AI integration: don’t start with agents

AI integration runs in five stages: assessment, general productivity, IT technical productivity, then technical debt and process automation, which run in parallel. Ordered by complexity.

Companies can be tempted to start with agents, however that is the most complex stage. General productivity captures the low hanging fruit and onboards the organization, improving the odds of a productive implementation of the later stages.

Stages

Assessment

Assessment of what the business requires and how IT supplies it.

• The business. How the organisation is set up, which core processes run on software, and how the managers of those areas see AI.
• IT. Its people and resources, the software stack, the plan, the technical debt, what is already planned for AI, and what has been attempted before.

The output is the plan revised and sized.

General productivity — app

Enterprise LLM applications are licensed per user and integrated into the workplace suite. As of August 2026 the main products are Claude, ChatGPT Enterprise, Microsoft Copilot and Google Gemini.

Market leaders have deployed across their organisations. Accenture is rolling out to about 743,000 people. JPMorgan reached 200,000 users in eight months. Goldman Sachs went firmwide after piloting with 10,000.

The implementation:

1. Standard training
2. Discovery and development workshops
3. Support applying it to each user’s processes

General. Document review and summary, data extraction, drafting, analysis in Excel, meeting minutes, searching across mail and files, meeting preparation.

Examples: tendering, contract administration, procurement, cost and planning. A long specification turned into a requirements checklist, supplier quote comparison, contract price adjustment.

Advanced. Working in teams on a shared project, deep research, simple automations — scheduled or recurring — and connections to the company’s own applications.

IT technical productivity — coding tools

AI coding tools are LLM applications for software development. As of August 2026 the main tools are Claude Code, Codex, Cursor, GitHub Copilot and Gemini Code Assist.

The tool reaches the code, the command line and the systems themselves. For a technical user that is the working interface to the company’s software: analysis, queries over its data, and automation of IT’s own operational work.

Rolled out to IT, and open to any user with the technical capability for it.

Technical debt — software

Assessment provides IT’s plan, the software each core process runs on, and the processes running on none. Old software and no software are the same debt.

AI coding tools change the cost equation of writing code, testing it and migrating data, and with it the total cost of producing an application. Work that could not justify a project before can now, on either route.

Two delivery routes:

• Outsourced. An off-the-shelf product, adapted or not, or development commissioned from a software company.
• Internally led. IT team with a contracted developer, AI coding tools, in short cycles with the user present.

Process automation — agents

Anthropic, OpenAI, Google, Microsoft, Amazon, xAI and Mistral all have platforms for building agents, and frameworks like LangChain build them without tying to one model. As of August 2026 there are deployments in production.

The default implementation would be an agent running a simple process end to end, deciding and implementing, with a human in the loop managing the exceptions. A different type would be an agent as the interface between people and the systems, capturing data in natural language where the work happens.

Steps:

1. Selection of processes
2. Assessment and documentation
3. Rationalisation
4. Building, testing and deployment