Workflow and system integration in 2026: why the connection layer is now the limit on AI

The average enterprise runs 957 applications and has 27% of them connected. Meanwhile 86% of IT leaders say that without proper integration, AI agents add complexity rather than value. The model is not the constraint any more. The wiring is.

The average enterprise runs 957 applications and has 27% of them connected. In the same survey, 86% of IT leaders said that without proper integration, AI agents add complexity rather than value.

Those two findings come from Salesforce's 2026 Connectivity Benchmark, and they define the problem better than anything else published this year. The models work. The connection layer is where the work now is, and it is where the returns are stuck.

One thing to be clear about before going further: that survey of 1,050 IT leaders at organisations of 1,000 or more employees was conducted in October and November 2025 and published in February 2026. It is the best available measurement of connected-application share and there is no newer edition. We are using it with the collection date printed, because the alternatives are to quote it as though it were fielded this year or to leave the best measurement out.

The number depends on what you count

Three 2026 application counts, three definitions, all correct

Stacking these in one argument is the fastest way to lose a technical reader.

Source: Salesforce 2026 Connectivity Benchmark; Zylo 2026 SaaS Management Index; BetterCloud 2026 State of SaaS.

Three 2026 reports give three application counts and all three are correct. 957 counts all enterprise applications. Zylo's 305 counts licensed SaaS discovered through spend, against a median of 240. BetterCloud's 118 counts IT-managed SaaS, up 11% year on year, with mid-market jumping from 116 to 164.

This matters because the integration argument is often made by stacking incompatible numbers. If you quote 957 applications and then a 118-app orchestration statistic in the same paragraph, someone will check. Pick one definition and hold it.

The more useful figures sit underneath the headline counts. BetterCloud found 56% of total apps carry IT approval, and an average of 27 AI-powered SaaS applications per organisation, roughly 22% of the portfolio. Zylo found 36% of SaaS licences left unused and business units controlling 81% of SaaS spend against 15% managed directly by IT. The sprawl is not an IT procurement failure. IT does not control most of it.

What the connection gap costs

The applications, and the ones that talk to each other

• Applications the average enterprise runs. 957 apps. Up from 897 the year before. Enterprises of 1,000 or more employees.
• Share currently connected. 27%. IT teams spend 36% of their time designing, building and testing custom integrations.

Fieldwork October to November 2025, published February 2026. It is the best available measurement of connected-application share and there is no newer edition, so the collection date is printed rather than hidden. Source: Salesforce 2026 Connectivity Benchmark Report, 11th annual, 1,050 IT leaders across nine countries, conducted with Vanson Bourne.

Source: Salesforce 2026 Connectivity Benchmark Report, 11th annual, 1,050 IT leaders across nine countries, conducted with Vanson Bourne.

The Connectivity Benchmark's 36% of IT team time spent designing, building and testing custom integrations is the closest thing to a cost figure in the public record, and it is a time figure rather than a dollar one. There is a reason for that.

The $3.5M average cost of a custom point-to-point integration, which appears in almost every integration business case written in the last five years, comes from MuleSoft research fielded in 2021. We looked hard for a 2026 equivalent. Every current result traces to vendor content marketing with no survey base behind it. The same is true of iPaaS adoption rates: no 2026 survey-based figure with a disclosed method exists, and the market-sizing firms publishing CAGRs do not publish their methodology.

So the honest position for a 2026 business case is that the sector has not published a current integration cost benchmark, and the number has to come from your own last three builds. That is a weaker argument to make in a board paper and a much stronger one to survive scrutiny.

On the manual-work side there is one 2026-fielded figure. Frends, working with Sapio Research, surveyed 611 IT and business decision-makers across Germany, Finland, Sweden, Norway, Denmark and the Netherlands in April 2026. It puts 7.6 hours a week per knowledge worker on manual tasks, which it frames as 44 working days a year, rising to 8.5 hours in Germany. It is vendor-commissioned and Europe-only. The agency is named and the method disclosed, which is more than most, and it should be quoted with both caveats.

Where orchestration actually is

Orchestration and returns, quarter by quarter through 2026

KPMG runs a US series and a global series with different bases. The series is named on every row.

Source: KPMG AI Pulse US and Global series, Q1 to Q3 2026; Salesforce Agentic Enterprise Index, August 2026; McKinsey and QuantumBlack, State of AI in 2026.

The single most useful trend line in 2026 is KPMG's orchestration figure, and it needs care because KPMG runs two parallel series. In the US series, organisations orchestrating multiple agents across workflows went from 9% in Q1 2026, 237 respondents fielded 17 February to 17 March, to 18% in Q2, 204 respondents fielded 28 April to 25 May. It doubled in a quarter, from a very low base.

