An internal AI platform sat untested while leadership drifted toward the safe default of buying Microsoft Copilot. We turned it into an evidence-backed executive decision to build, by treating the demo as a decision instrument, not a product showcase.
4.5 wksfrom engagement start to the executive decision to build at the AGM
Client
A global construction and project management consultancy
Led by
Blackstone& founding team member, prior engagement
Engagement
Delivery, Product leadership
Sector
Construction & Infrastructure
01 / Context
A stalled proof of concept and a decision drifting toward buy
Executive leadership was under pressure to settle a build-vs-buy decision on AI, and Copilot looked like the low-risk answer. Meanwhile an internal AI platform existed as a proof of concept: more than six features between 60 and 97 percent complete, built by a prior team, never tested by the product team, resting on an unvalidated data layer, and no stakeholder endorsements.
Behind the platform sat validated demand: 103 stakeholder interviews from an internal discovery programme, a core segment of more than 7,600 project managers within a potential user base of 17,000, and a funded risk-management systems programme that covered AI contract review.
02 / The Challenge
A fixed date and a non-technical audience
The annual general meeting on 15 May was the decision point. That left a four-and-a-half-week sprint to give a non-technical executive audience enough evidence to commit to building.
The cost of doing nothing was concrete: default to Copilot, and write off the platform investment along with the differentiated capability it represented.
03 / The Approach
Treat the demo as a decision instrument
We redefined the job first: arm the executive audience to make a defensible decision. Finishing the software could wait. We validated the data layer before polishing any feature; everything else depended on it.
Then we cut. The part-built features became three: a foundation, a differentiator, and a killer feature. Each was judged against a single test, could Microsoft do this, and features that failed it were cut even when they worked. The contract-analysis capability took the third slot over a further-along, revenue-tied alternative, and the three survivors were shaped into a live demo of under fifteen minutes. We replaced improvised demos with pre-tested golden queries and recorded fallbacks, and we secured endorsements from the business users who would live with the platform.
The cut
04 / Outcomes & Impact
The executives committed to build
At the AGM, the executive audience committed to build over buy. At the start of the engagement the platform was untested, its data layer unvalidated, and no stakeholder had endorsed it. It is now a funded direction.
A build phase followed as a direct continuation of the engagement.
05 / Capability Transfer
The working practices stayed with the team
The handover covered more than the three polished features. The working disciplines went with them: the feature-triage process, the test-suite discipline, and the evaluation loop that scored every feature against the questions the business actually asks. Those practices continued with the client team into the build phase.
By the numbers
0
Stakeholder interviews behind the validated demand the decision stood on
6 → 3
Part-built features cut to three, demoed live in under 15 minutes
17,000
Potential users across the business the platform now builds toward
How we made the call
The decisions behind the result
Anyone can list what was built. What matters to a buyer is the judgement: what we chose, what we rejected, and why.
Cut the features that argued for the competition.
Anything Copilot could match strengthened the case for buying Copilot. Three polished features with a clear gap beat six shaky ones, so a working presentation feature was cut.
Choose the headline feature for its decision value.
The contract-analysis capability beat a further-along, revenue-tied alternative: a stronger Copilot gap, a funded programme connection, a named stakeholder, and its own self-contained dataset.
Validate the data layer before polishing anything.
Every feature depended on the index beneath it. Thin data would have made the flagship assistant a generic chatbot regardless of polish.
Pre-test every demo moment.
A live failure in front of the executive audience was the single highest-impact risk. Pre-tested queries with a recorded fallback for every feature removed it.
* This case study reflects work delivered by a Blackstone& founding team member prior to the formation of Blackstone&. Client identity withheld; details generalised to protect confidentiality.
See yourself in this story?
Technology is the science. Partnership is the art. Let’s write the next one together.