Key Takeaways

• Nobody needs 3,200 integrations. You need the handful behind your recurring work, usually under ten. The catalog size tells you the ceiling; your weekly workflows tell you the actual list.
• Choose by workflow, not by tool. "Connect HubSpot" is a tool decision. "The Monday pipeline report needs HubSpot and Stripe" is a workflow decision, and it hands you the exact connection list with a reason attached.
• Every integration is a capability and a responsibility. Read access answers questions; write access takes actions. Connect read-only workflows first, add write access where a human approval gate makes it safe.
• Audit before you expand. After the first month, the question is not "what else can we connect" but "which recurring work still involves a human copying data between tabs". That is where the next integration pays for itself.
• A missing integration is a path, not a wall. When your tool is not in the catalog, there is a request-and-bridge route rather than a dead end.

The wrong question at the start

A team signs up, opens the integration catalog, sees a few thousand logos, and asks the natural question: "which of these should we connect?" It is the wrong question, and it produces the classic failure mode: either analysis paralysis, or a dozen connections made "just in case" that nobody has a plan for.

The right question is one level up: "which of our recurring workflows should the AI employee own?" Answer that, and the integration list writes itself. This post is the method: how to derive your list from your workflows, in what order to connect, and what to do when a tool you rely on is not in the catalog.

If you are on day one and just want a safe starting set, read choosing your first 3 integrations first. This post picks up where it ends: turning the first wins into a deliberate map of your whole stack.

How do you derive the list from your workflows?

Write down your team's five most annoying recurring tasks, then list the tools each one touches. The union of those lists, minus what you already connected, is your integration roadmap in priority order.

A worked example for a 20-person B2B team:

Five workflows, eight integrations, and every single connection has a sentence explaining why it exists. Compare that to "we connected twelve tools during onboarding" with no sentence attached.

Two useful side effects of doing it this way:

• You discover which tools are load-bearing. Stripe appears in three of the five rows above. Connecting it unlocks multiple workflows at once, so it goes first.
• You get a natural test for each connection. The workflow is the test. If the Monday report arrives correct, HubSpot and Stripe are configured right. No abstract "is it working" question.

What order should you connect in?

The order is simple: read-only workflows first, write actions second, and each write action behind an approval. This sequencing is about trust-building, both yours in the AI employee and your security reviewer's in the rollout.

Stage one: read and report

Start with the workflows where the AI employee only pulls and summarizes: reports, digests, weekly reviews. Wrong output is visible and costs nothing; you just correct it in the thread. This is where most teams get their first "it just did my Monday morning" moment.

Stage two: write with approval

Then add the workflows that end in an action: drafting the reminder email, updating the CRM record, filing the Linear issue. A proper AI employee drafts these and waits for a human yes. Viktor is review-first by default, so the sensitive action sits in the thread until someone approves it.

Stage three: expand by audit

After a month, run the expansion audit. Do not browse the catalog for inspiration; interrogate your own week instead:

@Viktor look at the recurring work you have done for us this month and tell me: where did a human still have to fetch data or paste something between tools? Which integrations would remove those steps?

The answer is your next batch of connections, ranked by hours actually saved rather than by what looked shiny in the catalog.

Who should connect what?

Integrations are authorized per person, which means "connect Stripe" really means "someone with Stripe access authorizes it". Two rules of thumb keep this clean:

• The workflow owner connects the tool. The person who runs the pipeline report authorizes HubSpot. Ownership of the connection then matches ownership of the output, which matters the day a token needs reauthorizing.
• Do not over-provision on day one. Connect what stage one needs, nothing more. Every credential the AI employee holds is something your security review has to account for, and a security review is easier to pass with a short, justified list than a long speculative one.

The mechanics of authorizing, OAuth versus API keys and what each means, are covered in how to connect any tool to your AI employee.

What if the tool you need is not in the catalog?

A missing logo is not a dead end. With 3,200+ integrations the common stack is covered, but every team has one niche tool, and there is a standard path for it.

The path, in order of effort:

1. Search for the underlying service, not the brand name. Tools get acquired and renamed; the integration may exist under the platform's name.
2. Check whether the tool speaks MCP or has an API. Viktor can work with tools through MCP servers or their APIs even when there is no packaged integration. The full walkthrough is in how to connect tools your AI employee doesn't support yet.
3. Ask for it. If a tool you need is not connected, Viktor can build the integration. Say what the workflow is, not just the tool name, because the workflow defines which read and write operations actually matter.

The practical consequence: do not disqualify an AI employee, or delay a rollout, because one tool in your stack is missing from a catalog page. Ask about the path instead, and judge the vendor by the quality of the answer.

Frequently Asked Questions

How many integrations should we connect in the first month?

Most teams land between three and eight: the tools behind their top recurring workflows. The number itself matters less than the rule that every connection maps to a named workflow. If you cannot say which recurring task a connection serves, you do not need it yet.

Should we connect everything during onboarding to save time later?

No. Connections made "just in case" are credentials with no owner and no test. Connect what your first workflows need, prove those, then expand from the audit of where humans still shuttle data by hand. Adding an integration later takes a minute, so there is no prize for stockpiling.

What is the difference between read and write access, practically?

Read access lets the AI employee pull and summarize: reports, digests, lookups. Write access lets it change things: send, update, create. Read-only workflows are safe to run from day one. Write actions belong behind the approval gate, where the AI employee drafts and a human confirms.

Who in the team should authorize an integration?

The person who owns the workflow it serves, using their own access. That keeps permissions honest, because the AI employee acts within what that person could do anyway, and it means the connection has a clear owner when a token eventually needs refreshing.

What happens when an integration stops working?

The most common cause is an expired authorization, and the fix is the owner reconnecting in a minute. If a workflow that used to arrive goes quiet, check the connection first. Why your AI employee can't access a tool walks through the diagnosis in order.

Our most important tool is niche and not in the catalog. Is that a dealbreaker?

Usually not. If the tool has an API or an MCP server, Viktor can work with it even without a packaged integration, and you can request the integration be built. Describe the workflow you need, and judge the answer you get: a concrete path within days is the mark of a vendor who takes the long tail seriously.