The Short Answer
Kaufman Rossin's mid-market survey, fielded December 2025 with NewtonX among 100 U.S. decision-makers, found 94% of companies already use generative AI tools while 2% have AI embedded across the business. The same sample shows why: 94% run AI tools and only 33% run integration or API management. With ibl.ai you own all the code and the data, so one governed platform layer serves every department.
The distance between those two numbers is not enthusiasm gone stale. It is a layer nobody bought.
What do the 2026 mid-market AI surveys actually measure?
Four separate studies, four methodologies, and one consistent shape — provided the numbers are quoted with the question that produced them.
| What was measured | Result | Survey (fielded) |
|---|---|---|
| Companies already using generative AI tools | 94% | Kaufman Rossin / NewtonX, 100 U.S. decision-makers (Dec 2025) |
| Companies testing or deploying AI | 83% | Kaufman Rossin / NewtonX (Dec 2025) |
| Companies with AI embedded across the business ("Operators") | 2% | Kaufman Rossin / NewtonX (Dec 2025) |
| Companies running integration / iPaaS / API management | 33% | Kaufman Rossin / NewtonX (Dec 2025) |
| AI integrated into operations / embedded in core processes (AI-using organizations only) | 86% / 36% | RSM, 1,030 U.S. and Canadian leaders (Mar 2026) |
| Average AI tools per mid-market organization | 4.2 | Freshworks, 12,021 IT decision-makers, over 9,000 of them mid-market (Mar 2026) |
| Mid-market AI spend lost to complexity overhead | 25% | Freshworks (Mar 2026) |
| CEOs saying AI has value / reporting a company-wide strategy | 98.5% / 7% | VirtuousAI with Chief Executive Group (Jan 2026) |
Two corrections to how these get repeated. The Kaufman Rossin report was published in 2026 but fielded in December 2025, over a one-month period, among companies with 20 to under 1,000 employees and $5 million to under $1 billion in revenue.
The publication date was 14 May 2026, which is why the survey is often cited as a May 2026 finding about adoption that was measured five months earlier.
And the 94% and the 2% answer different questions. 94% is generative AI tool usage, which the report says typically "started informally, driven by individuals."
The 2% is the top rung of the report's four-stage ladder, 14% dabbling, 52% testing, 31% building, 2% operating, with a further 1% not started.
Why does 94% adoption produce only 2% enterprise-wide operation?
Because the survey that found the 94% also asked which data and technology tools the company runs, and the answers do not line up.
AI tools: 94%. Knowledge search: 79%. Business intelligence: 75%. Then the foundational half of the stack drops off — integration, iPaaS and API management at 33%, data science and ML platforms at 31%, workflow automation at 31%, data movement and transformation at 29%.
Kaufman Rossin calls this the AI iceberg, and states the consequence plainly: "Only a third of the companies we surveyed have built connections between the two parts of the AI iceberg, which means manual intervention is still needed to make use of AI outputs."
Their worked example is an FP&A analyst downloading data from NetSuite into Excel, pasting charts into a slide tool, then exporting a PDF for management. The model wrote the narrative. A person still carried it between four systems.
That is what a 94%-to-2% gap looks like at the desk level. Productivity improves inside one step of a workflow, and the workflow itself is unchanged.
Is the mid-market AI gap a department problem or a platform problem?
It is usually described as a department problem, and the survey data does not support that description.
Departmental politics is the intuitive story: compliance picked one vendor, operations picked another, nobody coordinated.
But the barriers mid-market leaders actually named are cybersecurity concerns (43%), AI talent shortages (42%) and legacy system integration (41%) — not interdepartmental friction.
Silos show up in the data as an architectural fact rather than a cultural one. 38% of mid-market companies still operate with siloed data, where departments run systems that do not connect, and only 16% describe their data as governed and integrated.
VirtuousAI's January 2026 survey with Chief Executive Group puts the same finding in CEO language: 86% cite lack of AI expertise as a barrier and 81% report difficulty integrating AI with existing systems.
So the corrective is worth stating directly. Departments did not refuse to cooperate. Nobody built the layer through which they would have cooperated, and four separately purchased tools have no shared layer by construction.
What did Accenture's October 2026 earnings call say about why AI deployments stall?
It described the same wall from the side of the people paid to climb it, two weeks after this analysis first ran.
On Accenture's Q4 fiscal 2026 earnings call on 1 October 2026, chief executive Julie Sweet located the demand in foundations rather than models.
"Much of our growth today comes from continuing to build their digital core, data foundations and the enterprise AI stack that they need to use AI at scale," she said, adding that "clients remain at very different stages of readiness."
Lack of interest is not the explanation. Sweet reported that "nearly 100 additional clients initiated their first advanced AI work with us this quarter, bringing the fiscal 2026 total to more than 400," on quarterly revenue of $18.68 billion, up 7% in local currency.
