Your business just got AI. How to implement it properly

Your business has bought enterprise AI licences. I would start by asking each person to pick a recurring task in their own role that they understand well enough to check. A weekly update, a research brief or preparation for a meeting gives them something concrete to work on. They can see what improves, what fails and how much effort the review takes.

Rebuilding the customer relationship management system is a much bigger commitment. It affects shared records, reporting, integrations and other people's work. A licence gives you access to a tool. Learning to use it well takes a clear job, useful information, appropriate access and a way to judge the result.

Our recommendation is to start with your own workflow, make it useful to the team, then improve the handovers between teams. The pace depends on the work and the evidence. There is no fixed timetable, and some organisations will already have shared workflows mature enough to start further along.

Onboard it like a new colleague

Treating AI as a new colleague is a useful starting mindset. Explain the job, who the output is for, where to find the relevant information and what a good result looks like. Show it an example. Make the limits explicit and review the work. A person remains responsible for the result.

An AI agent is an AI system that can use tools to carry out steps towards a task. Depending on its setup, it might read files, create a draft or change something in another application. That makes the brief and the boundaries more consequential than they are in a chat that only returns text. Check which capabilities your organisation has enabled before designing a workflow around them.

For a weekly update, the job might be to summarise progress and blockers using this week's project notes. The output is a draft for the project owner. An approved previous update shows the expected length and level of detail. Sending it to colleagues is a separate action with its own permission.

Choose a workflow you can inspect

Begin with the outcome you need and how you produce it today. Write down the inputs, the steps, the people involved and the parts that cause delay or rework. You may find that the first improvement is a clearer source document or a better handover. Choose the AI task after you understand that problem.

Keep the first attempt narrow enough to inspect. Drafting an update from notes you know is easier to assess than asking an agent to redesign the entire reporting process. Anthropic's guidance on building agents recommends starting with the simplest approach that works and adding complexity when it is justified. A good first task also gives you a clear place to stop and review.

• Anthropic's guidance on building effective agents

Give shared work an organised home

Agree where the project lives before people start making their own copies. Keep the brief, current state, source material, examples and output together, with a named owner. Give each type of information a clear scope. Company rules belong in a shared home, personal preferences belong with the person, and project instructions belong with the project.

• Company-wide information covers business goals, shared definitions, approved tools and common working rules.
• Personal information covers the person's role, writing preferences and habits. It should not silently override shared project requirements.
• Project information covers the task's purpose, owner, current position, source files, examples, decisions and completion checks.

These are organisational choices, not a universal folder structure that every AI product reads automatically. Chat applications may use project instructions and uploaded files. File-based agents may read named instruction files and folders. OpenAI documents global and project instructions for Codex through AGENTS.md. Claude Code documents its own personal, project and organisation instruction scopes. Configure and test the product you use.

• OpenAI documentation on AGENTS.md instructions
• Claude Code documentation on persistent instructions

Give it relevant context and a working agreement

Context is the information available to the AI for its current task. The useful question is what it needs to do this job well. Give it the goal, the current facts and a representative example, then point to deeper material it can read when needed. Anthropic's context-engineering guidance favours a focused set of useful information and retrieving additional detail as the task requires it.

Markdown is plain text with headings and lists. It is a practical format for instructions used by file-based agents because people can read it, edit it and move it between systems. A short start-here file can explain the project and point to the sources. Give shared documents an owner, record material changes and remove obsolete instructions from the active reading path.

Write a working agreement for each project. It should state the goal, the human owner, the files to read, what the agent may do, when it must ask and how the work will be checked. Use concrete boundaries. For example, reading notes and saving a draft may be allowed, while sending the update or editing a shared record needs approval.

The agreement tells the agent how to behave. Product permissions determine what it can access or change. Check both. A sentence in a Markdown file cannot substitute for a permission setting that prevents an unwanted action.

• Anthropic on selecting and maintaining useful context

Save the method and add the tools it needs

Once the task works, save the repeatable method as a skill. A skill packages task instructions and may include examples, resources or scripts. In the Agent Skills format, SKILL.md is the main instruction file. For the weekly update, the skill might tell the agent which notes to read, how to group progress and blockers, how to check claims and where to save the draft.

