Field Notes / AI
AIDraft · being revised
The AI adoption maturity model.
Two offset tracks, five levels each, and a twelve-question assessment. Where your people and your organisation actually are, and what to do next.
This model is built for small and mid-sized companies of 30 to 250 people (up to around 500, if the back office is 120 people or fewer). Enterprise frameworks assume a dedicated AI team, a centre of excellence and a large budget. Most companies this size have none of those, and still need a way to see where they are and what to do next.
The core idea is that AI adoption runs on two tracks. The individual track is how people adopt AI in their own work. It moves fast, the barrier is culture and skill, and the cost is modest. The organisational track is how the company works with AI together: first in shared workflows, then on a common AI infrastructure that connects its systems and data, and in the end in what it sells. It moves slowly, because each level has prerequisites that compound. Each track has five levels, and each level is described through four requirement areas: culture, process and data, technology, and risk and governance.
The tracks are offset. The organisational track does not start until most people work with AI at individual level two. People have to believe in the tool before they will share a way of working with it, let alone run a process through it. Read the overview across: each organisational level needs the individual level beside it, broadly at the start and at least among the people who own the work further up.
The reasoning behind the model is in the companion note, AI capability is built, not installed.
| Individual track | Organisational track |
|---|---|
| 01Casual ExplorerLow | Not yet. The organisational track starts when most people are Deep Workers. |
| 02Deep WorkerModerate | 01Shared AI WorkflowsLow-Moderate |
| 03Connected OperatorSignificant | 02Standalone Task AgentsModerate |
| 04Tool BuilderHigh | 03Common AI InfrastructureSignificant |
| 05Agent ArchitectHigh | 04Autonomous AgentsHigh |
| Beyond any one person. From here the gains belong to the organisation. | 05Agentic OrganisationTransformative |
Where are your people, and where is your organisation? Twelve questions, a score on each track and what to do next.
The individual track
How people adopt AI in their own work. The biggest gains come at levels three and four, when people connect AI to their tools or build their own.
Casual Explorer
Uses online chat tools (ChatGPT, Claude, Gemini) for quick queries, replacing Google searches, getting quick answers, brainstorming ideas.
LowCultureCuriosity. Willingness to try.
Process & dataNone. Ad-hoc personal use.
TechnologyAccess to a chat interface. Browser or app.
Risk & governanceMinimal. General awareness that company data shouldn't go into unknown tools.
Value · LowMinor time savings on lookups and drafting. Nice-to-have, not transformative.
CostNear-zero. Free tiers or basic subscriptions.
Examples- Quick fact lookups
- Brainstorming names or ideas
- Grammar and translation help
- Explaining concepts
Deep Worker
Sends significant files and context. Spends 5-10 minutes crafting a single prompt. Uses AI as a thinking partner for substantive work.
ModerateCultureTrust in AI quality. Investment mindset toward prompting.
Process & dataStructured personal workflows. Saved prompts and templates.
TechnologyPro-tier subscriptions. File upload capability. Long context windows.
Risk & governanceData classification becomes critical. Which tools are approved for customer data, financial data, personal data (GDPR)? Clear go/no-go list required.
Value · ModerateMeaningful time savings on analysis, drafting, and review. Hours per week reclaimed.
CostSubscriptions (€20-50/user/month) + training investment for prompt craft.
Examples- Drafting strategy documents
- Analyzing datasets
- Reviewing and editing contracts
- Building presentations from raw input
Connected Operator
Connects AI to personal productivity stack, mail, calendar, tasks. Gets proactive suggestions and has actual work executed.
SignificantCultureComfort delegating to AI. Willingness to share personal data with AI systems.
Process & dataDefined personal routines. Clear boundaries on what AI can and cannot do.
TechnologyAI-integrated productivity tools (Copilot, Claude integrations, Zapier AI). API access or connected apps.
Risk & governanceGDPR and confidentiality review of connected tools. What data flows where? DPA agreements with AI vendors. Clear policy on AI access to email, calendar, CRM.
Value · SignificantStep change. AI handles routine work autonomously. Individual capacity expands materially.
CostHigher-tier subscriptions (€30-75/user/month) + integration tools + targeted training.
