The 6-step plan

1. Pick the wedge use case — one workflow, one team, measurable.
2. Define success in numbers — before scoping anything.
3. Get the data ready (or honestly accept what you have).
4. Choose the build path — agency, in-house, or hybrid.
5. Run the project the right way — weekly demos, fixed scope.
6. Launch, learn, and pick the next one — AI is a programme, not a project.

If you’re running a mid-sized business in 2026 — somewhere between 50 and 500 employees — you’re probably stuck in the same place as most of your peers: you know AI matters, you’ve seen vendors pitch ten different solutions, your team has tried ChatGPT, but you haven’t shipped anything that meaningfully changes how the business runs.

The mid-market sits in an awkward gap. Too big to wing it on free tools. Too small for the “€500K AI transformation” engagement that consultancies pitch to enterprise. You need something practical: one specific workflow improved, in production, this quarter, without burning a year of budget figuring out the how.

This is that playbook. Six steps. About 90 days from start to first production system. Used at companies between 80 and 600 staff. No fluff.

Step 1: Pick the wedge use case

Almost every failed mid-market AI project starts with the wrong scope. “We want AI for everything” or “We need an AI strategy” sound serious in board meetings. They produce nothing.

Pick a wedge: one specific workflow, one team that owns it, where the value is measurable.

Good wedge candidates:

Customer support: deflection rate, first-response time, ticket-to-resolution time.
Sales operations: lead-to-opportunity conversion, lead enrichment time, proposal generation time.
Finance: invoice processing time, reconciliation lag, error rate on AP.
Operations: dispatch time, document-processing throughput, scheduling-conflict rate.
Marketing: ad campaign creation time, content turnaround, lead-list research time.

Bad wedge candidates:

“Make us more efficient.” Too vague.
“Predict customer churn.” Sounds great, requires a year of clean data and a team to act on it. Most mid-market companies have neither.
“Automate everything in finance.” Too big, no clear team owner, no way to know if it worked.
The CEO’s pet project that nobody else cares about.

The wedge has three properties: specific (one workflow), owned (one team is responsible for the outcome), measurable (you can put a number on the before-and-after). Without all three, the project will drift.

Step 2: Define success in numbers, before any vendor talks to you

Pick one or two metrics, write them down, and refuse to start the project until they’re agreed.

Examples:

“Reduce average customer-support ticket-to-resolution time from 18 hours to under 4 hours for tier-1 issues.”
“Process 80% of incoming invoices automatically, with a human review queue for the remaining 20%, with under 1% error.”
“Cut sales-rep time spent on lead enrichment from 3 hours/day to 30 minutes/day.”

Notice what these have in common: a baseline (the current number), a target, and a definition of done. They take an hour with the team that does the work today, and they pay for themselves over the entire project. Vendors who can’t commit to numbers are the ones who don’t plan to deliver them.

The other thing this does: it gives you a kill criterion. If at week 6 the prototype isn’t getting close to the numbers, you stop — before you’ve sunk the rest of the budget. Without numbers, projects drift on for months past the point they should have been killed.

Step 3: Get the data ready, or accept honestly that you can’t

The number-one cause of cost overruns in mid-market AI: data. Specifically, “we have the data, it’s in our systems” — followed by 4 weeks of discovery showing the data is fragmented, dirty, scanned, in a 1996 ERP, or doesn’t actually exist.

Before kick-off, do a 60-minute data audit:

List the data the AI needs. (e.g. order history, customer profile, invoice line items, support ticket text.)
For each one, where does it live? (Specific system, specific table, specific document folder.)
Who has access? Can the project team get access in week 1?
What format? (Database, API, CSV, PDF, scanned image, handwritten?)
How clean is it, honestly?

If most of the data is in modern systems with APIs, you’re a green-light project. If most of it is in PDFs and scanned documents, the AI work is the easy part — the data preparation will be the cost driver. This article explains how data quality drives cost.

The right vendor will do this audit with you in week 1. The wrong one will quote without checking and surprise-bill you later.

Ready for a real number?

Estimate your custom AI project in 30 seconds

Three questions, an instant cost range and timeline based on real shipped projects. After 30 minutes on a discovery call you have a written fixed-price quote.

Try the cost estimator →See the timeline

Step 4: Choose the build path

Three options for mid-market companies. Pick one based on your team and your timeline.

Option A: Specialist agency (recommended for first project)

Time-to-production: 6–12 weeks.
Cost: €25K–€150K depending on scope.
Best for: first or second AI project, when speed matters and you’re not yet sure how big a programme this becomes.