The global series tells the returns story. Across 2,145 senior leaders fielded 28 April to 25 May 2026, only 7% reported established ROI, down from 8% in Q1. And 49% said they had delayed or scaled back AI agent deployments when costs started outweighing benefits. The organisations with full real-time visibility into what their AI systems cost to operate were five times more likely to report established ROI, 15% against 3%.

By Q3 2026, fielded 23 July to 26 August across 2,131 leaders, 55% were operating a formal AI harness layer, rising to 86% among those reporting established ROI, while only 12% consistently assess AI value against cost across the organisation. The pattern across three quarters is consistent: the organisations that built a governed layer between the agents and the systems are the ones reporting returns.

Salesforce's own telemetry gives the physical measurement rather than the self-report. Between February 2025 and April 2026, the share of agent actions crossing into a secondary business function rose from 0.9% to 5.9%. Agents are beginning to reach across systems. Almost all of them still do not.

Gartner's April 2026 view of agent sprawl adds the governance gap: only 13% of organisations believe they have adequate AI agent governance. Gartner also forecasts that by 2028 the average global Fortune 500 enterprise will have over 150,000 agents in use against fewer than 15 in 2025. That second number is a prediction, not fieldwork, and should be labelled as one.

Why the EBIT line has not moved

McKinsey's State of AI in 2026, fielded 4 May to 8 June 2026 across 1,719 participants in 97 nations, found 37% attributing at least some EBIT impact to AI use, essentially unchanged from the year before. Scaling of agents at large enterprises rose from 27% to 40% over the same period.

More agents, same profit impact. The explanation is in the same survey: nearly three-quarters of high performers report fundamentally redesigning workflows because of their AI use, against about one quarter of everyone else. Connecting two systems so an agent can read both of them produces a faster version of the existing process. Whether that shows up in EBIT depends on whether anyone changed what the process is.

This is the part of the integration argument that gets skipped, because it is organisational rather than technical. The wiring is necessary. It is not sufficient, and a programme that stops at the wiring will report 37% too.

What to fix before the next agent

What to fix before the next agent is scoped

Ordered by how often the absence of it is what stops the programme.

• A named owner for the connection layer. Only 12% of organisations treat integration as a central governance layer. Integrations get built where the pressure is and nobody owns the resulting graph.
• An inventory of what is connected and what is not. 27% of applications are connected on average. Most organisations cannot name which 27%.
• Governance on the APIs that already exist. 27% of APIs are ungoverned; 54% of organisations have centralised governance, so roughly half run the layer without one.
• Cost visibility on the agents already running. Organisations with full real-time cost visibility were five times more likely to report established ROI, 15% against 3%.
• A decision about which workflow changes, not just which systems connect. Nearly three-quarters of high performers redesigned workflows; about a quarter of everyone else did. The EBIT line has not moved for the second group.
• A dollar figure for what an integration costs to build. No current published benchmark exists. The $3.5M figure in circulation is MuleSoft research fielded in 2021. Use your own last three builds.

Source: Frends with Sapio Research, State of Integration and AI 2026, fielded April 2026, n=611 across six European countries; Salesforce 2026 Connectivity Benchmark; KPMG Global AI Pulse Q2 2026; McKinsey State of AI in 2026.

Source: Frends with Sapio Research, State of Integration and AI 2026, fielded April 2026, n=611 across six European countries; Salesforce 2026 Connectivity Benchmark; KPMG Global AI Pulse Q2 2026; McKinsey State of AI in 2026.

The Frends finding that only 12% of organisations treat integration as a central governance layer explains most of what we see in practice. Integrations get built where the pressure is, by whoever is under it, and nobody owns the resulting graph. Five years later the graph is the constraint and no single team can describe it.

The work is unglamorous and it is the same every time: a named owner for the connection layer, an inventory of what is connected, governance on the APIs that already exist, and cost visibility on the agents already running. The Connectivity Benchmark found 27% of APIs ungoverned and 54% of organisations with centralised governance, which means roughly half are running the connection layer without one.

Where an integration layer is the right answer instead of a replacement, the argument is about tail risk rather than cost, and we have written that out separately in wrapping legacy systems instead of replacing them. If the underlying problem is that there are too many applications rather than too few connections, start with the application rationalization playbook. Our data foundation practice builds the layer the agents read from, and the AI readiness assessment is where we map what is actually connected before anything gets built on top of it.

