Rolling out Claude properly across an SMB

The default failure mode

Pattern we see at least once a fortnight: an owner reads about Claude (or sees it on LinkedIn, or watches a competitor talk about AI on a podcast), buys 5-10 Claude.ai Teams licences, hands them out, and waits for the productivity gains to land. They don’t. The early adopters get value, the rest of the team doesn’t open it after week two, and the renewal conversation 11 months later is uncomfortable.

The problem isn’t the tool. Claude is genuinely transformative when it’s embedded properly. The problem is that “hand out licences” isn’t a rollout - it’s a procurement event. A real rollout has three layers: structured Projects per function, custom Skills, and a written governance policy with a named owner. Skip any of the three and the adoption curve flattens fast.

Layer one: Projects per function

Claude Projects are scoped workspaces with curated context - documents, instructions and prior chats - that the model uses for every conversation in that project. Set up correctly, a Project is a purpose-built version of Claude that already knows your business inside that context.

The default rollout pattern for an Australian SMB:

• Operations Project - SOPs, your org chart, your tech stack, role descriptions. Used for process design, role clarity, internal comms drafting.
• Sales Project - your pitch deck, ICP definitions, objection-handling notes, win/loss patterns. Used for proposal drafting, discovery prep, follow-up sequencing.
• Marketing Project - your brand voice, content pillars, recent campaigns, customer language. Used for content drafts, campaign briefs, copy review.
• Finance/Ops Project - your reporting cadence, approval workflows, basic policy docs. Used for analytical reviews, board pack drafting, budget conversations.
• Customer Service Project - your tone, SLAs, common issues, refund/escalation policy. Used for response drafting, triage assistance, knowledge gaps.

Each Project takes 30-90 minutes to set up properly. The leverage comes from the team using a Claude that already knows their function, not a generic chat assistant they have to brief from scratch every conversation.

Layer two: custom Skills

Skills are reusable instruction sets that you build once and invoke across Projects. The patterns we install most often:

• Brand voice skill - the tone, vocabulary, don’t-use-words and formatting rules your business writes by. Invoked any time the team is producing customer-facing copy.
• Framework skills - the proprietary frameworks your business uses to structure thinking (your sales methodology, your ops review cadence, your content brief format).
• Document templates - proposal structure, board pack outline, weekly leadership update format. Skill encodes the structure; team fills in the substance.

The compounding effect of skills is large and underappreciated. They turn Claude from “a smart assistant” into “an assistant that knows our way of doing things,” which is materially harder for a competitor without the same skill library to replicate.

Layer three: governance and the named owner

Before licences land in user inboxes, write a one-page AI policy:

• Which AI tools are approved for company use
• What categories of data can and can’t be put into them
• How customer/client information is handled and whether it crosses borders
• Who owns AI-generated IP (default: the company, mirroring employment IP terms)
• How a suspected breach or misuse is reported

One page. Signed off by the owner, IT and (where applicable) legal. Distributed alongside the licence, not three months after. Most data and IP problems we’ve seen with SMB AI rollouts came from shipping the tool without the policy.

The other half of governance is appointing a single named AI lead. Doesn’t have to be full-time - usually it’s an existing ops or product person who genuinely cares about the topic. What matters is that there’s one human accountable for adoption, governance and the next 90 days of rollout. Without one, AI work dies between meetings every time.

The 90-day adoption pattern

Once the three layers are in place, the rollout itself runs on a predictable curve:

• Days 1-14 - launch session per function (45 min), walk the team through their Project, show three real workflows live.
• Days 14-45 - weekly office hours run by the AI lead. People bring real problems, the lead shows how to solve them with their Project. Wins get shared in a #ai channel or its equivalent.
• Days 45-90 - the AI lead instruments which Projects are getting used and which aren’t. Underused Projects get iterated on. Custom Skills get built where the team is repeating the same prompts.

By day 90, the rollout is either landed - daily-active across the team, with stories you can tell - or it isn’t. A landed rollout gets you 5-10x the value of the licences-only version of the same subscription.

When to add Claude Code (and when not to)

If your business has internal engineers, Claude Code becomes the obvious next surface. The integration with the rest of the rollout is tighter than people assume - the same governance policy applies, and the engineering productivity uplift typically justifies the cost inside a quarter.

If you don’t have engineers, Claude Code isn’t relevant yet. Don’t buy it because it’s in the marketing copy. It’s genuinely useful only inside engineering workflows.

Signals the rollout is working

Three weeks after launch: are people opening Claude daily? Eight weeks after launch: have at least three custom Skills been written based on real team prompts? Twelve weeks after launch: can a random team member describe a workflow they’ve personally changed because of Claude? If all three are yes, the rollout is real. If any one is no, that’s the layer to debug.

How XLev installs this

A standard XLev Claude implementation runs 6-10 weeks. Week 1 is the AI Strategy Workshop (if you don’t already have a plan). Weeks 2-4 install Projects, Skills and governance. Weeks 5-10 run the adoption programme alongside your team, with weekly office hours and Skill iteration. We exit when the rollout is landed - not when a calendar date hits.

See the full service description on Claude Implementation, or take the AI Readiness Scorecard to see where you sit today.