Short answer: A workable AI roadmap fits on one page and runs 12 months in four phases, each ending in a go or stop decision. Pick one workflow and measure its baseline in weeks 0 to 4. Launch it in months 2 to 3, measure it and add a second in months 4 to 6, then grow a small portfolio.
A workable AI implementation roadmap for a company of 20 to 500 people fits on one page. It runs 12 months in four phases, and each phase ends in a go or stop decision.
In weeks 0 to 4, you pick one workflow, measure its baseline, and put an AI policy and staff training in place. Then you build and launch that workflow with named approval points (months 2 to 3), measure it and add a second one with an internal playbook (months 4 to 6), and grow to a small portfolio with governance and a vendor review (months 7 to 12).
What should an AI implementation roadmap contain?
Less than most templates suggest: the first workflow and its owner, the outputs of each phase (things you can point to, not activities), a decision gate per phase with criteria agreed in advance, a budget for the year, and the rules on tools, approvals and training. It does not need 40 use cases, a vendor shortlist or a three-year vision deck.
Why 12 months, when some guides promise results in six weeks? The first workflow can be live within a quarter; the rest of the year is for owning it. RAND’s study of why AI projects fail advises leaders to be prepared to commit a team to one problem for at least a year, and warns that projects on an overly accelerated timeline tend to fail.
Are you ready to start? A five-point AI readiness checklist
Answer each with a plain yes or no.
Owner. Yes if one person who runs the workflow can give it half a day a week for three months and can change the process. No if it is a committee, or IT alone.
Data access. Yes if someone can pull 20 to 50 real past cases from the systems involved, this month, with permission. No if access needs a new contract or an IT project.
One measurable workflow. Yes if it has a start, an end and a number that should move. No if the scope is “AI for marketing.”
Budget. Yes if money is approved for the first build and a year of running costs, plus the owner’s time. No if the plan is to “see what the pilot costs.”
Policy. Yes if staff know which AI tools they may use, which data must never go into them, and who checks output. No if there are no rules, or a ban nobody follows.
Five yeses: start now. A no on owner or data access: fix it first, since Gartner predicts that through 2026, organizations will abandon 60 percent of AI projects unsupported by AI-ready data. A no on policy: write it in week one. In Microsoft and LinkedIn’s 2024 Work Trend Index, 78 percent of AI users brought their own AI tools to work, 80 percent at small and medium-sized companies.
Ideas are rarely the gap. In Bitkom’s 2026 survey of 603 German companies with 20 or more employees, only 11 percent of AI users named missing use cases as an obstacle.
Weeks 0 to 4: what do you decide first?
This phase costs little money and a lot of attention.
Choose the first workflow with the people who do the work. Our guide to choosing your first AI workflow has the scoring table.
Record the baseline: two to four weeks of cycle time, hands-on minutes per case, rework rate and backlog.
Write a one-page AI policy: approved tools and company accounts; data that must never go in, such as patient data or privileged client documents; what a person checks before output leaves the company; telling customers when they talk to a chatbot; how to report a mistake.
Start AI literacy measures: a short session per role on what the tools do, where they go wrong and what the policy says. Record who attended.
Agree the gate criteria for the year, while nobody is attached to the project yet.
A policy that only forbids pushes AI use out of sight. As Wharton’s Ethan Mollick put it: “There are more reasons for workers to not use AI publicly than to use it.” Name where experimenting with approved tools is welcome.
In the EU, literacy measures are a legal duty. Article 4 of the AI Act, as rewritten by the Digital Omnibus in force since 27 July 2026, requires providers and deployers to take measures to support their staff’s AI literacy, without guaranteeing any specific level. The Commission’s AI literacy Q&A says no certificate is needed and an internal record of trainings is enough. Our EU AI Act guide covers the rest. This is general information, not legal advice.
Outputs: workflow description with owner · baseline · published policy · training record · approved budget · gate criteria.
Gate 1 (week 4): go if the owner has time in the calendar, the baseline is recorded, 20 or more real cases are in hand and the policy is out. Wait if data access is unresolved.