Examples- AI-drafted email replies
- Calendar optimisation suggestions
- Automated meeting prep briefs
- Task prioritisation from inbox
Tool Builder
Especially in ops, finance, and analytics, uses agents (Claude, Cursor, Replit) to build custom tools for their own work.
HighCultureBuilder mindset. Comfort with iteration and imperfect first versions.
Process & dataBasic version control habits. Documentation of what was built and why.
TechnologyAccess to coding agents (Cursor, Claude Code, Replit Agent). Internal data access. Basic deployment capability.
Risk & governanceReview of what internal data custom tools access. Basic code review practices. Who maintains tools when the builder moves on?
Value · HighMultiplier effect. One person builds tooling that saves hours weekly, for themselves and often their team.
CostDeveloper tool subscriptions + training + some internal data access setup. Still modest vs. org track.
Examples- Custom reporting dashboards
- Data transformation scripts
- Automated reconciliation tools
- Personal workflow automations
Agent Architect
Designs and deploys actual agent systems, multi-step, persistent, acting semi-autonomously across tools and data sources.
HighCultureSystems thinking. High comfort with autonomous AI. Able to define clear guardrails.
Process & dataDefined testing and validation routines. Monitoring of agent outputs. Fallback procedures.
TechnologyAgent frameworks (LangGraph, CrewAI, custom). MCP connections. Orchestration tooling.
Risk & governanceAgent action boundaries clearly defined. What can agents do autonomously vs. what requires human approval? Audit trail for agent actions.
Value · HighPersistent autonomous capacity, like a junior analyst running 24/7 for one person or team. The gain stays local, and it is a key-person risk until the organisation owns what was built.
CostPlatform costs + API usage + training. Higher than earlier levels but still individual-scale.
Examples- Research agents that monitor and summarise industry news
- Financial monitoring agents with alert triggers
- Multi-step data pipeline agents
- Personal CRM automation agents
The organisational track
How the company works with AI together. It starts when most people are Deep Workers, with teams sharing how they use AI. Level three is the turning point: a common AI infrastructure where systems are connected, skills are live and work runs under one set of rules. Everything above it is built on that, and it is where most SMEs stall.
Shared AI Workflows
Teams share prompts, skills and instructions, and work from shared folders or a Git repository. Common AI workflows, still run in each person’s own tools.
Low-ModerateCultureMost people already at individual level two, and they believe the tool helps. Sharing a good prompt or skill is normal, not showing off. Teams agree to do recurring work the same way.
Process & dataShared libraries of prompts, skills and templates for recurring team tasks. A shared folder or Git repository holds the instructions, templates and reference material. Team routines written down: notes and actions, handovers, first drafts, reviews.
TechnologyTeam or business plans with shared projects or workspaces. Skills or custom instructions that can be distributed. A shared drive or Git repository. No integration into core systems yet.
Risk & governanceData rules apply to shared context, not just to single chats. A named owner per library who reviews and retires what is shared. Access control on shared workspaces.
Value · Low-ModerateIndividual wins become team practice. Output gets more consistent, new people get up to speed faster, and the team’s way of working gets written down, which is the start of a defined process.
Investment · Low-ModerateTeam licences, a few hours a month to curate the library, and one named owner per team. No change to core systems.
Examples- A shared skill for customer proposals in house style
- A team prompt for month-end commentary
- Meeting notes and actions from one common template
- A Git repository with the team’s instructions, templates and reference documents
Standalone Task Agents
Small, contained tasks handled by AI agents, classifying, routing, prioritising. Standalone with limited integration.
ModerateCultureWillingness to trust AI on low-stakes decisions. Defined ownership of agent outputs.
Process & dataClear task definitions. Input/output specifications. Basic quality monitoring.
TechnologyAgent or automation platform. API connections to relevant data source. Simple evaluation setup.
Risk & governanceA named owner for each agent. Acceptable error rate agreed before go-live. Basic output quality monitoring. Human review of edge cases.
Value · ModerateRemoves specific manual tasks. Useful but isolated, doesn't change how work flows.
Investment · Low-ModerateCan often be built in days. Narrow scope means limited integration work.
