How long does it take to build an AI roadmap?

A focused 90-day framework works for most mid-market companies: 2 weeks for audit, 2 weeks for use case identification, 4 weeks for a pilot, and 4 weeks for evaluation. The key is starting small and proving value before scaling.

How much does AI implementation cost for a mid-market company?

For off-the-shelf AI APIs and tools, expect $62K to $141K in the first year. For custom models, $100K to $250K. The biggest cost most companies underestimate is data engineering, which typically runs 2-3x the model development cost.

What are the biggest mistakes mid-market companies make with AI?

Three common mistakes: buying an AI platform before defining the problem, hiring a Head of AI before having a strategy, and treating AI as a technology project instead of a business decision.

How to Build an AI Roadmap When You're Not a Fortune 500

Most AI roadmap advice is written for companies with $50M innovation budgets, dedicated ML teams, and data infrastructure they’ve been building for a decade. If you’re running a mid-market company with five engineers, real product deadlines, and a CEO who just got back from a conference with strong opinions about AI, that advice doesn’t help you.

This is a framework for the actual situation: limited engineering time, imperfect data, and executives who need to see results before they’ll fund the next phase. It’s designed to get you from “we need to do something with AI” to a working pilot in 90 days.

The 90-Day AI Roadmap

Weeks 1-2: The Honest Audit

Before you write a single line of code or talk to a single vendor, you need to know what you’re working with.

Data inventory. Where does your data live? How clean is it? Who owns it? Most mid-market companies think their data is in better shape than it is. The “customer database” is often three spreadsheets, a CRM with 40% duplicates, and a legacy system that nobody has export access to. That’s not a judgment, that’s just a common reality, and you need to know what you’re actually working with before you commit to anything.

Systems map. Document every system, every integration, every manual process that involves data moving from one place to another. The manual processes matter most. Those are usually where AI can help first.

Team capabilities. Be honest. Do you have anyone who’s built and deployed a machine learning model in production? Not in a Jupyter notebook, in production. There’s a massive gap between those two things. Knowing where your team actually is determines what you can build versus what you need to buy or outsource.

Business process review. Talk to operations, sales, customer service. Where are people spending time on repetitive decisions? Where are they copy-pasting between systems? Where do they say “I just know from experience” when making a call? Those are your AI candidates.

This audit takes two weeks if you’re focused. Don’t skip it. Every failed AI initiative I’ve seen started by skipping the audit and jumping to solutions.

Weeks 3-4: Identify Three Use Cases

Three, not ten or twenty.

Take everything from the audit and score it on two dimensions: business impact and implementation effort.

Business impact means revenue, cost savings, or risk reduction that someone with a budget cares about. “Improve efficiency” is not a use case. “Reduce customer churn by identifying at-risk accounts 30 days earlier” is a use case.

Implementation effort means what it actually takes to get something working. Consider: data availability (do you have what you need, and is it clean enough?), technical complexity (off-the-shelf API, fine-tuned model, or custom build?), organizational change (does someone’s job change, and will they accept that?), and integration requirements (does this need to plug into existing systems?).

Plot your candidates on a 2x2 matrix. You want the top-left quadrant: high impact, lower effort. Pick three from that quadrant. If you don’t have three in that quadrant, you might not be ready for AI yet, and that’s a valid finding.

For each use case, write a one-page brief: the problem, the data required, the expected outcome, how you’ll measure success, and what it costs to find out if it works.

Weeks 5-8: Build One Pilot

One. You identified three use cases so you have options if the first one hits a wall. But you’re only building one at a time.

The pilot has to use real data. Not sample data, not synthetic data, not “we’ll clean it up later” data. Real data, with all its messiness. If the pilot can’t work on real data, it won’t work in production either, and you should find that out now when it’s cheap.

The pilot has to have real users. Someone in the business has to use the output to make a decision or complete a task. A model that sits in a notebook generating accuracy scores is not a pilot. It’s a science project.

Set a clear success metric before you start: “the model correctly identifies X in Y% of cases” or “this process takes Z% less time with the tool.” If you can’t define success upfront, you can’t evaluate the result.

Four weeks is enough for a focused pilot. If it’s taking longer, you’ve probably scoped it too big.

Weeks 9-12: Evaluate and Decide

This is where most companies get it wrong. They fall in love with the pilot, or they declare it a failure too quickly.

Evaluate against the success metric you defined, nothing else. Did it hit the bar or not? If it did, plan the production path (which is a whole separate effort ). If it didn’t, diagnose why. Was it a data problem? A model problem? A user adoption problem? Each has a different fix.

Also evaluate what you learned about your organization. Did the team struggle with the tools? Did the business stakeholders engage or check out? Did you discover data quality issues that go beyond this one use case? These meta-lessons are often more valuable than the pilot itself.

Then decide: scale the pilot, iterate on it, pivot to use case number two, or stop. All four are valid outcomes. The worst outcome is continuing to invest in something that isn’t working because you’ve already spent money on it.

