← All insights
Best PracticesNovember 12, 20265 min read

AI Implementation Roadmap: Five Phases From Pilot to Scale

By · AI agents for European teams

The team behind AgentWorks, the AI workspace for European organisations.

Reviewed November 12, 2026

Share
AI Implementation Roadmap: Five Phases From Pilot to Scale

Team arranging an AI implementation roadmap

An AI implementation roadmap is a phased plan with owners, timelines and approval gates. It takes you from a scored list of use cases to one governed pilot, then to live workflows and a portfolio you can scale. The roadmap that works starts with one recurring workflow, keeps a human approving every consequential action, and settles data residency and compliance before the first run, not after it. Below you will find the five phases, how long each usually takes, what to decide at each gate, and the EU rules that shape the plan.

TL;DR

  • Gartner defines an AI roadmap as a plan of activities, timelines and responsibilities, measured against adoption, business value and foundational maturity.
  • Run five phases with four gates: Discover, Define, Build, Deploy, Scale. Do not move on until the gate is signed off.
  • Score use cases on volume, repetitiveness, data availability, error tolerance and cost. Start with one workflow that is frequent and rules-based.
  • Settle GDPR, the EU AI Act, data residency and human approval rules before the pilot, because retrofitting them is slower than building them in.
  • Budget for the run, not just the build: usage-based pricing makes cost per workflow run the number to watch.

Table of Contents

What an AI implementation roadmap is, and what it should deliver

Gartner describes an AI roadmap as a strategic plan that sets out the activities, timelines and responsibilities for implementing and scaling AI. It also says success is tracked against three things: adoption goals, business value delivered, and the maturity of foundations such as strategy, governance, engineering and data.

That definition rules out two common mistakes. A roadmap is not a list of tools to buy, and it is not a one-off project plan. A project plan ends at go-live. A roadmap carries on through monitoring, review and the decision to retire use cases that do not pay back.

For a business team, three outcomes are worth targeting from the start:

  1. Hours returned. Time removed from a named, recurring task, measured against a baseline you record before the pilot.
  2. Error and cycle-time reduction. Fewer corrections, faster turnaround on replies, quotes, reports or onboarding steps.
  3. Control. Every automated action traceable, with a named person accountable for approvals.

Write these down as numbers before you build anything. A pilot without a baseline cannot prove it worked.

The five phases at a glance

Five phases from AI pilot to scale

The AI4EA phase-gated execution model is a clear structure to borrow. It defines five phases with four gates: Discover (2 to 6 weeks), Define (4 to 8 weeks), Build (8 to 16 weeks), Deploy (4 to 8 weeks) and Scale (ongoing). The gates approve the definition, the build, the deployment and the scale-up in turn.

Those durations suit a large organisation building custom systems. A team configuring agents on an existing workspace can compress Define and Build considerably, because there is no model to train and no infrastructure to procure. The gates still matter just as much.

PhaseTypical durationMain questionGate decision
Discover2 to 6 weeksWhere does AI fit our work?Shortlist of scored use cases
Define4 to 8 weeksWhat exactly will the pilot do, and how do we measure it?Approved scope, baseline, owner
Build8 to 16 weeksDoes it work on real data with real approvals?Pilot meets success metric
Deploy4 to 8 weeksCan it run inside live workflows safely?Governance and training in place
ScaleOngoingWhich use cases earn more funding?Portfolio review, retirements

Phase 1: Discover your use cases

Start by breaking work into tasks, not departments. BDO advises splitting each process into individual tasks and matching each to the right capability: rules-based tasks such as data entry, routing and approvals suit plain automation; judgement-based tasks such as analysis and exception handling suit a copilot with a human in the loop; multi-step workflows such as client onboarding or research synthesis suit AI agents with governance and oversight.

Then score. The Australian National AI Centre publishes a scoring matrix for pain points with five criteria, each scored from 1 to 5:

  • Volume: how often the task occurs.
  • Repetitiveness: whether the process is consistent.
  • Data availability: whether digital records exist.
  • Error tolerance: the consequence of a mistake.
  • Current cost: the effort or resources the task consumes.

The same guidance adds a useful rule: if a simpler solution delivers 80% of the benefit at 20% of the cost or complexity, start there.

A worked example

An agency handles a shared inbox with 300 client emails a week. Scoring it: volume 5, repetitiveness 4, data availability 5 (the emails already exist digitally), error tolerance 3 (a wrong reply to a client is awkward but recoverable if a person approves it), current cost 4. Total 21 out of 25. Compare that with drafting bespoke strategy documents: volume 1, repetitiveness 2, error tolerance 2. The inbox wins as the first pilot, and the strategy work stays with your people.

Keep the shortlist to three use cases. You will pilot one or two.

