Building an AI workflow for a business function means turning a recurring task into a tooled, repeatable sequence — connected to your data and your tools — instead of retyping a prompt every time. An isolated prompt gives a one-off, variable result; a workflow gives a reliable, measurable, shareable one, because the method, the context and the sources are encoded once and for all. This article sets out the difference between prompt and workflow, the six-step design method for tooling a business function (identify the task, structure the steps, connect the data, make it reliable, measure, capitalise), and gives examples by department — legal, finance, HR, marketing, support, operations — illustrative and generic.
This is a deep dive into Step 2 of our pillar guide, “How to bring Claude into your company's processes”. One framing point: METASENSE is a Creative Tech agency based in Vélizy-Villacoublay, and our standing on AI does not come from a slide — we use these models ourselves, every day, to build platforms and produce assets. What we pass on here is a method you can transpose to your own business functions: the knowledge of your field stays yours, we bring the method and the command of the tool.
Worth remembering — A prompt solves one case; a workflow solves a category of cases, every time, the same way. The jump in value does not come from a “better prompt”: it comes from structure (steps + data + guardrails) and from capitalisation (the workflow becomes a shared asset, not know-how locked in one person's head).
Tool a first business workflow →
AI workflow or isolated prompt: what is the real difference?
A prompt is a one-off request; an AI workflow is a tooled, repeatable sequence that handles a recurring business task by bringing the right context, the right data and the right guardrails to every run. The distinction is not cosmetic: it decides whether AI stays a gadget or becomes a real gain. As long as a task depends on finding “the right prompt” buried in a chat history, it is neither reliable, nor transferable, nor measurable.
| Isolated prompt | AI workflow per business function | |
|---|---|---|
| Unit of work | One request, one result | A repeatable sequence for a category of tasks |
| Context | Restated manually every time | Shared and persistent (via a team Project) |
| Data / tools | Copy-pasted by hand | Connected to your internal sources through MCP |
| Know-how | In one person's head | Encoded in a reusable Skill |
| Execution | One step, supervised | Multi-step, with checkpoints (Cowork) |
| Result | Variable, untraceable | Reliable, measurable, shareable |
| Transferability | Low (leaves with the person) | High (becomes an asset of the organisation) |
Three working definitions, to fix the vocabulary for the rest of the article:
- An AI workflow is a tooled, repeatable sequence of steps that handles a recurring business task, where a prompt is only a one-off request.
- A recurring task is a task that comes back regularly with a stable structure (same inputs, same criteria, same output format) — the ideal candidate for a workflow.
- Tooling a task means giving it the context, the data and the controls it needs so the result is reliable without reinventing it at every run.
The right instinct: do not try to “prompt better”. Look for the recurring task you do ten times a week in the same way — that is the one that deserves a workflow. A better prompt improves one case; a workflow settles the whole category.
How do you design an AI workflow for a business function, step by step?
Designing an AI workflow for a business function takes six steps, in this order: (1) identify the high-leverage recurring task, (2) structure the steps, (3) connect the data and tools, (4) make it reliable with guardrails, (5) measure the gain, (6) capitalise it as a shared Skill. Every step is done with the experts of the function concerned: they know the real inputs, the edge cases and the compliance requirements; the method and the tooling are our part. The workflow that comes out is yours, tooled — not a recipe pasted on top.
| # | Step | The question it settles |
|---|---|---|
| 1 | Identify the task | Which recurring, high-leverage task deserves to be tooled? |
| 2 | Structure the steps | What is the sequence: inputs → processing → output? |
| 3 | Connect data & tools | Which sources and tools must the workflow rely on? |
| 4 | Make it reliable | Which guardrails, and which human checkpoints? |
| 5 | Measure | How do you prove the gain (time, quality, volume)? |
| 6 | Capitalise | How do you freeze the workflow and share it with the team? |
Step 1 — Identify the right task: which one to tool first?
