AI & Automation

How to Build a 12-Month AI Roadmap for a Mid-Sized Company

November 11, 2026 · 4 min read · Etzal Group Team

Most AI roadmaps fail for an ordinary reason. They are written as a list of technologies, and a list of technologies has no owner, no deadline and no number that tells you whether it worked. An AI roadmap that survives a year looks different. It is a short list of workflows, each with a person responsible, a measurement taken before anything changes, and a date on which somebody has to decide whether to continue.

This guide lays out a twelve-month plan that a company of fifty to five hundred people can actually run.

Start from workflows, not from tools

Open the roadmap with a simple inventory. For each department, write down the recurring tasks that take the most hours, cause the most rework, or make customers wait. Do not write "use generative AI in customer service". Write "answering order status questions takes four people three hours each per day".

A useful workflow entry has five fields:

  • What happens: the task in one sentence.
  • Volume: how many times per week.
  • Time per case: measured, not guessed.
  • Cost of an error: what goes wrong when it is done badly.
  • Where the data lives: the system, spreadsheet or inbox involved.

Twenty to thirty entries is enough. The inventory itself often takes two weeks and is the most valuable document the project produces, because it is the first time the company sees its own repetitive work in one place.

Choose three workflows for the year

Score every entry on two axes. The first is value: hours saved, revenue protected or errors avoided. The second is feasibility: is the data reachable, are the rules clear, and can a person check the result quickly?

Then choose three:

  1. A quick, low-risk one. Something internal, such as summarising meeting notes into tasks or drafting replies that a person approves. Its job is to teach the organisation how an AI project runs.
  2. One that customers or revenue can feel. For example, faster first response to enquiries or quotation drafting.
  3. One foundation. Usually a clean, searchable knowledge base or a tidy customer data set. It is unglamorous and it decides whether year two is possible.

Resist the fourth. Every additional project divides the attention of the same few people.

Quarter by quarter

Quarter 1: baseline and design. Measure the three workflows as they are today. Write the target for each in plain numbers, such as "first response under ten minutes for 80% of enquiries". Decide who approves AI output and in which cases it may act alone. Write the stop condition.

Quarter 2: the first pilot. Run the low-risk workflow with a small group. Keep a person approving every output for the first weeks. Record every correction, because the corrections are your evaluation set later.

Quarter 3: the customer-facing pilot and the foundation. Start the second workflow using what the first one taught you about review and data. In parallel, build the knowledge base or clean the data the third item needs.

Quarter 4: decide. For each workflow, compare the result with the baseline. Scale what met its target, redesign what came close, and close what did not. Then write the next year's inventory with real experience behind it.

Fund in stages

Do not approve a twelve-month budget in January. Approve the baseline work first. Release the pilot budget only when the baseline and targets exist on paper, and release the scaling budget only when the pilot met its number. Staged funding makes the roadmap honest, because each stage has to earn the next.

Include the costs that are easy to forget: the time of the staff who review output, integration work with existing systems, model usage fees that grow with volume, and maintenance after launch.

Write the stop conditions first

Before each pilot begins, write down what result would make you stop. For example: "If, after six weeks, a reviewer still corrects more than one in five drafts, we pause and redesign." Stop conditions written in advance are agreed calmly. Stop conditions invented after a disappointing result turn into arguments.

What the roadmap document contains

Keep it to a few pages:

  • The workflow inventory with scores.
  • The three chosen workflows, with owner, baseline, target and stop condition.
  • The rules for human approval.
  • The staged budget.
  • A one-page risk note on data privacy and on what the AI is not allowed to do.

If you want an outside view when building it, our AI Consulting engagement produces exactly this set of documents, and the fee is deducted when the work moves into implementation. You can also contact us with your inventory and we will tell you which entries look feasible.

The short version

Pick workflows, not tools. Measure before you change anything. Run three projects, not ten. Release money in stages, and decide in advance what failure looks like. A roadmap built this way is short, and it is still true in December.

Frequently asked questions

How many AI projects should a mid-sized company run in the first year?

Three is a sensible ceiling. One should be low risk and quick, one should touch revenue or customers, and one should build a foundation such as clean data or a shared knowledge base. More than three usually means none of them gets an owner with enough time.

Who should own the AI roadmap?

An operations or business leader who owns the workflows being changed, with IT as a partner. When the roadmap is owned only by IT it tends to become a list of tools. When it is owned by the people who run the work, it becomes a list of problems with numbers attached.

What should we do if a pilot does not reach its target?

Read the stop condition you wrote before the pilot began. If the result is under that line, stop and record why. A pilot that ends on purpose is cheap. A pilot that drifts for a year because nobody defined failure is the expensive kind.

Author

Etzal Group Team

Etzal Group Team, helping businesses grow with technology. Contact us.

Related service

WhatsApp AI Chatbot

We build, train, and run a WhatsApp AI agent that answers customers 24/7, qualifies leads, and follows up.