Phase 2: Define scope, owners and readiness

This is where most roadmaps are won or lost, and it is the least glamorous phase. Three jobs belong here.

Check readiness across four domains

The Frontiers Institute identifies four AI readiness domains: Strategy (alignment and governance), Organisation (leadership, capability and capacity), Data (trust, integrity and quality) and Technology (infrastructure, models and tools fit for purpose). For a workflow pilot, ask one plain question per domain:

  • Strategy: which business metric will this move, and who signs it off?
  • Organisation: who owns the outcome, and who will use the output daily?
  • Data: is the information the agent needs in a place and format it can reach, such as a mailbox, a shared drive or a shop back office?
  • Technology: do your existing tools connect, or does the pilot need integration work first?

Name an owner and a baseline

Every use case needs one accountable business owner, not a committee. Record the baseline now: minutes per task, tasks per week, error rate, current cost. Gartner also recommends a workforce plan that covers upskilling, AI literacy and champion roles, so pick one or two champions in the team that will use the output.

Settle the compliance frame

Decide now where data will be processed, which personal data the workflow touches, and which actions require human approval. Section on regulation below sets out what the rules require. Our GDPR and AI agents guide walks through the practical side, and the enterprise AI data security checklist covers the controls to confirm before go-live.

Phase 3: Build a pilot that can fail safely

A pilot has three elements: a narrow scope, a success metric fixed in advance, and a point at which you stop if it is not working. Keep scope to one workflow, one team and one or two integrations. Review at fixed intervals, for example every four weeks, and end the pilot if it misses the metric twice.

Configuring agents on a ready workspace shortens this phase. In AgentWorks you configure agents for recurring tasks and connect tools such as Outlook, Shopify and Google Drive, and agents can run on Claude, GPT, Gemini or Mistral models. For the inbox example above, an email triage assistant and an inbox reply drafter are the kind of templates you would start from, and the shared inbox use case shows the workflow end to end.

During the pilot, track four numbers:

  1. Approval rate: how often the person accepts the agent's draft unchanged.
  2. Edit rate: how often they change it, and how heavily.
  3. Time per task against the baseline.
  4. Cost per workflow run.

The approval and edit rates tell you whether the agent is ready for more autonomy. Do not loosen approvals because the pilot feels good. Loosen them when the numbers say so, and only for low-risk actions.

Phase 4: Deploy into live workflows

Deployment is where a pilot becomes part of someone's working day. Three things decide whether it sticks.

Human approval rules. Decide which agent actions need a person: anything that sends a message, books a meeting or changes a record is the sensible default. In AgentWorks, human-in-the-loop controls require approval for any agent action that sends, books or changes data, and audit logs record what happened. Our piece on human-in-the-loop AI control explains how to set those rules by risk level.

Governance written down. Who can create agents, who can change prompts, who reviews logs and how often. The AI agent governance framework and the guide to prompt governance give you templates for these policies.

The hand-off. A pilot run by one enthusiast does not scale. Move ownership to the team that lives with the workflow, with training and a named contact. We cover this in from pilot to production.

If the roadmap covers a whole company rather than one team, the 90-day plan for a 500-person company and the company-wide adoption playbook give you the rollout sequence. If staff already use unapproved tools, read avoiding shadow AI first: a governed option that is easy to use is the fix.

Phase 4: Deploy into live workflows

Phase 5: Scale as a portfolio

Gartner advises that, over time, organisations should move from one-off projects to a portfolio approach focused on ongoing value creation. In practice, that means a quarterly review of every live use case with three outcomes: fund more, keep as is, or retire.

When you choose what to scale next, favour workflows that repeat across teams. A moderately valuable use case that can be copied to five departments often beats one spectacular use case that cannot be copied. Compare candidates on cost per unit of benefit, for example cost per resolved email or per processed order, rather than on headline savings.

In AgentWorks, one workflow can call several agents, as the one workflow, many agents use case shows, and sector pages such as AI for e-commerce and AI for manufacturing show what a second and third use case often look like.

Budget for the run, not just the build

The most common budgeting error is pricing the setup and forgetting the running cost. With usage-based AI, cost grows with volume: a pilot with 20 users says little about what 2,000 will cost. So model the full year before you commit.

A workable cost model has five lines:

  • Licences: per-person plan fees.
  • Usage: workflow runs and agent actions, multiplied by expected volume.
  • Integration and data preparation: getting information into a usable place.
  • Oversight: the time people spend approving and reviewing.
  • Training and change: onboarding, champions, documentation.

Our 12-month total cost of ownership model walks through each line with a worked calculation.

