We’ve worked on 100+ AI engagements inside mid-market companies over the last several years. I want to write down what we’ve learned, in plain language, for the operators we work with. Not the polished version. The version I’d tell a friend over coffee.
Most of what’s written about “AI for the middle market” is wrong
Vendor-flavored. Hype-flavored. Enterprise-flavored. Almost none of it accounts for the actual reality of running a $5–100M business. So we keep finding ourselves writing notes back to clients explaining what’s actually true. This letter is the long version of those notes.
The 80% AI failure rate is real. It’s also misleading
The 80% number gets thrown around like a market statistic. It’s directionally correct. But it implies AI is broken, which it isn’t. The 80% are failing because the businesses weren’t ready, not because the technology was.
The businesses that fail share a pattern:
- AI handed to IT instead of operations
- No specific workflow named
- No internal owner
- No measurement
- Tools deployed, no workflow change
Twelve months later: “we tried AI, it didn’t work.”
The businesses that succeed share a different pattern:
- Owner-led
- Specific workflow named
- One person on the hook
- KPI measured
- Ship something rough in 30 days, iterate, move to the next one
The technology is the same. The pattern is the variable.
The tools commoditize faster than people realize
Two years ago, working with frontier AI required real engineering investment. Today, anyone can buy Claude or ChatGPT. In 18 months, the price will drop another 5x, and capabilities most operators think are “frontier” will be baseline. The companies that have built their advantage on tool access are about to lose it.
The companies that have built their advantage on workflows that fit their business, refined over 12 months keep it. The workflows are the moat. The tools aren’t.
If you’re paying for AI tools and can’t name a specific workflow they’ve changed, you don’t have a moat. You have a subscription.
Owner-led businesses move faster than PE-backed ones, on average
Counter to the prevailing narrative. The PE-backed companies have access to capital and operating-partner pattern matching, but they also have committees and procurement gating that slow AI deployment. Owner-led businesses where the owner personally said “go” moved faster.
The exception: PE firms running a deliberate portfolio AI program — centralizing rollout across portfolio companies — outperform both pools. The structural advantage of PE shows up when the rollout is orchestrated, not when it’s left to portfolio companies individually.
The first wins are smaller than vendors claim
We’ve watched companies write $500K AI strategy checks and produce decks. We’ve watched companies spend $25K on a single sprint and recover an FTE-equivalent of cost. The sprint usually beat the strategy work — not because $25K is more than $500K, but because the sprint shipped and the strategy didn’t.
The fastest-payback first projects we see consistently:
- Quoting acceleration
- Document processing
- Field-to-office reporting
- Senior knowledge capture (especially when someone’s retiring)
If your first AI project doesn’t fit one of those patterns, ask hard why. Most companies should start with one of them.
Senior people stay senior. Junior people get more capable faster
The fear that AI replaces senior judgment is mostly wrong. Senior people get more leveraged. They review more, originate less, focus on the edge cases.
The shift we see with junior people is more interesting. Junior employees with AI assistance ramp materially faster than juniors without. The senior estimator’s pattern recognition becomes accessible to the next-generation estimator through a workflow. The junior associate can produce work product that used to require 5 more years of seasoning.
That’s good news. It’s also a problem if you’ve been operating on the assumption that “we need 5 years to grow a senior.” That timeline compresses.
The hardest part is rarely technical
I’d estimate 80% of the failure modes we see are operational, not technical. The wrong owner. The wrong first project. The team not aligned. The CFO not engaged. The “AI committee” that meets but doesn’t decide. The vendor relationship that doesn’t have an internal equivalent on the other side of the table.
The technical work matters. We have to build something that works. But the technical work isn’t the hard part for most mid-market companies. The hard part is the operational change required for the technical work to matter.
Compounding is what matters most
A company that ships 5 small AI projects this year ships 10 next year — not because they have more budget, but because the muscle memory makes each one faster. Each workflow integrated teaches the team about the next one. Each measurement system reveals the next opportunity.
The companies that wait until 2027 to start aren’t 12 months behind. They’re 12 months behind on a curve that’s getting steeper.
This isn’t a panic claim. It’s an honest reading of how capability accumulates. The decision to start is more important than the choice of what to start on. Whatever you pick first, you’ll learn enough that the second is faster.
The model we built was designed around what we observed
AI Office at $2,500/month entry exists because we watched too many mid-market companies pass on $250K projects they actually needed, and we wanted a model that let them say yes without the bet. The month-to-month model with no lock-in exists because we wanted clients who could leave if we weren’t earning it. The KPI guarantee on sprints exists because we’d rather work for free than ship something we knew was going to miss the target.
None of these are accidents. They’re the response to specific failure modes we kept seeing.
What we still don’t know
A short, honest list of where we still don’t have great answers:
- How fast the commoditization curve moves over the next 24 months — and what it does to the value proposition of “operating partner” vs. “tool vendor.” We think the gap widens; we won’t know for sure until 2027.
- Whether weekly advisory touchpoints scale to the consultant capacity we’d want. The first cohort of clients will tell us.
- What the right conversion rate from advisory to a full build actually is at steady state. The real number could be a lot higher or lower than we model.
We’ll keep writing as we figure it out.
The ask
If you’re an operator working on AI and any of this resonated, I’d love to hear what’s working for you — and what’s not. The strategy improves when the people closest to the work push back on the parts that don’t match reality.
Reach out directly, or grab a 30-minute call.
— Ross