Every organisation, regardless of industry, runs on a gap. The software is there: an ERP for finance, a CRM for sales, a ticketing system for support. But the work that moves between those systems still happens inside human heads and email threads. Someone reads a vendor message, figures out what it means, cross-checks a figure in a spreadsheet, decides whether it requires escalation, and then acts. That person is not inefficient. They are doing something that software, until recently, genuinely could not do.
Bain & Company has mapped this gap and valued it at $100 billion in the US market alone. Including Canada, Europe, Australia, and New Zealand brings the total to around $200 billion. Vendors have captured roughly $4 to $6 billion of this so far, meaning more than 90% of the opportunity remains untouched. That is not a gap in awareness. That is a gap in capability that is only now beginning to close.
The reason traditional automation tools failed here is simple. Rules-based systems, the kind that have existed for decades, can only follow scripts. They break the moment a situation is ambiguous or when information is spread across systems that do not natively talk to each other. AI agents are different because they can read context, reason across multiple sources of information, and act within a set of boundaries that a company defines. They do not need every scenario pre-mapped. They handle variation.
The distribution of this opportunity across business functions tells a clear story. Operations and cost-of-goods functions represent about $26 billion of the addressable market, largely because workforces are large and even modest automation rates across a big headcount add up fast. Sales represents around $20 billion, driven by sheer volume of employees rather than any unusual automation potential. Customer support and engineering functions sit at 40 to 60 percent automation potential, meaning a large share of tasks in those areas can realistically be handled without human involvement. Finance falls in the 35 to 45 percent range. Legal sits at 20 to 30 percent, held back not by technical limits but by the cost of errors.
That last point matters. Bain identifies six factors that determine how automatable a workflow really is, and consequence of failure is one of them. A support ticket that is mishandled is a bad customer experience. A tax filing that is mishandled is a regulatory problem. This is why the same AI capability gets deployed differently in different functions, and why high-stakes areas will keep humans closer to the process even as automation handles the volume.
What makes this shift strategically significant is what it does to the pricing model of enterprise software. For twenty years, companies paid per user account, per login, per seat. That model assumed value scaled with headcount. AI agents do not need accounts. They do not log in. When one agent handles the work of ten people, the old pricing model collapses, and vendors who have not moved to outcome-based or usage-based pricing are already seeing the pressure in renewal negotiations and seat count reductions.
The early evidence of this shift is already in the market. Companies like Glean, which coordinates employee requests across multiple systems rather than searching a single database, has reached $200 million in annual recurring revenue. Sierra, which resolves customer issues across enterprise systems rather than within a single ticketing tool, has crossed $150 million. These are not large companies by traditional standards, but their growth rates signal that enterprise buyers are willing to pay for cross-system capability, not just single-system efficiency.
For anyone managing a team whose job involves connecting one internal system to another, the question is not whether this automation will arrive but how quickly their organisation's data and processes are in a condition to support it. Agents need structured, documented information to work from. The informal knowledge that experienced employees carry in their heads, knowing which supplier to call, which exception to flag, which approval to seek, is precisely the knowledge that has not been captured in any system. Organisations that begin documenting these patterns now will be in a far better position when they go to deploy or procure these tools.
Bain's own timeline is blunt: the window for software companies is measured in quarters, not years. The same urgency applies to the enterprises that will be buying, deploying, or being disrupted by these tools. The gap is closing fast, and the companies compounding data from real deployments today will be materially harder to displace in twelve months.