narrow, observable and easy to revoke, with a named person responsible for every decision that touches money, customers or access.

Keeping the tool list safe over time

A tool list drifts. Someone asks for one more field, a vendor changes an endpoint, and a read-only tool quietly gains a write path. A short routine prevents most of that.

  • Keep one written register of every exposed tool, with its owner, permissions, returned fields and success criterion.
  • Review permissions on a fixed schedule and after any change to the source system.
  • Read the call logs for refusals and errors, not only for successes.
  • Re-run the fixed test questions whenever a contract, model or prompt changes.
  • Pause or remove any tool whose owner cannot say clearly what it is for.

If you are not sure which of your systems holds the first task worth handing to an agent, the free Be Human AI diagnostic is a practical place to start.

Frequently asked questions

What is the Model Context Protocol in plain business terms? It is a standard way for an AI application to find and use tools that another system offers, using a shared format for names, descriptions and inputs. For a business, this means one tool server can serve several agents without a separate custom connector each time. It does not decide permissions, so access rules still belong to the server and to the systems behind it.

Which tool should a company expose first? Choose a read-only lookup against one system of record that people already check by hand, such as order status or a document index. A wrong answer there is visible and easy to correct, and the result shows whether your contracts, permissions and logs behave as expected. Drafts come next, and writes come only after that.

Is MCP safe to connect to email, payments or customer records? It is safe only when each exposed tool is narrow, uses its own limited identity and returns the fewest fields the job needs. Payments, signatures, deletions and bulk exports should stay outside the first set and sit behind explicit human approval when they arrive. Assume that text in emails and documents may try to influence the agent, and design so that a single manipulated call cannot do serious harm.

How long does it take to define a first set of tools? There is no fixed duration, because it depends on how many systems are involved, how clean their data is and how quickly your team agrees on the success criterion. The work moves faster when the job is written in one sentence, the required fields are listed and a test set with known answers exists before any connection is made. Be Human scopes these elements in the first engagement so that the tools follow the role.

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 →