An AI implementation roadmap is the order in which a leadership team turns AI experiments into measurable business value: assess readiness, choose a few high-value use cases, prepare the data, decide whether to build, buy or partner, pilot against a baseline, then scale with clear ownership and governance. A realistic frame is 6–12 months from the first assessment to several use cases in production. The companies that get value do not run the most pilots; they focus on a few core processes and measure every one.
This guide covers the portfolio and the organisation. To take a single agent from pilot to production, see our guide to AI agents for business; we do not repeat its production checklist here.
The AI value gap in numbers
Almost every company uses AI somewhere. Far fewer can show what it returns.
BCG, 1,000 executives in 59 countries. Only 26% can move beyond proofs of concept.
McKinsey, State of AI 2025, around 2,000 respondents.
McKinsey. In any single function, no more than 10% report scaling agents.
MIT NANDA, 2025, as widely reported, on a narrow definition.
Read the MIT figure with care: it counts a pilot as a failure if it showed no measurable revenue or cost benefit within about six months, which says as much about how pilots are scoped as about the technology. The broader BCG and McKinsey numbers point the same way: adoption is high, scaled value is rare. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027.
What separates the minority that gets value is focus. BCG's leaders (4% at the cutting edge, 22% beginning to realise substantial gains) pursue about half as many AI opportunities as other companies and concentrate on core business processes, where they generate 62% of their AI value.
Many pilots without owners produce slides. A few use cases in core processes, each with an owner, a baseline and a go/no-go decision, produce value.
Step 1: Run an AI readiness assessment
An AI readiness assessment answers one question: which use cases can we deliver in the next six months, and what must be fixed first? Score six dimensions per business area, because finance may be ready while operations is not.
| Dimension | Ready (2 points) | Partly (1 point) | Not yet (0 points) |
|---|---|---|---|
| Strategy and sponsorship | An executive sponsor; two or three written goals AI should move | Interest at the top, but vague goals or owned by IT alone | No sponsor; driven by enthusiasm or vendor pitches |
| Data access and quality | Data for target processes is known, reachable and reasonably clean | Spread across tools, exports and spreadsheets; quality unknown | On paper, in personal inboxes or in systems nobody can query |
| Systems and APIs | Core systems have documented APIs and test environments | Some APIs; older systems need exports or workarounds | No APIs, no test environments |
| Security and compliance | Written rules on which data may reach which AI provider; security joins early | IT policies exist but do not cover AI; reviews happen late | No AI policy; staff paste data into public tools |
| Skills and ownership | Named business and technical owners; staff have time to test | Engineering capacity but no business owner, or the reverse | AI is everyone's side project |
| Budget and measurement | Budget for pilot and rollout; baselines recorded | Pilot budget only; baselines are estimates | No budget; nobody knows what the process costs today |
How to score yourself:
- Score with the people who run the process, plus IT and security.
- Score per business area, 0, 1 or 2 per dimension, rather than one score for the company.
- Write down the evidence (an API document, a baseline, a named owner) so you can repeat the assessment later.
- Add up the total (maximum 12). At 10–12, move to a pilot. At 6–9, pilot the strongest use case while fixing the weakest dimension. At 0–5, fix sponsorship, data access and policy first; a pilot now would measure your gaps, not the AI.
- Treat a zero on data access or security as a blocker, whatever the total.
For an outside view and a written roadmap, AI feasibility and readiness assessments are part of our software consulting.
Step 2: Choose use cases with a value-feasibility matrix
Collect use-case ideas from every department and score each on two axes. Value: the hours, cost, revenue or risk it moves, and how central the process is. Feasibility: whether data and systems are reachable, how risky the task is, and the readiness score for that area.
Start here
Your first wave: one to three use cases, ideally in core processes, taken to a pilot with a measured baseline.
Prepare
Strategic bets. Fix the data or integration gap first, then schedule them for months 4–12.
Buy or bundle
Use the AI features in tools you already pay for. Not worth custom engineering.
Drop
Park them openly. Each one kept alive takes attention from the first quadrant.