The first step is to spot a recurring, time-consuming task with a stable structure — not the most impressive one, but the one that combines the best leverage with the lowest risk. A good candidate comes back often, always follows the same logic, draws on information that is available, and tolerates a human review. Bad candidates: rare tasks, high-stakes decisions with no control, or tasks that already run smoothly, where AI would only add cost.
A simple grid to decide: cross frequency (how many times a week?), unit time (how many minutes saved?) and risk (what happens if an error goes undetected?). The ideal first workflow is frequent, time-consuming, and carries manageable risk.
Honest note — Good framing also says no. Plugging AI into a task that already runs smoothly, or into a decision where an error cannot be caught without human control, adds cost without value. One workflow that holds in production beats ten demos that impress in a meeting.
Step 2 — Structure the steps: what does the sequence look like?
Structuring a workflow means breaking the task into a clear sequence — inputs, processing, output — before you even think about the tool. A vague workflow produces a vague result. Write it down in black and white: which inputs arrive (and in what format), which transformations follow one another, what output format is expected, and where a human validates.
In practice, for a document analysis task, the typical sequence is: document received → key elements extracted against a grid → checked against internal rules → summary drafted in the standard format → human validation before circulation. That explicit statement is what makes the result reproducible — and it is that, not the model, which accounts for 80% of the final quality.
Step 3 — Connect data and tools: what does the workflow rely on?
A business workflow is only worth the data and tools it can reach: that is the role of MCP (Model Context Protocol), Anthropic's open standard that connects Claude securely to your internal sources and applications. Without that connection, AI works blind on whatever is copy-pasted into it; with it, the workflow draws on your real data (document management, project tracking, CRM, messaging) within the boundaries you set.
Three Claude building blocks combine here, each with a distinct role:
- Projects hold a team's shared context: reference documents, instructions, standards, so that every member starts from the same base.
- MCP connects the workflow to your internal tools and data — the “nervous system” that links Claude to your real environment.
- Anthropic's business-function plugins (11 official plugins: Sales, Marketing, Legal, Finance, HR, Data, Design, Engineering, Operations) offer starting points by function, to be adapted to your context.
Data guardrail: before you connect anything, agree with your compliance and security teams which data may be exposed and which must never be. The technical connection comes after the rule, not before.
Step 4 — Make it reliable: which guardrails for a safe result?
A reliable workflow is not a workflow without humans: it is a workflow where the human checkpoints sit exactly where an error is expensive. Reliability is built from concrete elements: an explicit criteria grid, an imposed output format, reference examples, and above all human validation at sensitive steps (before an external send, a decision, a circulation).
This is also the step where you test on real cases — including the edge cases supplied by the business experts — and adjust until the result is stable. A workflow that works one time in two is not a workflow: it is a demo. You iterate on varied inputs before you consider the task tooled.
Step 5 — Measure: how do you prove the gain?
Measuring means tying the workflow to a readable before/after indicator: time per task, volume handled, rework rate, quality. Without measurement, you cannot defend the use case at the next budget round, nor decide what to extend. Set the indicator from the start: how long the task used to take, how long it takes now, and what the team does with the time freed up.
Measurement serves two decisions: keep or cut (a workflow that does not prove its gain has no reason to stay) and extend (a proven gain in one function justifies tooling the next). This is also where you tie the gain to the cost: a workflow that saves an hour but burns a lot of tokens has to be arbitrated — a subject covered in the article “Governing the cost of AI tokens in the enterprise”.
Step 6 — Capitalise: how do you share the workflow with the whole team?
The last step turns a workflow that works into a reusable Skill, shared with the team — so the know-how does not stay in one person's head. A Skill (Claude) is a repeatable piece of know-how you teach Claude (instructions, standards, sometimes scripts) and that it applies automatically in the relevant conversations. That is what moves a workflow from “personal trick” to “organisational asset”.
This is the lever for autonomy: instead of a mega-prompt nobody maintains, the team builds a library of versioned, shared know-how. The full method to structure, test and roll out a Skill is set out in the companion article “Creating Claude Skills and sharing them across the organisation”.
Worth remembering — The six steps are not interchangeable. You identify before you structure, you make reliable before you measure, and you capitalise last. Capitalising a shaky workflow means spreading an error at scale.