With AgentWorks, pricing is published in euros. The Free plan costs €0 one-off, the Business plan is €29 per person per month or €290 per person per year, and extra workflow runs are available as one-off top-ups from €5 to €400. That makes the pilot cost simple to forecast: count the people, estimate the runs, and compare against the hours you recorded in the baseline.

The regulation your roadmap has to respect

For a business operating in the EU, or serving EU customers, two bodies of law shape the plan. Neither is a reason to delay. Both are easier to meet when designed in.

The EU AI Act timeline

The European Commission states that the AI Act entered into force on 1 August 2024 and became generally applicable on 2 August 2026. Its main milestones: prohibitions and AI literacy obligations applied from 2 February 2025, governance and general-purpose model obligations from 2 August 2025, and transparency rules from August 2026. The same page records that the AI Omnibus, in force since 27 July 2026, deferred the Annex III high-risk rules to 2 December 2027 and the Annex I rules to 2 August 2028, and extended certain simplified requirements for SMEs to small mid-cap companies.

For your roadmap, two practical points follow. AI literacy is already an obligation, so training belongs in Phase 2 and Phase 4, not as an afterthought. And any use case that might fall under the high-risk categories, such as recruitment or credit decisions, needs classification at the Define gate.

GDPR still applies in full

The French regulator CNIL states in its recommendations on AI system development that where personal data is involved, both the GDPR and the AI Act apply, and that controllers must implement appropriate security measures, including access authorisations, backup management and infrastructure management. In practice: list the personal data each agent touches, record the workflow in your record of processing activities, and carry out a data protection impact assessment before processing where the risk to individuals is likely to be high.

UK-based teams should note that the UK GDPR carries equivalent duties, and that serving EU customers can bring you under the AI Act as well. Take advice from your data protection officer or a solicitor for your specific case.

This is why we run all servers, AI models and integrations in the EU, never use customer data to train models, and build human approval and audit logs into the product. Details sit on our Trust Center.

Where to start this week

If the roadmap feels large, shrink it to a week of work:

  1. List ten recurring tasks and score them on the five criteria.
  2. Pick the top one and record its baseline.
  3. Name an owner and one champion.
  4. Decide which actions need approval.
  5. Set a four-week review with a stop rule.

Perspective

Most AI roadmaps fail on sequencing, not ambition. Teams pick a platform, then look for a use case, then discover the compliance questions in month five. We would reverse that order: governance first, one workflow second, scale third.

Our view is that governance multiplied by throughput is the real measure of an AI programme. Throughput without governance stalls the first time a regulator, a client or a works council asks where the data went. Governance without throughput is a policy document nobody uses.

That is why AgentWorks is built the way it is: EU-only servers, models and integrations, no training on your data, approvals for every agent action that sends, books or changes data, and audit logs behind each run. Combined with usage-based pricing, a team can nail the first workflow in weeks, show the numbers, and earn the right to scale.

Start small, keep a person in the loop, and let the metrics decide what comes next.

Erwin

Start your first governed workflow

If your roadmap's first gate is a single pilot, you can set it up in AgentWorks without a procurement cycle. The Free plan costs €0 one-off, and the Business plan is €29 per person per month, with your agents, integrations, approvals and audit logs in one EU-based workspace.

Agent-works

See the plans and run costs on the AgentWorks pricing page, or look at how a workflow with agents and approvals is structured before you start.

FAQ

How long does an AI implementation roadmap take?

The AI4EA model puts the phases at roughly 18 to 42 weeks from Discover to Deploy, with Scale continuing afterwards. Teams that configure agents on an existing workspace, rather than building systems, can usually run a single-workflow pilot in a fraction of that. Set the timeline per use case, not for the whole programme.

What should the first AI use case be?

Choose a task that is frequent, repetitive, digitally recorded and tolerant of recoverable errors, because those score highest on the National AI Centre's matrix. Shared inbox handling, meeting-note follow-ups and order emails are typical. Keep a person approving outputs from day one.

Do we need to worry about the EU AI Act if we only use AI tools?

Yes, in part. AI literacy obligations have applied since 2 February 2025, and transparency rules apply from August 2026. Classify each use case, confirm whether it touches a high-risk category, and keep records of what your agents do. A workspace with audit logs and approval controls makes that record-keeping simpler.

How do we measure whether the roadmap is working?

Track the three measures Gartner names: adoption, business value and foundational maturity. Concretely, that means active users and approval rates, hours saved and cost per workflow run against your baseline, and whether governance, data and ownership are in place for the next use case.

Sources

Made with BabyLoveGrowth's content platform

About the author

· AI agents for European teams

AgentWorks is the AI workspace for European organisations: chat, agents, workflows and tasks, with every model from OpenAI, Anthropic, Google and Mistral hosted in the EU, approval before an agent acts, and every price in euros.

Read more about AgentWorks