AI-assisted ERP implementation

AI is doing real work in ERP programs in 2026, but it is cutting work packages, not programs. Vendor and consultant claims put testing effort down 60 to 85%, and design and configuration effort down 80% at one named customer. No SI has published a completed program where AI halved the total bill. The widely quoted McKinsey figure (effort down "at least 50 percent") comes from a May 11, 2026 article that describes what agents "could" do, with no finished program behind it. SI fees are 40 to 60% of implementation cost, so a 50% cut in SI effort is worth 20 to 30% of your budget at most. Price the claim at that level, then put five questions in the contract.

AI inside the product and AI used to build the program are two different purchases

The sales meeting mixes two things. The first is AI that ships inside the software you will run after go-live: SAP Joule, Microsoft Copilot in Dynamics 365 Finance and Supply Chain, the agents in Oracle Fusion Cloud ERP, Workday Illuminate, Sage Intacct's close and AP agents. Those belong in the license negotiation, and they do nothing for the implementation bill. The second is AI the SI uses to deliver the program: test generators, mapping assistants, configuration agents, and knowledge tools. Only the second kind can lower the hours you pay for.

The pages that rank for this topic cover the first kind. Clockwise Software prices in-product AI additions at $5,000 to more than $100,000 by approach. Ultra Consultants and CMARIX describe forecasting, anomaly detection, and document matching inside the ERP. None of the three discusses what an agent does to an SI's staffing plan. This article is about the second kind.

Four work packages have agents in production, and each one still ends with a person signing off

The table below lists what we could find in use at named customers or available to buy as of November 2026. Every entry is a vendor or consultant statement, and the claimed savings come from the party selling the service.

The pattern across all four rows is the same. The agent drafts, a consultant checks, and the client signs. Sage and PwC said plainly that consultants validate AI proposals and that formal metrics on time or cost savings "have not yet been disclosed." Oracle's own release says expert reviewers examine the completed configuration, request changes, and approve it. The vendors are not claiming autonomy, and Opkey's CEO told CIO.com that fully autonomous operation is likely years away.

The rows also show what agents do not touch. Nothing in the table shortens the decisions a steering committee has to make, the data cleanup the business has to do, or the training the workforce has to absorb. McKinsey's own article names change management, not technology, as the main constraint on future ERP roadmaps. In our experience a tier 1 program needs 36 months from decision to go-live, and most of those months are spent on exactly the work the agents skip.

The McKinsey "half" is a projection, and the number that reached you lost its qualifiers on the way

The claim you heard in the sales meeting almost certainly comes from one source. McKinsey's "The end of ERP as we know it?" was published May 11, 2026. It says AI agents could reduce implementation effort by at least 50 percent, cut program duration by half, shrink design and build from six to nine months down to two to three, cut testing effort by about 80%, and cut training preparation by up to 90%. It also says most large enterprises spend between $100 million and $1 billion on an ERP migration, with payback often taking four to five years.

By the time the article reached trade and regional press, the conditional language was gone. Pulse wrote that "McKinsey's data indicates" a 50% reduction, and Asianet reported that AI "reduces" cost and duration by at least two times. Neither linked to the original. The original names no completed program, no sample, and no method. It is a point of view from a firm that also sells ERP transformation work, and it should be read that way.

BCG, which also sells ERP program work, published its own estimate a year earlier, in April 2025, and landed at 20 to 40% less effort for an organization-wide implementation. BCG's phase-level figures are closer to what the vendors in the table above are claiming: testing down 60 to 70%, requirements gathering down 30 to 60%, custom code development down 30 to 40%, and data cleansing and mapping down 20 to 30%. BCG also wrote that several of its examples came from its own ERP program or a single client. The honest range for total program effort, from the two firms most quoted on the subject, is 20 to 50%, and the lower end has more evidence behind it.

A 50% cut in SI effort is worth 20 to 30% of the program

Even if you accept the McKinsey number in full, it applies to effort, not to your budget. ERP Research puts SI fees at 40 to 60% of implementation cost. The rest is software subscription, internal staff and backfill, data cleanup, change management, infrastructure, and the dual-run period. Agents do not reduce the subscription, and they do not reduce the months your finance team spends cleaning vendor master data.

The arithmetic is short. On a $5 million single-region tier 1 program, SI fees are $2 million to $3 million. If the SI halves its hours and passes every dollar through, you save $1 million to $1.5 million, or 20 to 30% of the program. If the saving is BCG's 20 to 40% of effort instead, you save $400,000 to $1.2 million, or 8 to 24%. That is a real number worth negotiating for, and it is not half.

Two conditions have to hold for even that number to arrive. The SI has to pass the saving through rather than keep it as margin, and the saving has to show up in the work packages that are actually on your critical path. Design is 10 to 20% of implementation hours, so an 80% cut in design effort moves the total by 8 to 16 points at best. E13 covers why the quote has not moved; this article covers how to make it move.