Compare the task with and without the skill. Keep the version that helps and update it when a recurring failure reveals a gap. Read downloaded skills before using them, including any scripts and access they require. Our guide to AI skills explains the format and how to assess whether a skill improves the work.

• Pulsar Learn, what is an AI skill?
• Agent Skills specification

Model Context Protocol, or MCP, is a standard way to connect an AI application to tools and information. You might use a connection to read the approved project notes where they already live. Add the connections required for the workflow and set appropriate permissions. Availability depends on your product, plan and administrator settings.

Hooks run at defined points in an agent's workflow. Code-based hooks can perform a fixed check, such as blocking a proposed send action unless an approval record is present. Some products also offer hooks that ask a model to make a judgement, so not every hook is deterministic. A before-action check can stop an action where supported. A check after the action cannot undo what has already happened.

The weekly-update example can begin with read access and a saved draft. If sending is added later, a human owner can approve the message and a configured check can validate that approval before the tool sends it. The team implementing that check must test it, including the case where approval is missing or refers to an earlier draft.

• Model Context Protocol introduction
• Claude Code guide to hooks

Check the whole job before expanding it

Measure the job from start to finish. Include preparing the inputs, reviewing the output, correcting mistakes and maintaining the workflow. Look at factual accuracy and usefulness alongside time and cost. Try normal work and awkward examples, such as missing notes, conflicting dates or an update with no progress to report. Keep failed examples so you can tell whether a later change fixes them.

Microsoft's adoption guidance combines practical use cases with training, success measures, champions and peer support. OpenAI's accounts of enterprise adoption similarly emphasise literacy, workflow design and evaluation. They do not establish that every organisation should follow the same sequence or expect a particular saving.

At Pulsar, our social-media workflow keeps the content brief, brand rules, source material, examples and checks in project files. The agent prepares the content and an independent reviewer scores it. Tom reviews the exact revision before anything is published. A correction can then improve the shared instructions or the checker for the next piece. This is an example of a bounded workflow, not a claim about measured productivity gains.

• Microsoft's adoption guidance
• OpenAI, how enterprises are scaling AI
• Pulsar Learn, how to keep AI-generated code maintainable

Coordinate within the team, then across teams

When a workflow is useful, show the team how it works and let them try it with their own tasks. Keep a small register of workflows, owners and status so people can see what is proposed, being tested or ready to reuse. Check that register before building. A named owner should maintain the shared skill, source folder and worked example.

This gives colleagues a way to contribute without creating competing versions. Set aside time for practice and feedback. If the method only works for its creator, find out what they know that has not yet been written down. Update the instructions, then have another person try again.

Next, look at handovers. For a sales-to-delivery workflow, agree what the customer brief must contain, who checks it and what counts as acceptance. Decide who fixes missing information and who maintains the connection. Cross-team work needs the affected owners involved because the consequences extend beyond the person trying the tool.

That progression, from your work to your team to the handovers between teams, is our recommended starting point. It lets the scope grow alongside practical understanding. Our consulting and training work uses the same discovery order of goal, current workflow and pain points before choosing an intervention.

• How Pulsar approaches consulting and training

Keep the operating knowledge when the tool changes

A control plane is the shared set of instructions, workflows and checks that directs the work. For a small team, it can begin as organised files in a shared folder. When version history and coordinated changes become useful, keep those files in a repository, a version-controlled collection of project files. The team should be able to find the current working agreements, methods, examples and decisions.

Keep that business knowledge separate from the setup specific to an AI product. If you move from ChatGPT to Claude, you should be able to take the project knowledge with you. The way instructions load, and the configuration of connections, permissions and hooks, may change. Skills can also have dependencies. Run the same representative tasks again and compare the results before trusting the new setup.

Claude Code's documentation describes importing an existing AGENTS.md file through CLAUDE.md. The Agent Skills specification provides a shared format for supporting products. These make reuse easier while leaving product-specific setup and testing necessary.

• Claude Code on importing AGENTS.md instructions
• Agent Skills overview and supporting ecosystem

Give it one proper first assignment

Pick one task you can check today. Put the current sources and a good example in the project, write the working agreement and ask for a draft. For the weekly update, that could be the following brief.

Check the result against the sources, keep the instructions that helped and log the workflow where your team can find it.