From Scattered Pilots to a Real AI Roadmap: A Plan for Mid-Sized Companies

Most mid-sized companies I walk into already have AI. What they don't have is a roadmap. Someone in marketing pays for ChatGPT Team on a corporate card, a couple of analysts quietly use Claude for cleaning data, and the CFO has Microsoft 365 Copilot switched on because it came bundled. Nobody can tell you what any of it is worth. That is the normal starting position in 2026, and it is fine. The problem is staying there.

A roadmap is not a slide that says "become an AI-first organization." It is a sequence of decisions about where to spend attention, in what order, with checkpoints that let you stop or double down. Below is the structure I use with companies between 100 and 1,000 people. Treat the timeline as a default, not a law.

Before the roadmap: three things you need in hand

You cannot plan against fog. Spend the first two to three weeks gathering three concrete things.

• An honest tool inventory. Pull the expense reports and SSO logs. You will almost always find more AI spend than anyone admitted to, spread across personal-tier and team-tier subscriptions. This is your shadow-AI baseline and your first security conversation.
• A shortlist of pain, not a wishlist of tech. Ask department heads where their teams lose hours to work nobody enjoys. You want verbs and volumes: "we answer 400 near-identical support tickets a week," "we spend three days a month reconciling vendor invoices." Ignore anyone who leads with a tool name.
• A named owner with real authority. Not a committee. One person, ideally reporting to the COO or CEO, who can approve a budget and kill a project. AI programs that report into IT alone tend to stall on procurement; programs that report into a single line-of-business leader tend to stay narrow. Pick someone who can see across functions.

Quarter 1: prove one thing, safely

The goal of the first quarter is not transformation. It is a single, defensible win that you can measure and repeat. Pick two or three use cases from your pain shortlist that share three traits: high volume, tolerant of imperfect output, and cheap to check. Support-ticket drafting, sales-call summarization, and first-draft marketing copy all qualify. Anything touching money movement, legal commitments, or regulated advice does not belong in Quarter 1.

Set up the boundaries first. Get a business-tier or enterprise agreement so your data is not training a public model. Microsoft 365 Copilot, ChatGPT Enterprise, Claude for Enterprise, and Google's Gemini for Workspace all offer this; the right one usually follows wherever your documents already live. Write a one-page acceptable-use policy that a non-lawyer can actually read. Decide what data is off-limits before people paste it in, not after.

Then measure a baseline you can compare against. If you skip this, you will never be able to prove value, and finance will be right to doubt you. For a support pilot, log current average handle time and first-response time for four weeks before the tool touches anything.

Budget expectation: Quarter 1 is small. Enterprise seats run roughly $30 to $60 per user per month in 2026. Twenty to forty seats plus a portion of your owner's time is the whole cost. If someone proposes a six-figure platform in month one, that is a red flag.

Quarter 2: turn the win into a repeatable pattern

Assume your pilot produced a real but modest result, maybe a 20 to 35 percent time reduction on the targeted task. Now the work is turning a good outcome into a boring, documented process other teams can copy.

Three moves define this quarter. First, write down what actually worked: the prompts, the review steps, the cases where the tool failed and a human took over. This becomes an internal playbook, and it is more valuable than any vendor's documentation. Second, name a champion in each department that will adopt next, someone respected who volunteers rather than being assigned. Third, start a lightweight measurement rhythm: a monthly one-pager per active use case showing hours saved, quality issues, and cost.

This is also when you confront integration honestly. A chatbot in a browser tab is easy. Getting AI to read your CRM, your ticketing system, or your document store is where the real friction lives. If a use case needs that connection, decide now whether you buy a tool that already integrates (Glean and similar enterprise-search products, or the native Copilot connectors) or you build a small integration. Do not let a promising pilot die because nobody owned the plumbing.

Quarter 3: scale what works, retire what doesn't

By now you should have two or three use cases with real numbers and at least one that underwhelmed. Kill the underwhelming one out loud. Publicly retiring a project that didn't pay off buys you more credibility than any success, because it proves the program has judgment.

For the winners, scaling means seats, training, and governance in that order. Roll out to the full teams that benefit. Run short, role-specific training; generic "intro to AI" sessions are forgettable, while "here is how our support team drafts a refund response" sticks. Stand up basic governance: who can procure new tools, how new use cases get reviewed, where sensitive data may and may not go.

Watch a failure mode here. Enthusiasm outruns evidence, and someone wants to automate a high-stakes process because the low-stakes ones went well. Hold the line. The traits that made Quarter 1 safe still matter.

Quarter 4: institutionalize and plan the next year

The final quarter is about making the program survive its founder. Move budget from experimental to operational lines so AI tooling is funded like any other software. Fold the acceptable-use policy and review process into normal IT and security governance. Publish a simple annual scorecard: hours saved, dollars saved or earned, adoption rate by team, incidents.

Then plan year two from evidence rather than hype. You now know which departments engage, which use cases pay, and where your data problems are. That is a far better basis for ambition than any analyst forecast.

What derails these roadmaps

The same handful of mistakes recur. Buying a platform before proving a use case. Measuring nothing, so wins evaporate into anecdote. Assigning AI to a committee that meets monthly and decides nothing. Chasing the flashiest use case instead of the highest-volume one. And treating adoption as a technology rollout when it is mostly a change-management problem wearing a technology costume.

A roadmap does not remove uncertainty. It gives you a cadence for spending your uncertainty deliberately, one quarter at a time, with the option to stop. For a mid-sized company that is not trying to win an AI arms race but does want real returns, that cadence is worth more than any single tool.