Months 2 to 3: how do you build and launch the first workflow?
Start with the boundaries. One page: what the system does, what a person approves, when it escalates, who can stop it. Put approval points wherever output leaves the company, an action is hard to reverse, or an amount crosses a threshold. Our post on human-in-the-loop approval points goes deeper.
Build inside the tools the team already uses, such as the inbox or the ERP, with the simplest design that works. Gartner predicts that over 40 percent of agentic AI projects will be canceled by the end of 2027, and suggests plain automation for routine workflows. Most first workflows are routine.
Before go-live, measure the error rate on a test set of past cases with known answers, run it alongside the old process for two to four weeks, and train the team on what to check.
Outputs: a live workflow in existing tools · boundaries page · test-set results · trained team · go-live note.
Gate 2 (month 3): switch over if the test set and parallel run meet the thresholds set at Gate 1 and reviewers can check output in the time planned. If close, extend once by up to a month; if not, stop. A stop here is cheap.
Months 4 to 6: how do you know it worked, and what comes next?
After six to eight weeks live, measure the same four numbers on a similar mix of cases. Report what changed, including review time and cost per case. Our guide on how to measure AI ROI has the method. Keep the metrics few: in S&P Global’s 451 Research survey, companies tracked more AI KPIs year over year, but projects were not necessarily performing better against them.
Then choose the second workflow by re-scoring the list. A neighbor that reuses the same integration, data or approval pattern often moves up.
Write the playbook while lessons are fresh: five to eight pages covering workflow selection, the baseline sheet, the boundaries template, test sets, the launch checklist and the policy. It is what makes the third workflow faster than the first.
Outputs: results report against the baseline · second workflow in build · playbook version 1 · updated policy.
Gate 3 (month 6): expand if workflow one met its threshold. Fix first if rework rose. Retire it if nothing moved and no fix is in sight, and write down why.
Months 7 to 12: how do you scale from one workflow to a portfolio?
A portfolio, not a backlog. List live and planned workflows, each with an owner, a metric, a status and a review date. In our view, two to four more workflows in this period is a realistic aim for a company this size. More than that usually means some have no owner.
Light governance. An AI register of each system, its owner, data, approval points and vendor. A 30-minute monthly review with the CEO and owners, and an incident log. NIST’s voluntary AI Risk Management Framework (Govern, Map, Measure, Manage) is a good reference; the certifiable ISO/IEC 42001 standard, published December 2023, is worth it mainly if customers ask. In the EU, check whether any workflow touches a high-risk area such as hiring or credit: according to the European Commission, those rules apply from 2 December 2027, and the vast majority of AI systems in use are minimal risk.
A vendor review. Check each vendor’s data terms, prices and exit path for your prompts, test sets and data. Models retire on someone else’s schedule. Anthropic’s deprecation page promises at least 60 days’ notice; on 30 September 2026 it announced that the Claude Sonnet 4.5 version dated 29 September 2025 retires on 30 November 2026. Each migration means rerunning your test set. Be wary of labels, too: Gartner estimates only about 130 of the thousands of agentic AI vendors are real.
Outputs: portfolio list · AI register · monthly review · vendor review · year-two plan and budget.
Gate 4 (month 12): keep, fix or retire each workflow, and approve year two.
Who does what?
Role | What they do |
|---|---|
CEO or founder | Sets the one-year goal and budget · chairs the four gates · names the owners · says openly what happens with freed time |
Workflow owner | Runs the workflow day to day · records the baseline · signs off boundaries and approval points · reports at each gate |
IT and security | Grants data access · approves tools and accounts · checks data protection and vendor terms · keeps the AI register |
The team | Shows where the time goes · tests on real cases · reviews and approves output · flags errors and proposes the next workflow |
An outside partner can build and train, but the owner stays inside. BCG found that many CEOs see the need for accountability but leave ownership spread across functions. This table closes that gap.
What does the research say about companies that scale and companies that stall?
