How to Build an AI Workflow From Your Skills

An AI workflow is several skills linked into a whole process, such as lead to proposal or month-end close. Each step is one skill with its own context, knowledge and app connections, and a person approves at set points. You build one by getting each skill working on its own first, then defining what each step hands to the next.

If a skill is a new starter's job card, a workflow is the run of job cards that takes a piece of work from start to finish. Each card is still its own job, with its own steps and checks. The workflow says what order they come in, what each one hands to the next, and who signs off along the way.

A workflow isn't one giant skill. Keeping each step as a focused skill makes it easier to test and fix, and a change to one step doesn't put the others at risk.

Each step in a workflow has six parts:

• The skill: the method for this step.
• The context: who the work is for, such as your ideal customer or approval rules.
• The knowledge: the documents the step reads, such as templates or past proposals.
• The apps: what the step touches, such as your CRM or email.
• The output: what this step hands to the next.
• The approval: whether a person signs off before the work moves on.

There's no separate workflow engine involved. Your AI tool runs each skill, either when someone asks or on a schedule. It reaches your apps through its own connectors. Several AI tools can run tasks on a schedule: Claude's scheduled tasks (Anthropic: Get started with Claude Cowork), scheduled automations in Microsoft 365 Copilot Cowork (Microsoft: Use Copilot Cowork) and routines in Grok Bot (xAI docs: Skills, routines and automations).

Here's a sales workflow laid out step by step:

The handoffs are what make it a workflow. Each step saves its output somewhere the next step can find it, such as the deal record or a folder for the client. The next skill's instructions say where to look.

Summit Training, a fictional corporate training provider, gets a new enquiry from Harbour Foods, a fictional food distributor.

• Qualify. The qualification skill checks the enquiry against Summit's ideal customer and updates the CRM with a score and the reasons. Harbour Foods qualifies.
• Research. The research skill gathers what's public about Harbour Foods and notes which case studies are most relevant. It saves a one-page brief to the client folder.
• Prepare. The meeting skill reads the brief and drafts an agenda and five questions for the account manager, saved to the client folder.
• Propose. After the meeting, the proposal skill reads the notes and the brief and drafts the proposal from Summit's template and price list. It checks every price against the price list, then stops for the account manager to review.
• Contract. Once the proposal is accepted, the contract skill drafts the agreement from past statements of work. A person approves it before it goes to the client.

The approvals are written into the skills themselves, as steps in each skill's instructions. The harness is what brings each step's skill, context and documents together. Finance workflows follow the same pattern with tighter approvals. A month-end run might reconcile the bank, raise invoices, chase overdue accounts and prepare a cash-flow snapshot, with every payment held for a person's explicit approval.

• Choose one process with a clear start and end, such as enquiry to proposal.
• Write the steps as separate jobs, one skill for each.
• Build each skill and test it on its own before linking anything (How to Create an AI Skill).
• For each step, write the handoff: what it produces, where it saves it and what the next step reads.
• Mark the approval points: anything that's sent, paid, signed or changes a record.
• Check that your AI tool has a working connector for each app in the workflow.
• Run the whole workflow once by hand, step by step, with a real example.
• Schedule only the steps that suit it, such as a weekly overdue-invoice check, after they've passed your tests.

• A normal run. Take a real lead from enquiry to draft proposal. Check each handoff landed where the next step expected it.
• Incomplete input. Use an enquiry with no company size. The qualification step should flag the gap, and the later steps shouldn't carry on as if it were filled.
• An unrelated request. Partway through, ask the AI for an internal newsletter. None of the workflow's skills should take over.

A skill tells your AI what to do. Your knowledge base shows it how, in detail. Your company context tells it who it's for. The Contexta harness brings them together, in every AI.

Contexta supplies the skills, company context and knowledge links for each step. Your AI runs them, and it works with the apps it's already connected to, such as your CRM, Xero or Slack, through its own connectors. Contexta doesn't connect to those apps itself, and it isn't a workflow engine. Each step draws on the right space: the sales space for proposals, the finance space for invoicing, each with its own home in your SharePoint or Google Drive and its own rules.

Because each step is its own skill with its own version history, you can improve one step and restore it if a change goes wrong, without touching the rest. You can build workflows from your own skills, or have our advisory and implementation team build sales, delivery and finance workflows with you through Contexta Managed AI. We run our own sales, delivery and finance workflows this way. See How It Works and Product.