What are some examples of AI workflows by department?
Every function has recurring tasks with a stable structure: here are generic examples by department to help you spot your own candidates. These examples are illustrative — the right workflow for your organisation is designed with your experts, from your real inputs and your constraints. They exist to start the thinking, not to be copy-pasted.
| Department | Candidate recurring task | Shape of the workflow (illustrative) |
|---|---|---|
| Legal | First pass on supplier contracts | Sensitive clauses extracted against a grid → checked against internal standards → note of points to watch → mandatory review by a lawyer |
| Finance | Reporting preparation / summary of documents | Data aggregated → put into the standard format → commentary drafted → controller validation |
| HR | Pre-screening of applications | A batch of CVs read against a criteria grid → summary of strengths and points to watch → human HR decision |
| Marketing | Adapting briefs and content | Structured brief → variants per channel based on brand guidelines → editorial review |
| Support | First qualification of requests | Sorting and categorisation → suggested answer drawn from the knowledge base → agent validation before sending |
| Operations | Meeting-notes summaries / follow-up | Notes read → decisions and actions extracted → project tracking updated through MCP |
What all these examples share: the human keeps the decision, AI takes on the repetitive preparation. That is exactly the logic of a case we ran with Bertrandt — an ideation and design workshop for an AI assistant for CV analysis, where AI serves a precise HR need, co-built with the teams concerned. (Real case, citable.)
Honest note — None of these workflows replaces the expertise of the function: they remove the repetitive part to give time back to the part that requires judgement. A workflow claiming to decide in place of a lawyer, an HR director or a controller is a bad workflow.
Where do the Claude building blocks fit in a business workflow?
Four Claude building blocks cover the needs of a business workflow: Projects for shared context, Skills for know-how, MCP for data and tools, and Cowork for multi-step execution. Understanding what each one does avoids the classic mistake — trying to do everything with a single “mega-prompt”.
- Projects — team context. A space that holds shared reference documents, standards and instructions, so everyone starts from the same base.
- Skills — encoded know-how. The method and standards for a task, taught to Claude once and applied automatically (the asset you capitalise at step 6).
- MCP — the connection to data and tools. The open standard that plugs Claude securely into your internal sources (documents, CRM, project tracking, messaging).
- Cowork — multi-step execution. Anthropic's feature that runs tasks in several steps (generally available since April 2026), for workflows that genuinely chain several actions.
The useful distinction: Skills = the internal “how to do it”; MCP = access to the outside world. The two are complementary, not competing. When a workflow chains several steps and several tools, you enter orchestration territory — covered in the companion article “Orchestrating AI: from the isolated prompt to multi-step agents”.
The right instinct: you do not bring in all four blocks on principle. A simple workflow can run on a Project and a Skill; MCP and Cowork are added when the task justifies it, not to look modern.
Where do you start to tool a first business function?
Start small: one function, one high-leverage recurring task, one workflow that holds — then you capitalise and extend. Trying to automate everything at once is the most common and most expensive trap. Proof through use convinces better than any top-down instruction. A realistic path:
- Pick a pilot function where the pain is clear and the value measurable.
- Identify 1 recurring task with high leverage and manageable risk (step 1).
- Structure the sequence with the experts of that function (step 2).
- Connect the data you need, once compliance has been framed (step 3).
- Make it reliable with guardrails and human checkpoints (step 4).
- Measure the before/after gain (step 5).
- Capitalise as a Skill and extend to the next function (step 6).
That is precisely the logic we applied with Grand Angoulême: a conference and a co-design workshop on agentic AI workflows built for non-specialist participants — we bring the method, the participants bring their business function. (Public-sector client, citable.) It is also what we pass on through skills transfer (modules with Strate, the design school, and Gocad Lab: 4 workshops, 24 staff): the goal is your autonomy, not your dependence. (METASENSE has a real training and skills-transfer activity; it is not certified.)
Worth remembering — The first workflow does not need to be spectacular; it needs to hold and to prove a gain. Once the proof is there, extending it becomes obvious to teams and leadership alike.