McKinsey. In the State of AI 2026, 37 percent of 1,719 respondents attribute at least some EBIT impact to AI, essentially flat on 2025. About 6 percent are high performers. Nearly three-quarters of them report fundamentally redesigning workflows because of AI, against one-quarter of others, and they are twice as likely to report committed senior leaders and defined processes to measure impact. In the same report, associate partner Tara Balakrishnan names the challenge: “to avoid simply accumulating tools or pilots and instead to build a repeatable capability to transform these efforts into new ways of working.”
BCG. Its Applied AI Index 2026, a survey of more than 1,300 executives, classes 7.5 percent of companies as future-built and 41 percent as scaling. Future-built companies concentrate AI into a single, coherent program, and 95 percent use clear KPIs or track the P&L value. In BCG’s CEO survey, 64 percent of CEOs say their company pursues AI pilots, but only 26 percent embed AI in broader business change, and only 14 percent clearly define the P&L impact for all AI initiatives. BCG’s “10-20-70” rule (10 percent of effort on algorithms, 20 on technology and data, 70 on people and processes) is a consulting heuristic, not a measured result.
Those that stall. In 451 Research’s survey of 1,006 respondents, fielded in late 2024, the share of companies abandoning most AI initiatives before production rose from 17 to 42 percent, and the average company scrapped 46 percent of its proofs of concept.
Read this with care. The surveys are self-reported, the leading groups are small, and correlation is not cause. They also lean large: 36 percent of McKinsey’s respondents work for companies with more than $1 billion in revenue. What transfers to a 60-person company is not the budget but the habits: one focused program, a named owner, a baseline, and workflows redesigned rather than decorated.
How much budget should each phase get?
Budget the year, not the pilot: the first build, twelve months of running costs and your own people’s time. The model is usually a small line. Among German AI users in Bitkom’s survey, 50 percent named data preparation and 41 percent integration as large cost items, against 8 percent for token consumption.
Phase | What you pay for · hypothetical figure |
|---|---|
Weeks 0 to 4 | Mostly internal time: owner, workshop, baseline · training · company accounts for approved tools |
Months 2 to 3 | First build: scoping, integration, test set, parallel run · about €18,000 including internal time in our worked example |
Months 4 to 6 | Running workflow one, about €1,000 a month in the same example · second build, often smaller if it reuses integrations |
Months 7 to 12 | Running costs across the portfolio · further builds · a few days for governance and vendor review |
The figures come from the hypothetical example in what AI implementation costs in 2026, which has the full worksheet. Yours depend on how many systems you connect, the state of your data and how much output a person checks. For reference, a Postmodel implementation project starts at €10,000, or $12,000 in the US.
What are the most common AI roadmap mistakes?
The 40-use-case strategy deck. A long list feels like progress and delays the first decision. BCG describes strategic clarity as focusing on a few high-value, end-to-end plays. Finish one, then pick the next.
Buying tools first. Seats for everyone without a workflow, owner or training produce scattered use and no measurable result. In the Work Trend Index, only 39 percent of people using AI at work had received AI training from their company.
No stop criteria. Without thresholds agreed in advance, every pilot stays “promising” and nothing is decided.
Policy and training last. By then, habits have formed in personal accounts, and in the EU the literacy duty already applies.
One-page AI implementation roadmap template
Copy this into a document and fill in the right-hand column. The entries are a hypothetical example: a 140-person medical supplies distributor whose first workflow is customer orders arriving by email. Anything you cannot fill in is your first task.
