A $4.5M AI Units commitment, rebuilt at one seven.

The brief

The client is a mid-market industrial distributor with two hundred and ten million dollars in annual revenue and operations across the United States and Canada. The business had selected GROW with SAP for a phased ERP migration scheduled across an eighteen-month window. The deployment scope covered seven hundred and eighty named users across finance, procurement, sales, and warehouse operations.

The trigger for engagement was the second-round commercial proposal from SAP, which followed an executive workshop on agentic AI capabilities. The proposal included an AI Units commitment built around an aggressive Joule adoption forecast across the deployed user base, sized at four point five million dollars over the three-year contract term.

What was at stake was an unmeasured commitment to a consumption-based product line in a deployment context where no actual consumption pattern had yet been established. The AI Units commitment, if locked at the proposed level, would have produced a substantial overhang of unused commitment if Joule adoption fell short of the forecast — and consumption-based commitments are not refunded.

The CIO and head of finance jointly authorised an independent commercial review before any commitment was made to the AI Units line, while progressing the underlying GROW deployment on a parallel track.

The opening proposal

SAP's AI Units forecast assumed a per-user Joule consumption rate drawn from the SAP product team's reference deployment benchmarks, applied uniformly across all seven hundred and eighty users. The forecast resulted in a year-one consumption estimate of approximately three million chargeable AI Units, scaling to four point eight million in year two and five point four million in year three.

The proposed commercial structure was a three-year prepay covering the full forecast at the standard list rate, less a deployment-incentive discount. The four point five million dollar headline reflected this prepay model and was framed as the structure required to secure the discount level.

The reference deployment benchmarks SAP relied on came from large-enterprise customers with significantly different operational profiles. The distributor's workflow analysis showed that only a subset of the seven hundred and eighty users were realistic Joule adopters in years one and two, and most would touch a small handful of agentic workflows rather than the full Joule surface.

The defence

1. Workflow-based consumption recount

We reconstructed the AI Units forecast against the distributor's actual deployment workflows, restricting the chargeable consumption to documented use cases. Procurement assistant flows, finance-period close summarisation, and sales quote support were modelled at conservative volumes. The recount produced a year-one consumption estimate of seven hundred thousand AI Units, not three million.

2. Population segmentation

Of the seven hundred and eighty named users, approximately one hundred and ninety were realistic AI adopters at go-live based on role and workflow exposure. We segmented the population accordingly and modelled separate consumption curves per segment rather than the uniform per-user rate SAP had applied.

3. Pilot-then-scale structure

We replaced the three-year prepay with a structured pilot followed by milestone-based scale-up. Year one was sized to the recount; years two and three were preserved as commitments only if specific adoption milestones were hit, with downward flexibility if they were not.

4. Unused-unit rollover and conversion

A rollover mechanism was introduced for unused units within a contract year, and a conversion option was added that allowed unused AI Units to be redirected against other GROW line items at a defined exchange ratio. Both clauses removed the downside of forecast inaccuracy.

The settlement

Settlement closed at one point seven million dollars in three-year AI Units commitment, against an opening proposal of four point five. The reduction was approximately sixty-two per cent. The structure was a pilot-and-scale model with documented adoption milestones gating the year-two and year-three step-ups.

Five contractual protections were written in: a documented workflow-based consumption model in an annex; a per-segment consumption rate rather than a per-user rate; a rollover for unused units within a contract year; a conversion option for unused units against other GROW line items; and a written-out renegotiation right if adoption fell more than thirty per cent below the model.

Total elapsed time from the second-round proposal to executed commercial agreement was ten weeks. The GROW deployment proceeded on the original timeline and the AI Units commitment is now sized to actual workflow exposure rather than a vendor-supplied benchmark.

The lessons

• AI Units forecasts built from SAP reference benchmarks routinely overstate consumption in mid-market deployments by a factor of three or more. A workflow-based recount is the first commercial defence step.
• Per-user consumption rates are misleading for AI Units. Realistic adopter populations and per-segment rates produce far more accurate forecasts than uniform per-user models.
• Three-year prepay structures for consumption-based products are the highest-risk commercial framework in mid-market AI procurement. Pilot-and-scale structures with milestone gates remove most of that risk.
• Rollover and conversion clauses materially change the economics of any consumption commitment. Their absence in a standard SAP proposal is a negotiation prompt, not a fixed term.
• Adoption-milestone language should be written by the buyer, not the seller. SAP-drafted milestones routinely include unmeasurable success criteria that default to the seller's interpretation.

Cross-reference these lessons against the firm's SAP GROW topic page, the GROW Buyer Handbook white paper, and the broader analysis in the GROW Licensing reading cluster.