The reason why artificial intelligence (AI) has been the focus of endless keynote speeches, product roadmap announcements by ERP providers and development efforts across all categories of business software is that AI-embedded systems promise (and will ultimately deliver) software that more closely molds its operation to the needs of a specific enterprise and individual user. The technology is making business software more powerful and easier to use. This is a major breakthrough after decades of enterprises and individuals having to adapt to the limitations of the systems.
That noted, the current state of AI is primitive relative to what will evolve and become commonplace by the end of the decade. That’s not a bad thing, even though many are wondering if this is just another overblown technology that overpromises and underdelivers. However, the cautious approach to AI taken by business software providers is a product of lessons learned. Mindful of past technology failures such as Microsoft’s ill-fated original Clippy and Apple’s Newton, software providers are balancing the needs of a competitive rush to be first with all things AI against creating user frustration and reputational loss from half-baked technology.
Much of what has been offered in embedded AI functionality supports limited use cases. These have the advantage of being straightforward and are workable by most organizations without addressing foundational data challenges or significant change management requirements. While these advances may appear to be trivial, they are important for building trust in the technology. These modest first steps are also useful in training individuals to have AI-first thinking, similar to how people came to learn how to use the internet. Moreover, since much of what people do in finance, accounting, supply chain and operational roles is trivial and repetitive, even these limited use cases will have a notable impact on productivity as enterprises adopt them. Agents, in particular, will be an important delivery mechanism of AI technology.
Even so, while there’s been a lot of attention to AI in all of its forms, it’s not hard to miss that actual utilization of AI is relatively low. We’re at that stage of technology adoption best described as everyone is talking about it, but nobody is actually doing it. Yet, if you’re thinking that we’re in a hype cycle, you’re missing the wave that is only starting to build. As in the past, business software arrives well ahead of enterprise adoption, partly because of caution but also because organizations that use the technology have to adjust operations and attitudes to take full advantage.
ISG Software Research asserts that by 2028, almost all providers of ERP software will have incorporated AI to reduce workloads, speed processes and reduce errors. Agents are useful in this regard because they enable business software providers to extend and amplify the
As we pass the third anniversary of ChatGPT’s unveiling, it’s worth taking stock of advances in embedding AI in ERP systems over the past year, and specifically what’s happened with agents. Much of what is in production today is at a basic level of functionality, useful first steps, but with limitations.
In 2025, ERP providers began releasing agents that combine conversational interfaces with automation conforming to governed workflows. Finance and accounting agents draft journal entries, explain variances, summarize invoices, match purchase orders and orchestrate approvals with user-defined guardrails. Purchasing agents monitor catalogs and scrutinize contracts, suggest suppliers, create requisition orders and coordinate multi-way matches. Supply chain agents continuously monitor actuals to plans, respond to execution signals, flag shortages and able-to-deliver issues, along with suggested resolutions. Service and manufacturing agents translate demand signals into work orders, along with parts reservations and technician schedules. Project-centric agents answer natural-language questions about actual and forecast margins, asset utilization and availability.
In addition to embedded functionality, an increasing number of providers offer agent “studios” to compose task flows, connect APIs and create domain-specific copilots. These are typically low- or no-code designers, and the best offer methods to adjust the underlying code for fine-tuning the agent’s functionality and processes. To build trust and avoid reputational damage, agent autonomy is confined to situations that are near certain, with limited consequences in the event of an error. Today’s agents default wherever necessary to propose actions that are then executed to conform to an enterprise’s explicit policies.
For 2026, provider roadmaps are emphasizing efforts to expand the scope of end-to-end autonomy, multi-agent collaboration and deeper domain specialization. Providers are talking about agents that span entire processes, such as record-to-report, order-to-cash and procure-to-pay, but siloed data and process complexity will limit execution to relatively simple process environments. Coordinating handoffs across finance, supply chain, projects and service without users stitching steps together is theoretically possible and convincingly demonstrable, but individual enterprises will need to assess feasibility under specific circumstances. Negotiation-capable sourcing agents will run competitive events, evaluate risk and draft terms. Working-capital agents will steer cash by dynamically timing payments and collections. Manufacturing and service will see agents tied to digital twins, simulating schedule changes before executing on the floor, and field agents that predict entitlements, book parts and verify resolution with sensor feedback.
One aspect of the outlook for embedded AI that remains vague is pricing. For now, it’s safe to say that nobody knows what they’re doing, and there’s a lot of testing going on. At this stage of market evolution, many providers are willing to include basic functionality that doesn’t consume significant development or operating costs into the base subscription price, but even that isn’t universally the case. Providers are charging (or planning to charge) for specific capabilities because their value is beyond trivial and they consume not insignificant amounts of compute resources, but there is still a lot of uncertainty about the details. Providers are tinkering with various plans in the hope they will spur adoption, provide a sufficient return on investment, do not condition the market to expect free services and do not interfere with closing a sale.
Cloud-based ERP accounts for a very large majority of new ERP sales. On-premises accounts for a small percentage of new sales and a steadily declining minority of the installed base in midsize and large enterprises’ ERP systems. Incorporating AI in a legacy, on-premises ERP system poses a significant challenge because the process and data customizations typical of such deployments can make creating even simple agents a relatively costly, bolt-on effort. The gap between embedded AI functionality available in current cloud-based ERP systems and the cost and effort to duplicate this in an on-premises system will grow, increasing pressure to migrate to the cloud through the end of the decade.
I strongly recommend that finance department executives embrace a fast-follower approach to adopting AI generally, and embedded AI capabilities in particular. Enterprises that have on-premises ERP software should evaluate whether and when they will migrate to the cloud. In the process, weigh the costs and risks associated with migrating an ERP system against the ongoing costs of maintaining an on-premises deployment, along with the initial and ongoing costs of bolting on AI and agent-like functionality. Part of this assessment should consider the future expanding operational productivity deficit resulting from not having embedded AI-related capabilities in off-the-shelf cloud ERP.
Regards,
Robert Kugel