The rest of the market is announcements, previews, and two-person startups

The gap between what is in production and what is being presented is wide, and it is widest at the top of the market. Oracle announced AI Success Navigator with six agents on October 26, 2026, and did not state availability or pricing. The Fusion Setup and Configuration Registry underneath it is a public Git repository that is available now, which is a meaningful first step. The release, though, quotes Accenture, Deloitte, PwC, Grant Thornton, and Apex IT on what it will do, not on what it has done. Microsoft's Project Mia, the agent effort for Dynamics 365 Finance and Supply Chain implementations, was in private preview with partners in May 2026 with a public preview slated for later in the year.

Opkey's Design Studio, launched February 18, 2026, says early users saw Oracle and Workday project timelines shorten by as much as 50%, with no sample size or method given. The startups are earlier still. Trope, in Y Combinator's Summer 2026 batch, has two employees and scans configurations to flag gaps between what was promised and what is being built. Tato raised a $5 million seed round in October 2025 to capture meeting decisions and project chatter into a searchable record. Both are useful ideas for program oversight, and neither configures anything.

Gartner put a number on the sector's froth in June 2025. It predicted that more than 40% of agentic AI projects would be canceled by the end of 2027 for rising costs and unclear value, and estimated that only about 130 of the thousands of vendors calling their products agentic were genuine, a practice it labeled "agent washing" (Reuters via The Star). None of this means the tools are worthless. It means the burden of proof sits with the party quoting the saving.

The vendors are not equally far along, and the right choice depends on which product you are implementing. In our judgment, Oracle Fusion customers have the strongest position, because the vendor itself now maintains the configuration registry the agents run on.

Oracle NetSuite customers in the software industry in North America can buy SuiteExpress, announced October 26, 2026. Its agents prepare data for migration and recommend next steps under consultant oversight. NetSuite claims implementations "in weeks, or even days," gives no cost figure, and says other industries wait until 2027.

SAP customers get the most from Joule for Consultants when the SI has a large junior bench whose questions it answers. Sage Intacct customers in the mid-market get the clearest live example of agent-assisted configuration and testing, through PwC. Microsoft customers should expect Project Mia to reach general availability before relying on it. For Workday, Opkey is the only implementation-side agent claim we found.

For Infor CloudSuite, the only announcement is from a partner, not the vendor. Fortude said on August 6, 2026 that its agents draft implementation documents and automate repeatable build steps, with consultants checking outputs, and it published no numbers. For IFS Cloud, a search of IFS's 2026 announcements and ERP Today's coverage of its H1 2026 results found agents inside the product (Digital Workers, IFS.ai) and nothing on agents used to implement it, as of November 4, 2026. On Infor and IFS, any AI saving an SI quotes rests on the SI's own tooling, and the baseline question below matters most.

The AI section of the proposal and the hours table rarely match

Expect the AI content of an SI proposal to sit in the AI section, with the staffing plan behind it built the way staffing plans were built before agents existed. The reasoning is simple. An SI prices on hours, its tooling is new, and it carries the risk if the agents underperform, so it has every reason to describe the agents in prose and leave the hours alone. The reader's job is to force the two to meet.

Suppose an SI proposes 14,000 hours and says its testing agents cut test effort by 60%. Testing is commonly planned at 20 to 25% of implementation hours, so the test package is 2,800 to 3,500 hours and a 60% cut removes 1,680 to 2,100 of them. That is 12 to 15% of the total, and at a $200 blended rate it is $336,000 to $420,000 off a $2.8 million bill. If the proposal still shows 14,000 hours, the 60% is an illustration, not a price.

The second thing to watch is the shift in mix. When a package does get cheaper, it tends to shed junior hours and keep senior ones, so the blended rate rises and the bill moves less than the hours do. KORE1, a staffing firm, puts 2026 boutique SI rates at $150 to $250 an hour for Dynamics 365, $150 to $275 or more for NetSuite, and $200 to $325 or more for SAP.

Take a Dynamics 365 bid of 10,000 hours: 6,000 junior hours at $150 and 4,000 senior hours at $250 come to $1.9 million at a blended $190. If agents remove 40% of the junior hours and none of the senior ones, the bid falls to 7,600 hours and $1.54 million. Hours are down 24%, the bill is down 19%, and the blended rate is up 7% to $203.

Opkey's own description of Design Studio says consultants shift toward advisory work, and ABeam's account of Joule for Consultants describes junior staff validating their approach before escalating. Both are good for quality. Neither is a discount unless the contract caps the blended rate.

Five questions to put in the contract

The questions below belong in the RFP and the statement of work, not in a conversation after selection. Each has a weak answer that should tell you the AI claim is a slide, and a strong answer you can hold the SI to.

Question 1 does most of the work. An SI that can show the baseline is one that has measured its own agents on a prior program, and that is the only kind of evidence this market has produced so far. An SI that cannot show it is asking you to carry the risk of its tooling at full price.