Midmarket teams lose 25% of their AI budget to complexity. Support teams should buy accordingly.

A wave of 2026 research says the same thing: what stalls midmarket AI isn't the model — it's the integration, data, and process work around it. That has a direct implication for how support teams should choose AI.

Written by

Chris Vavra

There's a number in Freshworks' new Global Cost of Complexity report that should stop any midmarket leader mid-scroll: on average, midmarket organizations lose 25% of their AI budget to complexity overhead — integration troubleshooting, data wrangling, governance, and rework — before a single return shows up. In the US alone, Freshworks estimates that at roughly $16 billion a year. They call it the complexity tax, and the rest of the report explains how it gets levied.

The findings are worth sitting with, because they don't describe a technology problem. They describe an execution problem — and the same pattern is showing up across nearly every serious study of midmarket AI this year.

The complexity tax, in the data

Freshworks surveyed 12,021 IT decision-makers across six countries, more than 9,000 of them in midmarket organizations. The picture that emerges is of AI that generates work faster than it removes it. Eighty-six percent of midmarket IT leaders say managing AI complexity has actually increased their team's workload — the opposite of the promise. Eight in ten report that AI outputs introduce noise, errors, or rework, the phenomenon the report calls "AI slop." Roughly a quarter of the team's AI-related time goes to troubleshooting and integration firefighting rather than strategic work.

And most of it never reaches production. Only 15% of midmarket organizations have AI integrated across multiple core operations; 36% are still in pilots or haven't deployed meaningfully. Meanwhile 72% of executives expect ROI within eight months, even though most deployments take six to twelve months just to go live. The clock starts before the system does.

The report is refreshingly direct about the cause. In its own framework, it notes that the root problem is rarely the AI model itself — it's the data and systems environment the model is asked to work in. When customer history lives in one system, ticket data in another, and the AI tool can't cleanly reach either, the model is reasoning with a fraction of what it needs.

It's not just one report

If this were a single vendor's survey, you could discount it. It isn't. The same finding recurs across independent 2026 research, with striking consistency.

RSM's Middle Market AI Survey 2026 found the top barriers to scaling AI were data quality, cited by 53%, and integration challenges, cited by 47% — not model limitations. RSM's own framing is that fragmented pilots and governance gaps are symptoms of organizations that haven't yet changed how work gets done. Kaufman Rossin's State of AI in the Mid-Market found that just 16% of midmarket companies have reached a fully governed, integrated data state, with legacy integration the most-cited obstacle; as one of its authors put it, AI can't simply sit on top of existing processes. Broader analyses point the same way: RAND and MIT's 2026 work found roughly 80% of AI projects fail to deliver measurable business value, and Deloitte reported that 42% of companies abandoned at least one AI initiative in the prior year. One survey of teams that did make it to production found the single biggest success factor wasn't better technology — it was rebuilding the process around the AI.

Read together, these reports converge on an uncomfortable, clarifying conclusion: the differentiator in AI outcomes is not budget or model choice. It's execution discipline — integrated systems, governed data, and a clean path from pilot to production. McKinsey's data, cited in the Freshworks report, quantifies the payoff: top-quartile companies generate 3.5 times more revenue impact from AI than median performers. The gap is execution, not intelligence.

Customer support is where the tax bites hardest

Most of this research spans IT broadly. But customer support is arguably the sharpest case, because support is where the fragmentation is worst — a helpdesk, a CRM, a knowledge base, and a fistful of channels, all of which the AI has to reach to be useful.

The CX-specific data mirrors the general finding exactly. Gartner's AI implementation research found that 62% of underperforming AI customer-service projects trace to data-preparation problems, not the technology. Roland Berger's 2026 study of more than 550 senior decision-makers named legacy systems, integration complexity, and data quality as the most frequently cited barriers to scaling — particularly for organizations already using AI and trying to expand it. And the practical cost of getting a support AI wired into a midmarket stack is real: weeks of integration engineering, plus the ongoing knowledge-base upkeep that most teams forget to budget for.

This is the same lesson we've written about before from the model angle: a support AI's performance is governed far more by the knowledge and systems it's connected to than by which model sits underneath. The complexity research is that same truth arriving from the budget side.

What midmarket support teams should take from this

Here's the part that turns diagnosis into a buying decision. If the complexity tax is levied by integration burden, data friction, and tool sprawl, then the way to avoid it is to choose AI that minimizes all three — not the AI with the highest benchmark score.

Notably, midmarket teams already sense this. In the same Freshworks research, buying behavior has shifted decisively toward fit over horsepower: 34% name integrating AI into current workflows as their top priority for the next two to three years, while only 7% want to replace systems wholesale. Ninety percent prefer solutions with built-in workflows over heavy configuration. Eighty-eight percent prioritize industry-specific context over raw technical capability, and 54% are buying AI capabilities rather than building them in-house. That is a market voting, clearly, for AI that works inside the business it already runs.

For a midmarket support team, that translates into a short, unglamorous checklist for evaluating any AI option:

Does it work inside the helpdesk you already run, or does it require a migration? Every study here flags rip-and-replace as a complexity multiplier, and the midmarket's own stated preference is overwhelmingly to avoid it. Does it deploy in days, not quarters — before the executive ROI clock runs out? Does it add to tool sprawl and governance load, or fold cleanly into what you already manage? And is the pricing predictable enough to budget, rather than a meter that compounds as your volume grows?

None of those questions is about the model. All of them are about whether the solution was built with a midmarket team's reality in mind — a lean team, an existing stack, a thin margin for error, and no appetite for becoming a systems integrator. That's the real selection criterion, and the research is unanimous that it's the one that predicts success.

The tax is optional

The complexity tax feels like the cost of doing AI. The evidence says it's mostly the cost of doing AI wrong — of buying powerful tools that don't fit, then paying for the gap in integration hours, rework, and stalled pilots. Midmarket support teams don't have to pay it. The move is to stop shopping for the most capable model and start choosing the solution built for how your team actually works: inside your existing systems, fast to deploy, simple to govern, and priced to predict.

Get that fit right, and the 25% the rest of the market is losing to complexity is yours to keep.

Chris Vavra is the founder of Flexivity AI, which adds AI to the osTicket and Zendesk helpdesks midmarket teams already run — no migration, fast to deploy, on predictable pricing.