01
Benchmarks set the shortlist
Independent scores for intelligence, speed, and price narrow every lab’s catalog to a few candidates for the job.
Who decides what
You set the outcome. We set the standard. The evidence picks the model.
- The outcome, and by when
- You
- Your business sponsor brings it to intake, with the data and the constraints.
- The success criteria
- You and us
- Agreed before the first line of code: revenue, margin, hours reclaimed, adoption.
- The technical standard
- Us
- Set on every build: orchestration, evaluation, and vector store choices, and a clear call on what stays in house and what goes to a partner.
When we will tell you no
Saying where AI does not belong is part of the job. The discovery workshop maps where generative AI, agents, and data products create value, and where they don’t. We tell executives plainly what AI will, and will not, do for their business, and where a faster, cheaper model is enough, we use it.
Step by step
The engagement, from intake to playbook.
Six steps in four moves. The people who scope the work are the people who build it.
Move 01 Diagnose
Step 01 Intake
A conversation with the business sponsor about the outcome, the data, and the constraints.
From you The business sponsor, and the outcome you want, by when.
Step 02 Discovery workshop
Where generative AI, agents, and data products create value for you, and where they don’t.
From you The people who own the problem, and the data you have.
Step 03 Build-ready brief
A scoped brief, sequenced by dollar impact and risk, with success criteria set before the first line of code.
From you Agreement on the success criteria.
Move 02 Build
Step 04 Build
Forward deployed engineers embedded with your team turn the brief into a working system in a one to six week window: RAG pipelines, evaluation harnesses, agentic and MCP workflows.
From you Your own cases to test against, and time with the people who own the problem.
Move 03 Prove
Step 05 Rollout and hand-off
We stay through rollout and hand-off so the system is used, not just shipped.
From you The people who will use it, and the team that will run it.
Move 04 Compound
Step 06 Playbook
Every win becomes a reusable playbook and reference implementation, so the second deployment costs a fraction of the first.
From you The next use case on the roadmap.
Before it ships
What has to be true before it ships.
| Check | What has to be true | How it is met |
|---|---|---|
| Quality | Measured on your own cases, not assumed. | An evaluation harness built on your cases, run on every change. |
| Grounding | Where it answers questions, every answer can be checked. | Retrieval over your documents and data, with sources cited. |
| Cost and latency | Set before launch. | The model chosen per job on evidence: the strongest where the stakes are high, faster and cheaper models where they are enough. |
| Oversight | A person signs off where it matters. | Guardrails on agents that act inside your tools. |
| Data | Built to be used, owned, and measured. | The pipelines and data models underneath the system. |
Choosing the model
No default model. The job decides.
02
Your cases decide
An evaluation harness runs the shortlist on your own data and tasks, so quality is measured, not assumed.
03
Route by the job
The strongest model where the stakes are high, faster and cheaper models where they are enough, with cost and latency set before launch.
04
Re-run when the frontier moves
When a better model ships, we re-run the evaluation harness and switch if the numbers say so. A configuration change, not a rebuild.
What compounds
The second deployment costs a fraction of the first.
Playbooks and reference implementations compress time to production from months to weeks. Across a private equity portfolio, that is how the second company costs a fraction of the first.
A reusable playbook
A win in one team or company becomes the playbook for the next.
A reference implementation
Each build documented as a reusable pattern, ready for the next deployment.
One technical standard
Orchestration, evaluation, and vector store choices, held across builds.
Systems connected once
MCP servers connect your systems once, for every model you use, so a model switch is a configuration change, not a rebuild.
Questions
Questions, answered.
What does an engagement look like?
Intake with the business sponsor, then a discovery workshop, then a scoped, build-ready brief sequenced by dollar impact and risk. The build runs in a one to six week window, followed by hand-off. Success criteria (revenue, margin, hours reclaimed, adoption) are set before the first line of code, and we stay through rollout so the system is used.
How long does a build take?
After intake and a discovery workshop, we write a scoped, build-ready brief, sequenced by dollar impact and risk. The build then runs in a one to six week window.
Do you work with sponsors or with portfolio companies?
Both. We can diagnose across a portfolio with the operating team, and embed forward deployed engineers inside an individual company to build. See AI value creation for private equity.
Which AI models do you work with?
Whichever frontier model fits the job: Anthropic, OpenAI, Google, Meta Llama, Mistral, and open-weight models. We set the orchestration, evaluation, and vector store choices on every build. See every lab, compared.
What happens after launch?
We stay through rollout and hand-off so the system is used, not just shipped. Every win becomes a reusable playbook and reference implementation, so the second deployment costs a fraction of the first.
How do engagements start?
Most engagements come by referral. Tell us what you want to build on the get started page, or write to hello@promiseandproof.com.
Next
Tell us what you want to be true, and by when.