Examples- Support ticket prioritisation
- Lead scoring and routing
- Document classification
- Expense categorisation
Common AI Infrastructure
One governed environment the whole company works in. Core systems and data are connected, shared skills are live, and access, data rules and logs are set in one place.
SignificantCultureAI is treated as shared company infrastructure, not a collection of personal subscriptions. People work in the common environment by default, because it is better than the private tab. Builders in-house who can connect systems and write skills (individual level four).
Process & dataA decision on which systems and data to connect, and at what level of access, read-only first. The skills and templates from level one move into the common environment, each with an owner. Connected sources are documented and clean enough to be trusted.
TechnologyFor some, a business or enterprise plan (Claude, ChatGPT) with approved connectors and shared skills switched on. For others, their own setup: a gateway such as LiteLLM in front of the models they choose, with connectors (MCP) to CRM, ERP, documents and the data warehouse. Chat, agent and developer tools all behind it.
Risk & governancePolicy becomes a setting: model and connector allow-lists per team, read-only access where possible, guardrails for personal and confidential data, spend caps, and one log of who did what. A named owner for the environment, and someone who reads the log.
Value · SignificantThe enabler. Every person and every agent works with the company’s own data and skills, safely and under the same rules. Individual wins stop being local, and the next agent is built on what is already connected instead of from scratch.
Investment · Moderate-SignificantFrom a business plan with connectors and skills switched on, to your own gateway and setup. The building blocks are proven. The work is deciding what to connect, who can reach it, and who owns it. This is where most SMEs stall.
Examples- A company Claude workspace with CRM, documents and shared skills connected
- A LiteLLM gateway with per-team keys, budgets and logs in front of chosen models
- Governed, read-only access to the data warehouse for analytics
- Proposal, month-end and contract-review skills available to everyone
Autonomous Agents
Fully autonomous agents living in the tech stack. Take real actions, handle multi-step processes with minimal human oversight.
HighCultureHigh trust in AI. Mature accountability structures. Organisation-wide AI literacy. Comfort with AI making consequential decisions.
Process & dataHigh process maturity across functions. End-to-end workflow documentation. Automated quality assurance for agent actions. Continuous monitoring and alerting.
TechnologyBuilt on the common AI infrastructure from level three. Agent orchestration platform. Automated testing infrastructure. Observability and audit trails. Rollback capabilities. Production-grade deployment.
Risk & governanceComprehensive benchmark suites. Automated testing pipeline, agents cannot deploy without passing precision/confidence thresholds. Continuous monitoring of live agent decisions. Clear escalation and rollback triggers. Audit trails for all autonomous actions.
Value · HighEntire process areas run autonomously. Headcount implications are substantial. Organisation operates at a fundamentally different capacity.
Investment · HighRequires the common infrastructure from level three, plus defined processes, automated testing, monitoring and governance layers.
Examples- Autonomous customer service resolution
- Self-managing supply chain optimisation
- Automated financial close processes
- Proactive risk detection and response
Agentic Organisation
Work is redesigned around agents, and agentic capabilities are built into the business model: what you offer, how you price it and who you can serve. Agents drive revenue, not just cost.
TransformativeCultureLeadership asks two questions of every plan: can an agent do this work, and what can we now offer that we could not before? People are employed for judgment, relationships and oversight. The company thinks in agent capacity alongside headcount.
Process & dataWorkflows are designed around agents from the start, with human checkpoints placed on purpose. Your own data and process knowledge are treated as the asset the agents run on, and the thing competitors cannot copy.
TechnologyAgent orchestration across the stack. Agents that act towards customers, not only internally. Infrastructure to deploy, measure and retire agents quickly. Feedback loops from live use back into improvement.
Risk & governanceAccountability for agent decisions at every level, including those that reach customers. Automated compliance monitoring. Agent-driven revenue and risk reported to the board like any other business line.
Value · TransformativeBeyond efficiency. Offers, service levels or pricing that were not economic with people doing the work become possible. The advantage is hard to copy, because it rests on your data, processes and people, not on the model.
Investment · TransformativeA business model decision, not a technology project. It changes strategy, organisation, hiring and budgeting, and it can only be built on the levels below.