Typical use cases by department:
| Department | Example use cases | Value signal | Feasibility check |
|---|---|---|---|
| Customer support | Ticket triage, drafted replies, resolving known issues | Ticket volume, handling time | Helpdesk API, knowledge base quality |
| Finance | Invoice extraction and matching, expense coding, month-end preparation | Minutes per invoice, errors | ERP access, clear approval rules |
| Sales operations | CRM updates from calls and emails, account research, proposal drafts | Rep time on admin, data completeness | CRM API, email and calendar permissions |
| Operations | Order and supplier follow-up, scheduling exceptions, document checks | Cycle time, manual touches per order | ERP or purchasing system, supplier data |
| HR administration | Policy and leave questions, onboarding paperwork, letter drafts | Internal requests, response time | Personal data rules; no AI decisions about people |
| Legal and compliance | Contract review against a playbook, onboarding file checks | Review time, missed clauses | Written rule set, expert sign-off |
| Product and engineering | AI features in your product, coding agents, test generation | Revenue, retention, lead time | Product data, evaluation, review discipline |
Follow the BCG pattern: fewer, deeper bets in core processes. Two or three first-wave use cases is plenty, each with a business owner before it gets a budget. For the engineering row, see our article on AI coding agents in software projects.
Step 3: Get your data and integrations ready
Many AI projects stall here, not at the model. For each first-wave use case, answer five questions before the pilot starts:
- What data does the task need? What a person uses today, plus the outcomes that show whether the work was done right.
- Where does it live? Which systems and owners, and whether it is reachable through an API or warehouse, or only by export.
- Who may see it? The AI should never see more than the user it works for. Decide which data may reach which model provider, under which contract and retention settings.
- Is it good enough? Sample for gaps, duplicates and outdated documents. Poor data does not stop a demo; it stops production.
- How will the AI retrieve it? Most business use cases need retrieval over internal documents, not a custom-trained model. See advanced RAG and RAG vs fine-tuning.
Integration is usually most of the build. The Model Context Protocol (MCP), an open standard introduced by Anthropic in November 2024 and since widely adopted, lets you wrap a system once and reuse the connection across models and use cases. That reuse makes the second use case cheaper than the first.
If an external team will see personal data, plan for it. Kosovo, where our engineers work, is outside the EU and has no adequacy decision, so remote access to personal data is a transfer that needs a data processing agreement, the Standard Contractual Clauses and a transfer impact assessment. Our GDPR outsourcing guide lists the steps.
Step 4: Build, buy or partner
Three ways to get AI into a process; most portfolios use all three.
| SaaS AI features | Platform + configuration | Custom build | |
|---|---|---|---|
| What it is | AI inside tools you already use (helpdesk, CRM, office suite) | An AI or automation platform configured for your systems | Your own application on model APIs, with your integrations and rules |
| When it fits | Standard tasks inside one tool | Several systems, common patterns, moderate customisation | Core processes where the value is in your data and rules |
| Cost profile | Per-seat or usage licences | Platform fees plus configuration | Engineering build plus model and hosting costs |
| Control | Low: the vendor decides | Medium: within the platform's limits | High: you own the code and choose providers |
| Main risk | Relabelled features that do less than promised | Platform limits, lock-in | Build and upkeep effort for a use case not worth it |
Test vendor claims on your own cases. Gartner warned of "agent washing" and estimated that only about 130 of the thousands of vendors claiming agentic AI offer real agentic capabilities. Custom work usually means generative AI development on established models, connected to the systems from Step 3. "Partner" cuts across both build options, and a good partner tells you when a licence is the better answer.
How to choose an AI implementation partner
- Do you start with a readiness assessment, or with the tool you already sell?
- How will we measure success, and what is the pilot's go/no-go criterion?
- Who builds the evaluation set, from which of our real cases?
- How do you integrate our older systems, and who owns those connections afterwards?
- Which model providers do you use, and where does our data go?
- Is the pilot fixed price, and what exactly is in scope?
- Who owns the code, prompts and evaluation sets, and does the work happen in our repositories and cloud accounts?
Our answers: AI pilots are fixed price with an agreed evaluation set and a go/no-go decision; you get a proposal within 24 hours of the first call; NDA and IP assignment are signed before work starts; we work in your repositories and cloud accounts, and you own the code. Engineers start about 48 hours after you approve their profiles, with a two-week trial you don't pay for if you're not satisfied, from Prishtina on CET/CEST, the same hours as Berlin and Paris.
