The model worked. The pilot results were clean, the outputs were accurate, and Finance signed off on the methodology. The Chief AI Officer presented the findings to the board in Q3. By Q4 the deployment contract was signed. Fourteen months later the system was still running in a staging environment. The planners who were supposed to use it had rebuilt their Excel workarounds. Nobody flagged this to the board.
According to BCG, 74% of enterprise AI projects produce no measurable value. That number gets cited constantly and almost never gets explained. The cause is not model quality or a shortage of data scientists. The cause is a contract structure problem.
Most enterprise AI programs are designed to succeed in a sandbox and delivered by vendors with no outcome-staked fees tied to what happens in production. The firms that solve this run 12-week cycles with AI agents + named experts, structured so the delivery team cannot collect their performance tranche without pushing through to production.
They operate as an AI Native Operating Partner, not a T&M vendor. Most enterprise buyers have not worked with one yet, and the deployment contract they signed is the reason.
What the pilot actually tested
A pilot tests the model. It does not test the integration. Most enterprise AI pilots run on clean historical exports from the ERP, not live transactional data with schema inconsistencies and access-permission conflicts inside an SAP or Oracle environment.
They run in a controlled setting where the team manages the inputs. They do not test what happens when the allocation agent hits an exception type the training data never included.
The four failure points between sandbox and production are not model failures. Integration: connecting the agent to live ERP, TMS, or WMS data in a production schema. Exception handling: building logic for scenarios the agent was not trained on, which is where most real decisions happen.
Adoption: getting the planners who run the workflow to act on the agent’s outputs instead of working around them. Value verification: agreeing with Finance on what success means before week two, not after go-live. A pilot that succeeds on clean historical data has not tested any of these. Most deployment SOWs are written as though it already did.
Why T&M contracts stop at failure point three
A vendor on a time-and-materials contract has no financial incentive to drive a program through all four failure points. The SOW is written around deliverables: design documents, an integration specification, a go-live date. Exception-handling logic and planner adoption are not deliverables. They are what happens after the SOW scope ends.
Integration risk lives with the client. Exception logic is the operations team’s problem once the vendor hands over. Planner adoption is a change management task that starts where the vendor’s contract ends.
This is not about vendor ethics.
Vendors deliver what the contract defines. A T&M contract defines completion, not outcomes. So the team reaches go-live, the planners find workarounds, the go-live log gets archived, and the board hears about it when someone asks why the program never scaled. The pilot-to-production gap is a contract design problem, not a technology problem.
Eight questions to ask before a pilot becomes a deployment SOW
Before the deployment contract is signed, these questions separate vendors built to solve the four failure points from those designed to hand them back.
Start with proof of production. Can the vendor show a live deployment on real ERP data, not a demo environment? Ask for the go-live log, the exception history, and the name of the senior who was accountable for the outputs. If the answer is a slide deck, that is the answer.
Then go to ownership. Who owns exception-handling logic when the agent encounters a scenario outside its training data? What is the documented escalation path and the SLA? Who owns the integration work if the live schema does not match the pilot export?
These are questions about liability, not methodology, and vendors who hedge on them are telling you something. Then the adoption loop. What does the planner feedback cycle look like in weeks two through eight, and does it have the authority to change the agent’s logic mid-cycle? Planner adoption that is left to the client after handover is not planner adoption. It is hope.
Finally, the money. What percentage of the vendor’s fee is contingent on a Finance-validated production outcome, not a go-live milestone? What happens contractually if the agent recommendation is wrong and the client acts on it?
What is the exit clause if the program stalls between failure point two and three, and how much of the fee is already collected by then? The answers to those three questions tell you more about the vendor’s confidence in their own model than any reference call will.
How Tranche 3 changes what gets prioritized in weeks six through ten
A delivery team with a performance tranche staked on a Finance-validated production outcome behaves differently in weeks six through ten than a team on T&M. The SAP integration work that costs more than it bills is not optional.
The exception-handling logic that slows the sprint becomes the sprint. Planner adoption is not a post-handover task. It is something the delivery team owns because the outcome-staked Tranche 3 does not clear without it.
This is the structural effect of outcome-staked commercial terms: they realign who carries the production risk. An AI Native Operating Partner running 12-week cycles with AI agents + named experts has a direct incentive to push through all four failure points, because the contract makes it economically irrational to stop short of production.
Future Works structures its engagements this way, with outcome-staked fees built into the contract before the engagement starts.
The 74% statistic will not improve until deployment contracts are designed around the failure points, not just the pilot results. The pilot tells you the model works. The contract tells you who is responsible for making it work in production. Those are different questions, and in most enterprise AI programs, only one of them gets asked.
If you are selecting an AI deployment partner, the eight questions above are where the evaluation should start. Start at Future Works.
Matt Leta, CEO, Future Works.