Examples- A service that only paid off for large customers, now offered to small ones
- Pricing by outcome or uptime instead of by hour
- Always-on service levels no staffed team could deliver
- New market entry scoped as agent deployment, not a hiring plan
The organisational track rests on the individual one
The two tracks are not independent. A company cannot share AI workflows its people do not use, or run a process through a tool its process owner does not trust. When the organisational track gets ahead of the people, AI initiatives turn into theatre: pilots that look good in a steering group and are quietly worked around on the floor.
The dependency runs one way. The individual track can and should move ahead without waiting. The organisational track cannot move ahead of the people. As a rule of thumb:
- Level one, shared workflows, needs most people at Deep Worker.
- Level two, task agents, needs the owners of those tasks to be Connected Operators.
- Level three, common AI infrastructure, needs builders in-house who can connect systems and write the skills that run on it.
- Level four, autonomous agents, needs people in-house who can design and run agent systems.
The assessment flags it when your organisational ambition is ahead of your people.
Process-integrated AI runs on the side
Earlier versions of this model had AI embedded in business processes as level three: suggested replies in the support tool, auto-generated reports, reorder suggestions. I no longer think it is a step on the track. It is AI used as part of the classical automation toolkit, alongside rules, integrations and RPA.
That can be a strong case on its own. A support team of eight might work like twelve, and a five-day close might take two. Where the case is there, do it, and judge it like any other automation project: a defined process, an owner, a quality gate and a baseline.
But it does not move you towards an agentic organisation. Each integration is built for one process in one system, and the capability stays there. It does not give people a shared place to work, connect the company’s data, or make the next agent easier to build. Common AI infrastructure does. That is why it, and not process integration, is level three.
What leadership must provide
On the individual track, the job is permission, resources and an honest read on whether it helps:
- Clear go / no-go guidelines (do first). Explicit rules on data classification: what types of data can go into which tools, which tools are approved for sensitive data (GDPR, confidential), and where it’s free-for-all. Removes the biggest adoption blocker, people not knowing what’s allowed.
- Training that meets people where they are (ongoing). Upskilling matched to current level and ambition. A finance controller at level one needs different training than a data analyst ready for level four. One-size-fits-all workshops waste time and money.
- Budget and tool access (do first). Clearly available budget for pro-tier subscriptions or company accounts. Nothing kills momentum faster than making people expense a €20 tool through procurement.
- Measure what people experience (quarterly). Individual impact is hard to put one number on, so ask. In a short pulse: how much does AI help you in your work, and what share of your work is AI-supported? Add one open question: name a task you now do differently. The first two show direction, the third shows whether it is real. High usage with low perceived help means the training is not landing on real work.
On the organisational track, the job is ownership and gates:
- One environment, one set of rules (do first). Decide where company AI work happens, what it can reach and who owns it. A business plan with connectors and skills is enough for some; others need their own gateway and setup. Either way, make the governed way the easy way.
- One owner per process. A named person owns the process, its data and the AI running in it, and decides changes. Without that owner the AI component belongs to nobody, and it degrades the first time the model or the process changes.
- Quality gates before go-live. Agree before the pilot starts what good looks like: a test set of real cases, the acceptable error rate, the confidence threshold for acting without review, and what happens when it is wrong. Run the test set again whenever the model, the prompt or the process changes.
- Measure impact, not activity. Pick the business measure before building: cycle time, error rate, cost per case, capacity per person. Take a baseline. The number of agents, prompts or users is activity, and says nothing about value.
- Decision rights for scaling. Who decides that a pilot goes into production, gets more autonomy or is switched off. Written down, so success is not decided by enthusiasm.
Run two workstreams
Push individual adoption hard and fast. Do not wait for the organisational track: your best operations person at individual level four may deliver more this year than a half-built task agent. Invest separately and deliberately in the organisational foundation, and let it follow the people, not lead them.
At organisational level five, the question moves from cost to revenue. Work is redesigned around agents, and what agents make possible shows up in the offer: services that were not economic before, pricing by outcome, customers you could not serve. That advantage is hard to copy, because it rests on the data, processes and people built up through the levels below. A small company with defined processes and fluent people can get there faster than a large one.
Assessment: where are you?
Twelve questions. A score on each track, a check of whether your organisational ambition is ahead of your people, and what to do next. Nothing is sent anywhere.