The work being paid for is the layer underneath the AI — the same layer the 33% integration figure says most mid-market companies have not built.
Why does a model with a 1M-token context still fail across departments?
Because context length solves recall, and the cross-department problem is reference.
A model can read a million tokens. It cannot know that the "customer" in your CRM, the "account" in your accounting system and the "requester" in your ticketing tool are three records for one organization, with different identifiers and different rules about who may see what.
None of that is in the weights, and none of it arrives by uploading documents. It is a structured, machine-readable map of your concepts, relationships and taxonomies, and it has to be built — the work set out in the enterprise AI data ontology.
This is why the pilot succeeds and the rollout does not. A pilot touches one system, so the missing map does not bite. Scaling means crossing systems, which is exactly where the absent relationships live.
The sequencing matters more than the tooling. Build agents first and you get the 94%. Build the layer first and you get a shot at the 2%. The longer case that the data layer, not the model, is the durable advantage is in intelligence is a commodity, the data layer is the moat.
What does a platform layer have to do that a point solution cannot?
Four things, and each one is a property of the layer rather than a feature of any tool sitting on it.
- One integration surface. Connections to the systems of record — HRIS, CRM, ERP, ticketing, accounting — are built once and reused by every agent, instead of once per vendor.
- One governance framework. Freshworks found mid-market organizations run an average of 4.2 AI tools and only 33% have a formal, consistently applied AI governance framework. Four security reviews produce four policies; one platform produces one.
- One identity model. Access binds to the company's existing identity provider and is enforced server-side, so a role change propagates everywhere rather than in four admin consoles.
- One memory boundary. What an agent learns about a person is stored under a scope the organization defines, not inside a vendor's account.
The economics follow from the same property. Freshworks estimates mid-market companies lose an average of 25% of AI spend to complexity overhead before any return — roughly $16.29 billion annually in the U.S. alone.
That overhead is the cost of not having the layer, paid in integration work and rework rather than in license fees.
The enterprise version of this argument, with the per-seat bill attached, is in SaaS fragmentation is the hidden cost of enterprise AI.
Why is shared memory the part that decides whether AI scales across departments?
Because an agent that cannot carry context past its own department is a better search box, not a colleague.
When each department runs its own tool, each tool keeps its own memory. The agent handling onboarding does not know what the agent handling policy questions established about the same employee last week, and no amount of model capability closes that.
The honest version of the fix is narrower than the marketing version, and it should be. On ibl.ai, agent memory defaults to the agent's own scope; only high-level facts — name, role, language, accessibility needs, broad goals — are stored globally and travel across agents.
Temporary state carries an expiry and is purged nightly, and duplicate detection runs semantically so the same fact is not stored five ways. Administrators read and manage every memory from one tab.
That combination is what makes cross-department context safe to enable. A global scope that captured everything would be a compliance problem wearing a convenience label, which is the same trade examined in finance AI with no audit trail is evidence, not speed.
How does ibl.ai give a mid-market company one AI platform layer instead of four tools?
By shipping the layer itself, and shipping it to the customer.
With ibl.ai you own all the code and the data.
The platform deploys on your own infrastructure with full source code, so the integration layer, the governance rules, the retention windows and the memory scopes are components your team reads and changes rather than vendor behavior you observe.
It is model-agnostic across any LLM, so choosing a better model later is a configuration change rather than a migration between four vendors on four renewal cycles.
Billing is usage-based with no per-seat pricing, so bringing a fourth and fifth department onto the same governed layer does not multiply the bill by headcount.
You can deploy anywhere: your own cloud, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity.
Agents reach systems of record through an MCP-based interoperability layer that assembles a per-user memory, under RBAC tied to your identity provider, with audit logs, retention policies and data residency options.
1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY.
Related reading: finance AI with no audit trail is evidence, not speed — the same mid-market gap measured inside one function; and SaaS fragmentation is the hidden cost of enterprise AI — the enterprise version, with the per-seat bill itemized; and agents that only read are demos — what the 2% looks like at the connector layer, once an agent has to write back into a CRM or ERP.
Related: PE Firms Put AI Agents in Healthcare. The Data Comes First — the same gap inside sponsor-backed healthcare companies, where the published AI result started from a clean dataset.
Sources: the 94%, 83%, 2%, the maturity ladder, the tool-stack percentages, the barrier percentages and the data-readiness figures from Kaufman Rossin's The State of AI in the Mid-Market, fielded with NewtonX in December 2025, and its 14 May 2026 publication date from Kaufman Rossin's press release; the Sweet quotes, client counts and quarterly revenue from Accenture's Q4 fiscal 2026 earnings call transcript; the 86%/36% and the sample from the RSM Middle Market AI Survey 2026; the 25%, $16.29 billion, 4.2 tools and 33% governance figures from Freshworks' Cost of Complexity research; the 98.5%, 7%, 86% and 81% from VirtuousAI's research with Chief Executive Group.