Field | Fill in (hypothetical example) |
|---|---|
Goal and budget | |
One-year goal | Three workflows live, each measured against a baseline |
Sponsor and 12-month budget | CEO · build, running costs and internal time: €____ |
Readiness: owner · data · workflow · budget · policy | Yes · yes · yes · yes · no (policy in week 1) |
Weeks 0 to 4 | |
Workflow 1 and owner | Email orders entered into the ERP · head of customer service, half a day a week |
Baseline | 1,800 orders a month · 7 minutes each · 4% corrected · 1-day backlog |
Policy, training and Gate 1 | Policy out in week 1 · 45-minute session per team · Gate 1 in week 4 |
Months 2 to 3 | |
Approval points | A person approves new customers, orders above €5,000 and anything unreadable |
Test set and go-live threshold | 200 past orders · 2% or fewer field errors · three-week parallel run |
Gate 2 | Month 3: switch over, extend once, or stop |
Months 4 to 6 | |
Results to report | Minutes per order · error rate · backlog · cost per order |
Workflow 2, playbook, Gate 3 | Supplier delivery-date confirmations · playbook by month 6 · Gate 3 in month 6 |
Months 7 to 12 | |
Portfolio and register | Workflows 3 and 4 from the re-scored list · IT lead keeps the register · monthly review |
Vendor review | Month 9: data terms · prices · model retirements · exit path |
Gate 4 | Month 12: keep, fix or retire each workflow · year-two budget |
Where to start
Run the readiness check with your leadership team this week. Any no is your first task. If all five are yes, choose the first workflow with our scoring method, and read why AI projects fail before you set your gate criteria.
Each workflow on the roadmap goes through the same four steps: Context, Workflow, Action and Learning. To check workflow one, book a fit check.
Frequently asked questions
What belongs in an AI implementation roadmap?
Less than most templates suggest: the first workflow and its owner, the outputs of each phase, a decision gate per phase with criteria agreed in advance, a budget for the year, and the rules on tools, approvals and training. It does not need 40 use cases or a three-year vision deck.
Why plan 12 months if the first workflow goes live in a quarter?
The rest of the year is for owning it: measuring results, adding a second workflow, writing the playbook and setting up light governance. RAND advises leaders to commit a team to one problem for at least a year, and warns that projects on an overly accelerated timeline tend to fail.
How do you know your company is ready to start?
Answer five questions with a plain yes or no. Is there an owner with time? Can someone pull 20 to 50 real cases this month? Is there one measurable workflow, an approved budget, and a policy staff know? A no on owner or data access is the first thing to fix.
What are the most common AI roadmap mistakes?
A strategy deck with 40 use cases, buying tools before there is a workflow and an owner, pilots without stop criteria, and policy and training left to the end. By then, habits have formed in personal accounts, and in the EU the literacy duty already applies.
Related reading
Sources
McKinsey & Company, QuantumBlack, The state of AI in 2026: On the road to ROI, August 25, 2026. Online survey, May 4 to June 8, 2026, 1,719 participants in 97 nations.
Boston Consulting Group, The Formula for Agentic AI Value: The Applied AI Index 2026, September 30, 2026. Survey of more than 1,300 CxOs and senior leaders.
Boston Consulting Group, CEOs Are Starting to See Value from AI. Now Comes Execution., July 22, 2026.
S&P Global Market Intelligence, 451 Research, Generative AI shows rapid growth but yields mixed results, October 27, 2025. Online survey of 1,006 respondents, October 21 to November 25, 2024.
Gartner, Lack of AI-Ready Data Puts AI Projects at Risk, February 26, 2025.
Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, June 25, 2025.
RAND Corporation, James Ryseff, Brandon F. De Bruhl and Sydne J. Newberry, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, August 13, 2024.
Bitkom, Erstmals nutzt die Mehrheit der Unternehmen KI, September 14, 2026. Telephone survey of 603 companies with 20 or more employees in Germany.
Microsoft and LinkedIn, 2024 Work Trend Index Annual Report: AI at work is here. Now comes the hard part, May 8, 2024. Survey of 31,000 knowledge workers in 31 markets.
Ethan Mollick, Making AI Work: Leadership, Lab, and Crowd, One Useful Thing, May 22, 2025.
Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal, July 24, 2026, Article 1(5) replacing Article 4 of the AI Act.
European Commission, AI Literacy: Questions & Answers, page updated July 27, 2026.
European Commission, AI Act, page updated August 3, 2026.
NIST, AI Risk Management Framework, released January 26, 2023.
ISO, ISO/IEC 42001:2023, AI management systems, published December 2023.
Anthropic, Model deprecations, accessed October 4, 2026.