Claude for business in practice: from a Teams service desk to Zoho Desk

Claude for business is easiest to judge when it works inside the systems your staff already use, not in a chat window on its own. This post covers the packaged options Anthropic now offers and where they stop. It then looks at a live IT service desk inside Microsoft Teams that routes tickets into Zoho Desk.

Most firms start with a subscription and a few people typing questions. That helps the individual, but the underlying process stays the same. Things change when Claude can read a request, check the systems that hold your records, and put the result where a person can act on it.

So the question is not whether to use AI in general. You need to decide four things:

• which job to hand over first
• which systems that job has to touch
• how much control your staff keep
• what it will cost

The sections below take these in turn. The figures come from Anthropic's own documentation and from the UK data protection regulator, so you can compare the options on facts rather than impressions.

There are three broad routes. They differ in how far Claude reaches into your systems and in how much set-up they need.

Route	What you get	Fits when
Chat plans	Claude on web, desktop and mobile, with Projects for business context and Skills for repeatable tasks	Individuals need help drafting, analysing and summarising
Packaged connectors	Claude working inside a fixed list of popular tools, with ready-made workflows	Your key systems are on that list
Custom integration	Claude built into your own workflow, reading from and writing to the systems you choose	The work runs through tools the packages do not cover

Chat plans are the cheapest way in: Anthropic lists Claude Pro at 20 US dollars a month, and a Project gives Claude persistent context about your business across every conversation inside it.

The limit is that someone still copies information in and out. Packaged connectors and custom integrations remove that step. That is where time savings start to show up in a team's numbers rather than in one person's day.

On 13 May 2026 Anthropic announced Claude for Small Business. It describes the product as "a package of connectors and ready-to-run workflows that put Claude inside the tools small businesses depend on". The named tools are Intuit QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace and Microsoft 365.

The package includes 15 ready-to-run agentic workflows. They cover finance, operations, sales, marketing, HR and customer service. There are also 15 skills, built on the repeatable tasks that owners said slow them down most.

How control works

You switch the package on inside Claude Cowork, connect your tools and pick the job. Anthropic's wording is clear: "Claude does the work; you approve before anything sends, posts, or pays." Cowork is a separate tab in the Claude desktop app, so staff who use it need that app installed.

The approval step matters because it keeps a person accountable for anything that leaves the business. You should keep that principle whichever route you choose.

If your finance, sales and documents already sit in the listed tools, this package is a sensible first test: the connectors are switched on inside Claude, not built.

Packaged connectors cover what the vendor has chosen to support. Claude for Small Business names seven tools. The Enterprise plan lists connectors to Google Drive, Gmail, Google Calendar, GitHub, Microsoft 365 and Slack. Zoho Desk does not appear on either list.

That gap is common. Many firms run their service desk, CRM or job management on a platform outside the big names. The work that most needs automating often lives there: ticket queues, customer records and escalation rules.

What a custom integration adds

A custom build connects Claude to the specific systems your process depends on. It can also carry your own rules: which categories exist, who owns each queue, and when a human must step in. A packaged workflow cannot know these, because they are unique to your organisation.

The trade-off is effort. A packaged connector takes seconds to switch on. A custom integration has to be designed, tested against real requests and maintained as your systems change.

So the practical test is simple. List the systems a job touches. If they are all on a supported list, start with the package. If the core record sits elsewhere, as with a service desk on Zoho Desk, you are looking at an integration project and should plan and budget for one.

Here is what that integration route looks like in production. For Asset Services Group (Message Direct), Svennis built an IT service desk on Claude inside Microsoft Teams, sitting in front of Zoho Desk, which reaches 99.7% first-time-right routing and a 40% efficiency gain, as their Head of Technology states on the record in the published Asset Services Group case study.

The pattern is worth studying because it is plain. Staff raise requests where they already work, in Teams. The ticket ends up in Zoho Desk, the system of record the IT team already runs. Claude sits between the two and decides where each request should go.

Why routing is a good first job

• It is bounded. There is a fixed set of destinations, so a right answer can be defined.
• It is measurable. You can count how often a ticket lands in the right place first time.
• It keeps people in charge. Claude sorts the request, and a human engineer still resolves it.
• It removes a known cost. Every misrouted ticket is passed around before anyone starts on it.

Few other jobs combine a clear success measure, low risk per decision and high volume so neatly. That makes routing a strong candidate for a first project.

If you are considering something similar, a handful of decisions matter more than the choice of model. Settle them before any build starts.

Meet people where they already are

If staff live in Teams, the assistant belongs in Teams. A separate portal adds a step, and every extra step lowers uptake. The same logic applies to Slack or email if those are your main channels.

Keep one system of record

The assistant should write into your existing service desk, not keep its own list of tickets. Reporting, service levels and audit trails then stay in one place, and your IT team does not have two systems to watch.

Write the routing rules down first

Claude can only route well against categories and owners that are clearly defined. If your team argues about where a type of request belongs, settle that before the build. Otherwise you are automating the disagreement.

Decide where a person must approve

Sorting a ticket is low risk. Resetting access, changing permissions or contacting a customer is not. Anthropic's own package requires approval before anything is sent, posted or paid, and that is a sound default for your own design.

Measure from day one

Record how often routing is right first time before you launch as well as after. Without a baseline, any efficiency figure is a guess.

Pricing depends on the route, and the cost structures differ in ways that affect budgeting.

Chat plans

Claude Pro is 20 US dollars a month per person. It comes with a usage allowance that refreshes over a rolling window of about five hours, which Anthropic adjusts over time, so it suits individual use rather than a shared queue of work.

Enterprise

The Enterprise plan works differently. It uses a single seat type, priced per user per month and billed annually. That seat covers Claude on web, desktop and mobile, plus Claude Code and Cowork. The seat fee covers access only. All usage is billed separately at standard API rates, with no included token allowance and no per-seat usage limits.

Other points to know:

• Self-serve Enterprise needs at least 20 seats. Sales-assisted needs at least 50.
• Both versions have the same features and the same seat pricing.
• Admins can set spend limits for the whole organisation and for individual users.
• Enterprise is also sold through the AWS Marketplace.

For a custom integration such as a service desk, the model cost is usually driven by volume of requests, not number of staff. Ask for an estimate based on your actual monthly ticket count, and set spend limits before launch. Our services and pricing page sets out how we scope and charge for integration work.

Data security is the most common worry. In Anthropic's survey of small business owners, half named it as their single biggest hesitation about AI. Anthropic states that it does not train on your data by default on its Team and Enterprise plans.

Enterprise adds controls that matter for a service desk handling staff details. These include audit logs, SCIM for identity management, custom data retention controls, a Compliance API and customer-managed encryption keys. The Compliance API gives programmatic access to activity logs, chat histories and file content, filtered by user and time range.

What the ICO expects

The ICO's Guidance on AI and data protection was last updated on 15 March 2023. It is now under review because of changes made by the Data (Use and Access) Act. It includes content on what to consider in a DPIA. It also covers how solely automated decision-making links to fairness, an area the Data (Use and Access) Act reshaped in February 2026, when Articles 22A to 22D of the UK GDPR replaced Article 22. A routing assistant that only sorts tickets for a human to resolve is a very different case from a system that makes decisions about people on its own.

Plan for people's rights too. Under the right to erasure, an organisation has one calendar month to respond, extendable by up to two months. It should also tell anyone it shared the data with, unless that is impossible or involves disproportionate effort. Your integration must let you find and delete records across every system it touches. Our guide to AI law in the UK covers the wider picture.

The service desk example works because the job has the right shape. Use the same test on your own candidates before you commit money.

1. High volume. The task happens many times a week, so small gains add up.
2. Clear right answer. You can say whether each outcome was correct.
3. Low cost of a single error. A wrong result is caught and fixed without harm.
4. Existing system of record. The output lands somewhere your team already works.
5. A named owner. One person is responsible for the rules and the results.

Ticket routing, enquiry triage and sorting inbound requests usually pass all five. Tasks that make final decisions about people, money or health usually fail the third test. Those should start with Claude drafting and a person deciding.

The same approach carries across sectors. Practices have their own high-volume, rule-based tasks, from appointment reminders to request handling. See how this applies to dental practices and veterinary practices.

Resist the urge to start with the most impressive idea. A modest job that works and can be measured gives you evidence for the next one.

You can make real progress in a few weeks without a large commitment. Work through these in order.

1. Pick one job. Apply the five tests above and choose the candidate that passes most clearly.
2. Map the systems. Write down every tool the job touches. Check each against the connectors Anthropic lists for Claude for Small Business and for Enterprise.
3. Take a baseline. Measure the current volume, error rate and time taken for a typical month.
4. Write the rules. Agree categories, owners and the points where a person must approve. Settle disputes now.
5. Check data protection. Decide whether you need a DPIA, confirm your plan's training and retention settings, and make sure you can meet an erasure request across every connected system.
6. Choose the route. If every system is supported, trial the packaged connectors. If your core record sits elsewhere, scope an integration and estimate model costs from your real volumes.
7. Set spend limits. Put limits in place before launch, not after the first invoice.
8. Review after a month. Compare results against the baseline, then decide whether to widen the scope or move to the next job.

Keep a person accountable for every step that leaves the business. The approval principle protects you while you learn where the tool is reliable and where it is not.