What we would do first

Pick the workflow that currently requires a person to copy a value from one system into another, and count how many times that happens in a week. That is your swivel-chair inventory, and it is usually shorter and more valuable than the integration backlog, because every item on it has a person who can tell you exactly what the rule is.

Then, before any agent work is scoped, ask which system the agent will need to write to. Reading across systems is a retrieval problem. Writing across them is an integration and governance problem, and it is the point at which most agent projects discover that the connection layer was the project all along.

Frequently asked questions

How many applications does a typical enterprise actually run?

It depends entirely on what you count, and the three most-quoted 2026 figures are counting three different things. Salesforce's 2026 Connectivity Benchmark puts the average at 957 applications across enterprises of 1,000 or more employees, counting all enterprise applications. Zylo's 2026 SaaS Management Index says 305, median 240, counting licensed SaaS discovered through spend. BetterCloud's 2026 State of SaaS says 118, counting IT-managed SaaS. None of them is wrong. If you cite more than one in the same document, define each or the comparison is meaningless.

What share of enterprise applications are actually integrated?

27%, according to Salesforce's 2026 Connectivity Benchmark. The same report finds IT teams spending an average of 36% of their time designing, building and testing custom integrations, and 82% of IT leaders naming data integration as a top challenge. Note the fieldwork date: that survey of 1,050 IT leaders was conducted in October and November 2025 and published in February 2026. It is the best measurement available on connected-application share, and it is not 2026 fieldwork.

What does a point-to-point integration cost to build?

There is no current published figure, and the one in circulation should not be used. The widely quoted $3.5M average cost of a custom point-to-point integration comes from MuleSoft research fielded in 2021. We searched for a 2026 equivalent and could not find one that traces to a survey rather than to vendor content marketing. The honest position is that the sector has not published a current integration cost benchmark, so the number has to come from your own last three builds.

Why do AI agents make the integration problem worse?

Because an agent that cannot reach a second system is a chatbot. Salesforce's 2026 Connectivity Benchmark found 86% of IT leaders saying that without proper integration, AI agents add complexity rather than value, and 50% reporting that their agents operate in silos. On the measurement side, Salesforce's own telemetry across February 2025 to April 2026 shows the share of agent actions that cross into a secondary business function rising from 0.9% to 5.9%. Agents are starting to reach across systems, from a very low base.

How many organisations have actually reached cross-workflow orchestration?

18% as of Q2 2026, up from 9% the quarter before, in KPMG's US AI Pulse series. That is the figure worth tracking, and it is worth being precise about which survey it comes from: KPMG runs a US series and a global series with different bases. The 9% to 18% doubling is the US sample, 237 respondents in Q1 fielded 17 February to 17 March 2026 and 204 in Q2 fielded 28 April to 25 May 2026. BetterCloud's 2026 State of SaaS, a different survey of 525 IT and security professionals, puts it the other way round: 90% of organisations still lack true cross-app orchestration.

Is integration actually what is blocking AI returns?

It is the most-cited single barrier in the only 2026-fielded integration-specific survey we found. Frends, working with Sapio Research, surveyed 611 IT and business decision-makers across six European countries in April 2026 and found 36% naming integration challenges as the top barrier to AI progress, while only 12% treat integration as a central governance layer. In the same study only 26% of AI projects delivered measurable positive P&L impact and 63% remained in investigation or pilot stage. It is vendor-commissioned and Europe-only, and worth reading with both caveats attached.

What should we fix before adding another agent?

Four things, in order. Name an owner for the connection layer, because Frends found only 12% of organisations treat integration as a central governance layer and that is where the sprawl comes from. Inventory what is already connected and what is not. Govern the APIs you have, since the Connectivity Benchmark found 27% of APIs ungoverned. And put cost visibility on the agents you already run, because KPMG's Q2 2026 global sample found organisations with full real-time cost visibility were five times more likely to report established ROI, 15% against 3%.

Does connecting systems actually change the financial outcome?

The evidence says redesigning the work does, and connection is a precondition for that. McKinsey's State of AI in 2026, fielded 4 May to 8 June 2026 across 1,719 participants in 97 nations, found 37% attributing at least some EBIT impact to AI use, essentially unchanged year over year. The same survey found nearly three-quarters of high performers had fundamentally redesigned workflows because of their AI use, against about a quarter of everyone else. Wiring systems together without changing the workflow produces a faster version of the current process, which is why the EBIT line has not moved.

The work behind this

Fifteen engagements in the case library carry this capability, across CRM, ERP, ticketing and BI environments that were never designed to agree with one another.

Every one names the client where we are permitted to and states the measured outcome: Workflow and system integration, 15 engagements.