Start with an AI readiness assessment
We score your readiness per business area, prioritise use cases on value and feasibility, and hand you a costed roadmap with the first pilot defined. You get a proposal within 24 hours of the first call.
Step 5: Pilot with a measurable outcome
A pilot is an experiment designed to produce a decision. If nobody defines the measurable benefit before it starts, nobody can show it afterwards. Write a one-page charter first.
Pilot charter
- Owner: one task, one business owner who decides go or no-go.
- Success metric: one number the owner cares about, such as minutes per invoice or share of tickets resolved without a person.
- Baseline: that number measured today, on real volume.
- Evaluation set: a few hundred real cases with known outcomes, awkward ones included, run on every change.
- Guardrails: read-only or sandbox access; a person approves anything irreversible.
- Timebox and budget: typically three to four weeks, fixed in advance.
- Go/no-go criterion: the quality, cost-per-task and risk targets that earn a production budget.
Measure, don't estimate: in a 2025 study by METR, which METR now describes as out of date, how much faster developers felt with AI differed from what was measured. Perceived and measured gains can diverge, so run the pilot in shadow mode and compare against the baseline.
For agents that act in your systems, the production controls are in the agents guide linked above; our AI agent development page describes our fixed-price pilots. A "no" at go/no-go is a result: you spent a pilot budget instead of a production budget, and you know why.
Step 6: Scale and govern
Scaling is where most value is, and where AI transformation programmes lose momentum. McKinsey's gap between experimenting and scaling is largely organisational. Six building blocks close it.
Operating model
A small central team owns shared foundations (integrations, logging, evaluation, provider contracts); business areas own their use cases.
Ownership
Every live use case has a business owner for the metric and a technical owner for operations.
Governance
An inventory of AI use cases, a light approval step for new ones, versioned prompts and models, a quarterly portfolio review.
Security
Least-privilege access, approvals before irreversible actions, logging and defences against prompt injection.
Regulation
Classify each use case under the EU AI Act and data protection rules before scoping, not after launch.
Change and training
Role-based training, clear word on what changes for whom, and adoption measured like any other metric.
The OWASP Top 10 for LLM Applications lists prompt injection as the top risk (LLM01), so treat every input as untrusted and enforce limits in code, not in the prompt. The EU AI Act places extra obligations on AI used in high-risk areas such as credit, employment and education; its timelines have been changing, so check current dates and your classification with counsel.
Give change management its own budget line: involve the team in building the evaluation set, show them the numbers, and be explicit about what happens to the time saved.
Measuring ROI and what AI implementation costs
A simple ROI formula
ROI = (value delivered − total cost) ÷ total cost
Value = hours saved × loaded hourly cost + other measurable gains (fewer errors, faster cash collection, added revenue)
Total cost = one-off build (pilot + rollout) + run cost (model usage, hosting, monitoring, upkeep)
Count hours net of review time, and only hours that are redeployed or avoid a cost.
An illustrative worked example
Illustrative figures for invoice matching in a finance team; replace them with your own baseline and internal cost per hour.
| Item (illustrative) | Figure |
|---|---|
| Invoices per month | 3,000 |
| Time saved per invoice, net of review | 4 minutes, so 200 hours per month |
| Loaded internal cost per hour (assumption) | €50 |
| Value per month | 200 × €50 = €10,000 |
| One-off cost: pilot €15,000 + rollout €75,000 | €90,000 |
| Run cost per month (model usage, hosting, upkeep) | €2,500 |
| Net benefit per month | €10,000 − €2,500 = €7,500 |
| Payback | €90,000 ÷ €7,500 = 12 months |
| ROI over two years | (€240,000 − €150,000) ÷ €150,000 = 60% |
The portfolio effect matters more: the second use case reuses the integrations, logging and evaluation harness of the first, so it costs much less to build while its value does not shrink.
What to track every month
Coverage (share of eligible volume handled), quality (evaluation score, error and escalation rates), value against the baseline, cost per completed task against budget, and adoption and incidents.