Three Mistakes Every Mid-Market Company Makes

1. Buying a Platform Before Defining the Problem

“We need an AI platform” isn’t a strategy. It’s a procurement decision dressed up as one. Companies sign six-figure annual contracts with AI platforms before they’ve identified a single use case. Twelve months later, the platform is shelfware and the “AI initiative” is a line item nobody wants to explain to the board.

Define the problem first. Then figure out whether you need a platform, an API, a point solution, or a few Python scripts running on a cron job. Sometimes the unsexy answer is the right one.

2. Hiring a “Head of AI” Before Having a Strategy

This is the enterprise playbook applied to mid-market reality, and it doesn’t work. At a Fortune 500, a Head of AI has a budget, a team, and organizational support. At a mid-market company, they’re usually a single person with no budget, no team, and a mandate to “figure out AI for us.”

Six months later, they’ve built some prototypes nobody uses, they’re frustrated, and they leave. You’ve spent $200K+ on salary and opportunity cost with nothing to show for it.

Get the strategy right first. Run a pilot or two. Then you’ll know whether you need an AI hire, what kind, and what their first project will be.

3. Treating AI as a Technology Project

AI projects fail most often for organizational reasons, not technical ones. The model works but nobody trusts the output. The system is accurate but it changes someone’s workflow and they resist it. The predictions are good but there’s no process to act on them.

AI is a business decision with technology components, not the other way around. The business owner should drive the initiative. Engineering builds the system. If it’s the other way around, you get impressive demos that never reach production.

What “AI Readiness” Actually Means

Vendors love to sell “AI readiness assessments.” They’ll audit your tech stack, your cloud infrastructure, your data pipelines. All useful information. But none of it tells you whether your organization is actually ready for AI.

AI readiness comes down to two things.

Data culture. Does your organization make decisions based on data, or on gut feeling? If leaders routinely ignore dashboards and reports, they’ll ignore AI outputs too. AI doesn’t replace judgment, it informs it. If nobody values data-informed judgment today, AI won’t change that.

Decision-making speed. AI generates insights at machine speed. If your organization takes six weeks to approve a process change, the bottleneck isn’t the AI, it’s the approval chain. Companies that benefit from AI are companies that can act on what the AI tells them.

These aren’t technology problems. They’re organizational problems. And they’re worth solving even if you never deploy a single model, because a data-driven organization that makes fast decisions will outperform one that doesn’t, with or without AI.

What This Actually Costs

Realistic numbers for a mid-market AI initiative over the first 12 months:

If you’re using mostly off-the-shelf AI APIs and tools:

Cloud infrastructure and API costs: $1,000-$3,000/month ($12K-$36K/year)
Data engineering and cleanup: $20K-$40K (this is the part everyone underestimates)
Integration development: $15K-$30K
Pilot development: $10K-$25K per use case
Training and change management: $5K-$10K

Total first year: $62K-$141K

If you’re building custom models:

Add a data scientist or ML engineer (contract): $15K-$25K/month
MLOps infrastructure: $2K-$5K/month
Additional compute: $1K-$5K/month

Total first year: $100K-$250K

These numbers are for doing it right. You can spend less, but you’ll cut corners on data quality or monitoring, and you’ll pay for it later when the model degrades in production and nobody notices for three months.

Compare this to what enterprise AI vendors will quote you. Proposals for “AI transformation programs” routinely run $500K to $2M and deliver less than what a focused team can build for a fraction of that.

When to Bring in Outside Help

Build internal capability when:

You have strong engineers who are interested in AI/ML
Your use cases are core to your business and will require ongoing development
You plan to make AI a competitive advantage over the next 3-5 years
You can invest 6-12 months in learning curve before seeing results

Bring in outside help when:

You need to move fast and can’t afford a 6-month learning curve
Your first use case requires specialized expertise your team doesn’t have
You need an objective assessment of vendors , tools, or build-vs-buy decisions
You want to validate your strategy before committing budget
You need someone who can translate between the business team and the engineering team

The approach that tends to work: bring in an outside perspective to set the strategy and run the first pilot, with explicit knowledge transfer to the internal team built into the engagement. The goal of any outside help should be to make itself unnecessary. If an advisor is still essential two years in, something went wrong.

The Real Competitive Advantage

Mid-market companies have advantages that enterprises don’t. You can move faster. You have fewer stakeholders to align. Your data might be smaller, but it’s often more focused and relevant. You can go from idea to production pilot in 90 days. A Fortune 500 takes 90 days just to get budget approval.

The companies that actually make progress with AI pick the right problem, move quickly, learn from the result, and iterate. That’s something mid-market companies do better than large ones.

You don’t need a $50M budget. You need a clear problem, honest data, and the discipline to prove value before you scale.

I’m Eric Brown. I work with mid-market companies as a fractional CTO and AI strategy consultant in Denver. If your team is trying to figure out where AI fits without a Fortune 500 budget, let’s talk .