Tooling a business workflow with your team
You have the licences and the recurring tasks; what remains is turning them into workflows that hold. METASENSE (Vélizy-Villacoublay) co-builds a first high-leverage business workflow with your experts — from identifying the task to capitalising it as a Skill — and trains your teams towards autonomy. Claude as the lead tool, with a method that adapts to other assistants.
Explore enterprise AI enablement & Claude workflows →
This article is a deep dive into Step 2 of the pillar guide “How to bring Claude into your company's processes”. To go further: capitalise as Skills, orchestrate several steps, govern the cost. And if the question is an assistant exposed to your customers: “Do you need an AI agent on your website?”.
FAQ
What is the difference between a prompt and an AI workflow?
A prompt is a one-off request that gives a single, variable result. An AI workflow is a tooled, repeatable sequence that handles a recurring business task by bringing the right context, the right data and guardrails to every run. The workflow makes the result reliable, measurable and shareable, where the prompt stays a personal trick.
How do you build an AI workflow for a business function with Claude?
In six steps: identify a high-leverage recurring task, structure the sequence (inputs, processing, output), connect data and tools through MCP, make it reliable with human checkpoints, measure the before/after gain, then capitalise it as a shared Skill. Every step is done with the experts of the function, who bring the field knowledge.
Which task should you automate first?
A frequent, time-consuming task with a stable structure, where an error can still be caught by a human review. Cross frequency, unit time saved and risk: the right first workflow combines high leverage with controlled risk. Avoid rare tasks, tasks that already run smoothly, and high-stakes decisions with no control.
What are some examples of AI workflows by department?
Legal: first pass on contracts against a grid. Finance: reporting preparation and summary of documents. HR: pre-screening of applications. Marketing: adapting briefs per channel. Support: qualification and draft answer. Operations: summarising meeting notes. In every case, the human keeps the decision and AI takes on the repetitive load.
How do you connect Claude to my internal data and tools?
Through MCP (Model Context Protocol), Anthropic's open standard that links Claude securely to your internal sources: document management, CRM, project tracking, messaging. Before any connection, agree with your compliance teams which data may be exposed and which may not. The data rule always comes before the technical connection.
What is the difference between a Skill and MCP?
A Skill encodes the internal “how to do it” of a task — your method and your standards, applied automatically by Claude. MCP gives access to the “outside world”: your internal data and tools. The two are complementary, not competing: the Skill carries the know-how, MCP carries the connection. A good workflow often combines both.
Should you orchestrate several steps, or is a single prompt enough?
It depends on the task. A simple workflow can run on a Project and a Skill. Multi-step orchestration (through Cowork) is justified when the task genuinely chains several actions and several tools. Orchestrating a shaky step only industrialises the error: you orchestrate once the workflow is reliable, never before.
How do you measure the gain from an AI workflow?
By setting a readable indicator from the start and comparing before and after: time per task, volume handled, rework rate, output quality. Measurement serves two decisions — keep or cut a workflow, and decide what to extend. It also ties the gain to the token cost, so heavy workflows can be arbitrated.
Can an AI workflow replace business expertise?
No, and that is not the aim. A good workflow removes the repetitive part of a task to give time back to human judgement, which keeps the decision. A workflow claiming to decide in place of a lawyer, an HR director or a controller would be a bad workflow. Business knowledge stays with your experts; AI tools it.
Do you need to be technical to design an AI workflow for a business function?
No. Design starts from knowledge of the business function, which your teams already have. The technical part (connecting data through MCP, structuring a Skill) can be learned and transferred. It is precisely because your staff are not AI experts that co-design and training towards autonomy create value.
Sources
- Anthropic — Cowork and plugins for teams across the enterprise (business-function plugins, Customize menu)
- Anthropic — Model Context Protocol (MCP)
- Anthropic — Plans & Pricing (Claude Team, Claude Enterprise, Projects, Skills)
- IntuitionLabs — Claude Skills vs. MCP: A Technical Comparison for AI Workflows