What AI implementation costs
Illustrative ranges at our published rates of €45–65 per engineer-hour, assuming 40-hour weeks and 8-hour days. They are planning examples, not quotes; licences and model usage come on top.
| Phase (illustrative) | Team and duration | Hours | Illustrative cost |
|---|---|---|---|
| Readiness assessment and roadmap: scorecard, use-case matrix, data and systems review, costed plan | 1 engineer-consultant × 2 weeks | 80 h | €3,600–5,200 |
| Pilot: one use case, one or two systems, evaluation set, go/no-go | 2 engineers × 3–4 weeks | 240–320 h | €10,800–20,800 |
| Production rollout: live integrations, approvals, logging, monitoring, training, handover | 3 engineers × 10–12 weeks | 1,200–1,440 h | €54,000–93,600 |
| Next use case on shared foundations | 2 engineers × 4–6 weeks | 320–480 h | €14,400–31,200 |
| Ongoing care: evaluation runs, prompt and model updates | 1 engineer × 2–4 days a month | 16–32 h per month | €720–2,080 per month |
AI pilots are fixed price; squads start from €15k a month. For what drives AI budgets, see the AI development cost guide, or get a range for your scope from the cost estimator.
A 6–12 month AI implementation roadmap
The six steps as a plan a leadership team approves phase by phase; each phase ends in a decision.
| Months | Focus | What happens | Decision at the end |
|---|---|---|---|
| Month 1 | Assess and choose | Readiness scorecard, use-case matrix, data and systems review, baselines, AI policy draft, sponsor and owners named | Which one to three use cases go to pilot? |
| Months 2–3 | Pilot | Pilot charters, evaluation sets, builds on real systems in read-only or sandbox mode, shadow runs, security and data review | Go/no-go per pilot against its baseline |
| Months 4–6 | First production | Rollout to a slice of volume, then wider; monitoring, runbooks, training; shared foundations; next use case prepared | Widen, improve or stop; confirm the second wave |
| Months 7–12 | Scale and govern | Second and third use cases on shared foundations, inventory and governance, monthly ROI reporting, quarterly portfolio review | What to expand, retire and plan for next year |
Smaller companies can compress this. The order should not change: readiness before pilots, pilots with baselines before production, owners before scale.
Frequently asked questions
What is an AI implementation roadmap?
A phased plan for bringing AI into the business so it produces measurable value: readiness assessment, prioritised use cases, data work, a build-buy-partner decision, pilots with baselines, then scaling with ownership and governance.
How long does AI implementation take?
A realistic frame is 6–12 months: about a month to assess and choose, two months of pilots, three months for the first rollout, then scaling on shared foundations.
How much does AI implementation cost?
Illustratively, at our rates of €45–65 per engineer-hour: readiness assessment and roadmap €3,600–5,200, pilot €10,800–20,800, production rollout €54,000–93,600. Licences and model usage come on top.
What does an AI readiness assessment cover?
Six dimensions scored per business area: strategy and sponsorship, data, systems and APIs, security and compliance, skills and ownership, budget and measurement. It shows what you can pilot now and what to fix first.
How do you measure the ROI of AI?
ROI = (value − total cost) ÷ total cost. Value is hours saved, net of review time, times your loaded hourly cost, plus other measurable gains; cost is build plus run. Record the baseline before the pilot.
Should we build, buy or work with a partner?
Buy for standard tasks inside existing tools, configure a platform for common patterns, build custom where the value is in your own data and rules. Choose a partner that measures against baselines and leaves you owning the code.
Turn your AI plans into a costed roadmap
Tell us your goals and the processes you have in mind. We assess readiness, choose the first use cases with you and propose a fixed-price pilot with an evaluation set and a go/no-go criterion.
Sources
- BCG, press release, 24 October 2024: Where's the Value in AI? 74% of companies struggle to achieve and scale value
- McKinsey, November 2025: The state of AI in 2025: Agents, innovation, and transformation
- MIT NANDA, The GenAI Divide: State of AI in Business 2025, as reported by Virtualization Review: MIT report finds most AI business investments fail, reveals GenAI divide
- Gartner, press release, 25 June 2025: Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027
- METR, 10 July 2025: Measuring the impact of early-2025 AI on experienced open-source developer productivity
- OWASP: Top 10 for LLM Applications (2025)
- Model Context Protocol: modelcontextprotocol.io
