D1Recommendation
We recommend that [company] license Trilivy.AI for – one-time, and deploy it first as [usecase] for [usecase_audience], with a go / expand decision at day 60.
The evidence in Parts A to C supports this recommendation:
- The problem is real and costly. Our knowledge is scattered, and generic AI cannot reach or govern it (Part A2). On our own numbers, recoverable time is worth – per month (Part B2).
- Buying beats building. – against an estimated $0.5M–$2.2M and 12–18 months to build an assistant with mobile, or $6M–$25M for the full platform (Part B3).
- The return is measurable. 3-year net value –, ROI –, payback – (Part B6).
- The technology holds up in the vendor reviews (A- overall, A for the retrieval core), and our own verification checklist stands at – confirmed (Part C0.4).
- The risk is bounded. Perpetual license, full source code, a replaceable AI model, our data in our databases, and a narrow first deployment.
- Upside not counted: Content Studio, Course Maker, mobile apps and Talk to a Person come with the platform (Part C5).
D2The 90-day plan: crawl, walk, run
We start narrow, prove value and expand only on evidence. The platform is built to be the foundation from day one, even though the first application is small.
| When | Stage | What happens |
|---|---|---|
| Week 0 | Approve | Sign license; name the owners (D4); Trilivy provisions our dedicated server; our AI Data Expert and team start the Workshop training. |
| Weeks 1–2 | Crawl | Our AI Data Expert selects and ingests the approved documents for [usecase]; create the agent instance; add 10–25 suggested prompts; record the baseline. |
| Weeks 3–4 | Validate | Test with 30–50 real questions, including missing-answer, conflicting-document, expired-content and unauthorized-access cases. Run the Part C checklist. |
| Weeks 5–8 | Pilot | Release to [pilot_group]; collect feedback; fix knowledge at the source. |
| Day 60 | Decide | Measure against the success criteria (D3). Go, adjust, or stop. |
| Days 60–90 | Walk | Wider audience; ingestion schedules; monitoring; escalation route (Talk to a Person); pick use case #2. |
| Later | Run | More applications on the same foundation; connect our existing assistants via MCP; agent actions introduced in stages (explain, recommend, prepare for approval, act). |
D3Success criteria at day 60
| Measure | Baseline | Target |
|---|---|---|
| Questions answered correctly with the right citation (test set) | n/a | [t_citation] |
| Time to find an answer | [base_time] | [t_time] |
| Repeated questions / tickets to [repeated_q_team] | [base_tickets] | [t_tickets] |
| Pilot users active weekly | n/a | [t_active] |
| User usefulness rating | n/a | [t_rating] |
| Unauthorized-access test cases leaked | n/a | 0 |
| Running cost (hosting + AI usage) per month | n/a | ≤ [t_cost] |
Expansion rule: we advance when the application meets its quality and access criteria, users find it useful, its cost is understood, and our AI Data Expert can maintain its knowledge.
D4Who owns what
Buying technology does not assign responsibilities, so we assign them here:
| Role | Responsibility | Name |
|---|---|---|
| Executive sponsor | Sets priorities, funds the work, resolves conflicts, decides on expansion | [sponsor] |
| AI Data Expert (mandatory) | Owns the knowledge: knows where it lives, decides what is authoritative, gets it handed over and keeps it current. No project without this person. | [ade] |
| Business owner | Defines the task, audience, acceptance criteria and outcome | [business_owner] |
| Knowledge owner | Approves sources in their domain, resolves content conflicts, manages validity dates | [knowledge_owner] |
| Platform delivery: [impl_mode] | Stands up and operates ingestion, retrieval, integrations, deployment and monitoring | [impl_name] |
| Security & governance | Permissions, appropriate use, audit, action boundaries | [sec_owner] |
D5Risks and mitigations
| Risk | Mitigation |
|---|---|
| No one owns the knowledge | A named AI Data Expert with [ade_pct]% of their time allocated, before signing |
| No one can run the platform | An external certified system integrator or our own IT team, named in D4 |
| Our documents are messy or out of date | Start with ~50 well-known documents; problems surface early and are fixed at the source |
| Confidential content reaches the wrong people | Chunk-level metadata filters, token-scoped access, and unauthorized-access tests before release |
| Wrong answers damage trust | Citations on every answer, strict mode, curated Q&A for critical facts, Talk to a Person escalation |
| Skills gap on the stack | Workshop training tracks, certified integrator support, full documentation, coding agents on the source (C6) |
| Vendor or AI-provider dependency | Full source code; replaceable model; self-hosted option |
| Scope creep | Crawl–walk–run gates; one use case at a time against its own criteria |
D6The decision options
| Option | What it means | Consequence |
|---|---|---|
| A. Approve (recommended) | License Trilivy.AI (–), start the 90-day plan | First grounded answers in weeks; decision on evidence at day 60 |
| B. Approve with conditions | Approve, subject to [decision_conditions] | Same plan, with a gate before signing |
| C. Defer | Revisit in [defer_months] months | About – in unrecovered time per month; a larger, messier corpus later; competitors keep compounding |
| D. Build in-house | Staff an internal program | $0.5M–$2.2M and 12–18 months before first production use |
D7Closing statement
Most of the success or failure of enterprise AI is decided before anyone writes application code. It depends on whether the company's own knowledge is organized, governed and reachable. That is the part that takes time, and it does not depend on which model wins next year.
Wherever we stand today, this works for us. If we are new to AI, we skip the years of pilots that others have already paid for and start with the foundation done right. If we already have pilots and vendors, this is the layer that unifies them, stops the duplicated knowledge copies, and makes the assistants we already own more useful.
We would not be handing the organization a plan to admire. Our executives, product owners, technical team and early adopters would get a working platform in their hands from the first week, running on our own documents, with sources they can check. And we would own it: the source, the data and the freedom to extend it.
Waiting is not free. Each month we wait costs us about – in lost time, and the knowledge we will need keeps growing and leaving with people. We can begin with a bounded, measurable step, and the downside is small and known.
Generic AI answers questions. AI grounded in our data answers our questions.
We ask the committee to approve Option A.
D8Sign-off
Decision taken: [decision]
Signature and date:
Signature and date:
Signature and date:
Signature and date: