Activating AI in one department: AI in HR, support and sales, step by step

Activating AI in one department means choosing one repetitive workflow and connecting Claude to the system that team already uses. You prove the result there before any other team starts. The method is the same for AI in HR, for a support desk and for a sales team. Only the system changes: HR records, a helpdesk or a CRM.

This guide is part 2 of 3 in our series on Department Activation. Part 1, AI in the finance team: what Claude takes over and for whom, covers what Claude can take on. Part 3, Why AI works department by department: measured results in marketing and beyond, covers results. This post covers how the work is done: the stages, who does what, how long each step takes and how each step is checked.

Department Activation is two weeks of focused work on one team, called a sprint. In a sprint, Claude is given a small set of that team's routines, and the team learns to use them. A sprint has a clear start and a clear end. It covers one department, not the whole company.

Starting narrow is not only our view. Salesforce's guide to AI agents for HR advises releasing the tool "to a single department first". StaffCircle gives the same advice for HR: pick one process, give it a named owner, and extend only once that pilot holds.

The best first workflow is one that happens often, follows the same pattern each time and does not decide anything about an individual person. StaffCircle's step-by-step guide to AI in HR puts it plainly: pick one high-volume, low-judgement process, give it a named human owner who signs off every output, and extend only once that pilot holds.

Drafting and collation are the safe ground. StaffCircle states that AI in HR "works best on drafting and collation, not on decisions about individual people." It names three processes that should never be a first pilot: redundancy selection, disciplinary outcomes and pay awards. The same logic applies outside HR. A first sales workflow should tidy records, not set discounts.

A quick test helps you shortlist candidates. Ask three questions of each workflow:

• Does the team do it every day or every week, in roughly the same way?
• Can one named colleague check every output before it changes anything?
• Does it live inside a system the team already uses?

If the answer to all three is yes, the workflow is a candidate. If you want ideas by task rather than by department, our overview of AI by business task lists workflows across functions. Choose one anchor workflow for the first sprint, plus a handful of smaller routines around it.

Claude becomes useful in a department when it can read from and act in the program that department already works in. For HR that is the HR records system. For support it is the helpdesk. For sales it is the CRM.

Salesforce describes the same pattern: an HR agent "connects directly to your company handbook, CRM platforms, and HR service management software". It uses secure APIs and respects user permissions.

There are four ways to make each connection:

• Connector: a ready-made link between Claude and a program.
• MCP server: a small service built for your system that gives Claude defined actions it may take there.
• Plugin: a packaged set of tools and instructions added to Claude.
• Skill: a written routine Claude carries out the same way every time.

Svennis builds one of these for each connection. We show you the options with their cost, and you choose. The work starts from the live system, with the permissions your team already has.

The model matters less than the connection, but it still matters. Anthropic's models overview positions Claude Haiku 5.5 for high-volume tasks "such as classification, extraction, and routing". It advises starting with Claude Opus 5.5 for most workloads if you are unsure. Some workflows need to run on their own rather than with a person typing. Such a workflow becomes an agent, the kind described on our page on AI agents built for single processes.

A Department Activation sprint lasts two weeks, and each week has one job. In week one, 5 to 8 skills or one anchor workflow are written. In week two, the team uses them on real work, with us beside them.

A skill, in this sense, is a routine Claude carries out the same way every time. Each skill has three parts: written instructions, test cases, and a colleague who owns it. Test cases are real examples with the answer the owner expects, so anyone can check whether the skill still behaves. A skill without a named owner in the department does not go into the sprint.

Week one is mostly writing and testing. The skill owner explains how the task is done today, including the exceptions. We turn that into instructions, connect Claude to the system, and run the test cases until the owner accepts the results.

Week two is use. The team runs the skills on its own tasks, and questions come up fast. We adjust instructions, add test cases for the cases nobody mentioned, and train the team on its own work rather than on generic examples.

At Svennis we ask each skill owner to write the test cases with us before the connection is built, because a skill the owner has not tested is the one the team quietly stops using. That rule costs a little time in week one and saves the sprint in week two.

Activating AI in one department needs a few named people on the client side and a clear timetable. The client's time commitment is modest: about an hour a week from the colleague who owns each skill, plus the team's training on its own tasks.

On the client side, four roles matter:

• Skill owners: one colleague per skill, who explains the task, writes test cases and signs off outputs.
• Claude ambassador: a colleague who uses Claude well, shows good examples, answers questions and collects the ones that repeat, for about an hour a week.
• The team: everyone who will use the skills, trained in week two.
• The managing director: who signs the targets at the start and presents the results at the end.

On our side, we write the skills, build the connections, sit with the team in week two and run the checks.

The timetable runs in this order:

1. Claude accounts set up and secured first, usually after Pilot B (Claude Team for the whole team).
2. The sprint itself: two weeks on one department.
3. A check on day 60: who uses Claude, whether they use it well, how much and for what, read from the admin console.
4. A readout on day 90, presented by your managing director, not by Svennis: results beside the signed targets, the recommendation and the proposed decision.

Projects usually run three sprints, each on one department.

AI in HR is, in Zendesk's definition, "the application of artificial intelligence technologies to streamline and enhance human resources functions". In practice, a first HR sprint works best on requests that arrive every week and follow a written policy. Leave requests are a good anchor workflow.

Salesforce describes the flow simply. When an employee asks for time off, the agent "checks the balance, verifies the policy, and submits the request to the manager". Here is how that looks when Claude is connected to an HR records system such as Zoho People:

1. An employee asks Claude for three days off.
2. Claude reads the employee's balance from the HR records.
3. Claude checks the leave policy in the handbook.
4. Claude submits the request to the manager, who approves or declines.

The HR officer owns the skill. The test cases cover the awkward requests. Examples are a balance that is too low, a request that clashes with a policy rule and an absent manager. The decision always stays with the manager.

A second skill can draft performance review write-ups from a manager's notes. StaffCircle warns that a suggested rating a manager waves through "is a rubber stamp with extra steps". It recommends that the rating field start empty. Claude drafts the text; the manager sets the rating.

For the pilot, StaffCircle suggests two measures. The first is elapsed time from the conversation to a completed record. The second is how much of the draft the manager changed. One review cycle is usually enough to see whether the saving is real. Our page on AI for recruitment and HR covers further HR workflows.

Support and sales follow the same steps as HR: one anchor workflow, one connected system, one owner per skill. What changes is the system and the kind of work.

Support on a helpdesk

In support, the anchor workflow is usually triage. Zendesk notes that AI "takes on repetitive manual work like ticket tagging and routing." Claude connected to a helpdesk such as Zoho Desk reads each new ticket, tags it, and sends it to the right queue. This is the classification and routing work for which Anthropic positions Claude Haiku 5.5. The support lead owns the skill, and the test cases are past tickets with the queue they should have reached.

Sales on a CRM

In sales, the anchor workflow is often record upkeep. Jeremey Donovan, writing on LinkedIn, lists "Auto-populate and clean up CRM" as a generative AI use case in sales. The need is clear in the figures Persana AI cites: 75% of businesses depend on CRM data, and about 60% of that data ends up incomplete, wrong or old. Claude connected to Zoho CRM can update a deal after a call, fill missing fields and flag duplicates for a person to merge.

Time is the reason to start here. Matt Dornfeld writes that most sales reps spend only 28% of their week actually selling. Julie Woods-Moss writes that research shows AI adoption in sales lags behind marketing and operations. Record upkeep is a low-risk way to close that gap.

The first department you activate should be the one where the anchor workflow is clearest and the owner is ready. This table compares the three departments in this guide across the decisions you make before a sprint.

Decision	HR	Support	Sales
System Claude connects to	HR records and the handbook	Helpdesk	CRM
Typical anchor workflow	Leave requests, review write-up drafts	Ticket tagging and routing	Updating and cleaning CRM records
Skill owner	HR officer	Support lead	Sales manager or sales operations
What stays with a person	Ratings, pay, disciplinary and redundancy decisions	Sensitive or complex tickets	Pricing, discounts and merges
What to measure	Time to completed record, share of draft changed	Tickets reaching the right queue first time	Complete fields, time spent on admin
First check	One review cycle	Day 60 usage check	Day 60 usage check

Zendesk's advice on return on investment applies to every column: track your key metrics over time and compare the benefits with the cost. Agree those metrics before the sprint, so the day-90 readout compares results against signed targets rather than impressions.

Most department rollouts that stall do so because of order, not technology. StaffCircle states that most stalled AI-in-HR projects "are not tool failures. They are sequencing failures." The common failures, and the check for each, are these:

• No owner. A skill nobody owns drifts. The check: no skill enters the sprint without a named owner.
• Rubber-stamp approval. A manager accepts every suggestion unread. The check: decision fields start empty, and the person fills them.
• Staff using public chatbots. StaffCircle calls a manager pasting performance notes into a public chatbot "the most common real-world AI incident in HR." The check: secured company accounts before the sprint starts.
• Seats in the wrong hands. StaffCircle warns that limiting AI to HR and a few senior managers loses the saving, which sits with line managers. The check: the day-60 usage review from the admin console.
• Uncapped agent usage. An agent's model usage is billed by Amazon Web Services on use and is not capped unless you set a budget. The check: set a budget before the agent goes live.
• Blocked tools. Anthropic notes that an organisation's settings can make a supported tool fail; an administrator can disable web search, for example. The check: test every skill under the real account settings.

Treat vendor time-saving figures with care. StaffCircle notes that such figures for individual HR tasks are almost never sourced. Your own measures from the sprint are the ones that count.

For a UK company, the practical questions are data, billing and policy. All three are easier to answer for one department than for the whole business at once.

Data comes first, because HR holds the most sensitive records. Zendesk notes that HR handles names, identity numbers and performance reviews. Before the first record goes near Claude, check whether you need a data protection impact assessment. For HR data you usually will.

Also sign processor terms with the provider, and confirm how any transfer of data outside the UK is covered. StaffCircle adds three practical questions: where the data is physically hosted, who inside the vendor can see it, and what the vendor is independently certified to for AI management. For the legal side, our guide to AI law in the UK for businesses sets out what applies.

Billing comes in two currencies. Claude seats are bought from Anthropic and billed in pounds. An agent's model usage is billed by Amazon Web Services in US dollars. Plan the budget for both, and set a cap on the second.

Your internal AI use policy can come after the pilot. The data protection checks above cannot: they come before any real record is used. StaffCircle advises: "Write the AI policy from what the pilot showed you."

One department gives you a live cycle to write from. It shows which data staff actually need, which outputs need sign-off, and where the rules must be strict. The second department then starts from a policy based on evidence rather than guesswork.

You can prepare for activating AI in one department before you speak to anyone. Five steps cover the groundwork:

1. List the repetitive workflows in one department and score each against the three questions on volume, owner and system.
2. Pick one anchor workflow that drafts or collates, and does not decide about a person.
3. Name the colleague who will own each skill and agree their hour a week.
4. Choose a Claude ambassador for the team.
5. Write down two or three targets your managing director is willing to sign.

Then check where you stand in the wider sequence. Svennis works in four steps: the Workshop, a pilot (Pilot Claude A, an AI agent on Claude for one process, or Pilot Claude B, Claude Team for the whole team), the department Activation, then monthly operations. Activation starts from Claude accounts that are already set up and secured, which is why it usually follows Pilot B.

Svennis runs this work for you, one department per sprint. The Department Activation service page sets out the sprint, what you prepare and the published prices. Bring your shortlist of workflows and your named owners, and the first sprint can start from your own work.