Option B: In-house (recommended once you have 3+ projects in production)

Time-to-production: 6–12 months.
Cost: €350K–€700K Year 1 for a small team.
Best for: companies where AI has become a permanent capability and you want institutional knowledge to compound.

Option C: Hybrid (where most companies end up)

Time-to-production: 6–12 weeks for the first project, agency-led.
Cost: agency project + one senior in-house hire.
Best for: ambitious mid-market companies that want to ship fast AND build internal capability.

For most mid-sized businesses, the right answer for the first project is the agency path. You learn from someone who has shipped this before. You ship fast. If AI becomes a real programme — you start hiring. If it doesn’t, you haven’t locked yourself into a year of headcount.

The full agency-vs-in-house breakdown is here.

Step 5: Run the project the right way

Mid-market companies often run AI projects like enterprise IT projects: 6-month plans, monthly steering committees, quarterly milestones. That’s how you ship in 18 months instead of 12 weeks.

What works:

One clear technical owner on your side. Not a committee. Probably your CTO or head of engineering, possibly your COO or operations head if there’s no CTO.
One sponsor. Usually the head of the team that will use the AI. They unblock decisions when discovery raises questions.
Weekly demos, every Friday. Not status updates — live demos of working software. Even if it’s 5% complete in week 1.
Fixed scope after week 2. The technical spec gets signed off after week 1–2 of discovery. After that, scope changes are explicit, written, and re-quoted.
Real users in week 4 or 5. Not after launch. Get 3–5 actual users from the operating team running the AI against real tasks while it’s being built.
Production cutover at week 8–12. Not “Phase 1”, “Phase 2”, “Phase 3”. Real production, real users, real impact on the metric you set in step 2.

If the project is run this way, you know within 4 weeks whether it’s on track. If it’s not, you have time to course-correct or kill it before you’ve burned the budget.

Step 6: Launch, learn, and pick the next one

The first AI project is a foothold, not the destination. Once it’s live:

Measure against the success metric for 4–6 weeks. Did it hit the numbers? If not, why not?
Document what you learned. Where did the AI surprise you? What integration was harder than expected? What edge cases did your team handle that the AI didn’t?
Decide on the next 2–3 wedges. The first project earns the right to ask for the budget for the next ones. Don’t squander the moment.
Decide on internalisation. If you’re doing 3+ projects, hire a senior owner. If you’re doing 1–2 a year, agency continues to be cheaper.

The best mid-market AI programmes I’ve seen do three projects in Year 1, total budget around €100K–€200K, ship 8–12 weeks each. By Year 2 they’re running 5–6 projects in parallel, often with the in-house senior owner running the agency engagements.

Pitfalls I see weekly

Picking the wrong wedge. “Predict customer churn” before you’ve cleaned the customer data is a 12-month project. Pick something with cleaner data.
Skipping the success metric. Without a number, the project drifts.
Engaging vendors before scoping. Vendors will scope the project to fit their products. Define what you want first, in your own words, then talk to vendors.
Letting consultancies turn it into a strategy engagement. Mid-market doesn’t need a 60-page AI strategy. It needs three projects shipped.
Underestimating change management. The AI builds in 8 weeks. Getting your team to actually use it can take longer. Train, demo, iterate, listen.

A worked example (anonymised)

Company: 180-person specialty distributor in Europe.

Wedge: invoice processing in their finance team.

Success metric: cut invoice-to-payment lag from 9 days to under 3 days; reduce manual touches from 100% to under 30%.

Data audit: invoices arrived as PDF email attachments. AFAS as the ERP. Bank API for payment status. All accessible.

Build path: agency-led project, fixed price €34K, 8 weeks. One internal stakeholder (CFO) sponsored, head of finance was the technical owner on the customer side.

Run: weekly Friday demos. Real finance team users testing from week 3. Cutover at week 7, with their old process running in parallel for 2 more weeks for safety.

Result: 5.5 days payment lag (vs 9), 22% manual touch rate (vs 100%). Hit the targets. Year 2 they added two more projects on the same agency engagement.

Where to start, today

If you’ve got a wedge in mind, the cost estimator will give you a ballpark in 30 seconds. The timeline page shows what a typical 8–12-week project looks like week by week. The process page includes a readiness checklist.

Or skip the reading and just book a discovery call. 30 minutes, no pitch, free. By the end you’ll know whether your wedge is well-scoped, what the data audit looks like, and a rough number to take to your CFO.

How to Start a Custom AI Project in a Mid-Sized Business