B1The financial case in one page
- The alternative is far more expensive. Building only an AI assistant with mobile would cost an estimated $0.5M–$2.2M and take 12–18 months. Rebuilding the full platform would cost $6M–$25M and take 24–36 months.
- The worst case is small and bounded: the license plus a short pilot. A failed internal AI build typically costs $500k or more and a year.
- No subscription trap. The license is perpetual. Updates (18%/year from Year 2) are optional, and if we stop paying, we keep what we have.
- The return is countable: hours avoided, tickets deflected, onboarding shortened, future hiring avoided. Each can be baselined before and measured after.
B2The cost of doing nothing
Waiting feels free because no one sends an invoice for it. It is not free:
- The hours keep leaking. Every month, our people keep searching, asking and waiting the slow way. On our numbers, that is about – per month in recoverable time.
- Our knowledge gets harder to organize. Documents pile up without governance, systems get replaced and half-migrated, and experienced people leave without their knowledge being captured. Starting later means starting a harder job.
- Competitors compound. A competitor that started a year ago is not a fixed year ahead. Their knowledge base has been improving all year and keeps improving.
- Late programs get rushed. A program that starts under pressure tends to cut corners on its foundation to show quick results, which is how pilots fail in public.
B3Options considered
| Option | Up-front cost | Time to first use | Main risk |
|---|---|---|---|
| 1. Do nothing | $0 | n/a | Ongoing lost hours; knowledge loss; falling behind |
| 2. General AI subscriptions only (per-seat) | [subs_cost] | Days | Not grounded in our knowledge; recurring per-seat cost; no ownership |
| 3. Build an AI assistant in-house (ingestion, RAG, chatbot, admin, mobile) | $0.5M–$2.2M | 12–18 months | Delivery risk; 15–25%/year maintenance; engineers diverted from our core |
| 4. Build the full platform equivalent | $6M–$25M | 24–36 months | Large program risk; architecture still to be discovered |
| 5. Another AI platform vendor | [vendor_quote] | [vendor_time] | Usually subscription, closed source, vendor-held knowledge index |
| 6. Trilivy.AI (recommended) | – | Days to weeks | Needs our AI Data Expert, plus a certified integrator or our IT team (Part A4) |
Options 2 and 6 can coexist. Trilivy.AI connects our existing assistants to our knowledge, which makes the subscriptions we already pay for more useful.
What building option 3 would cost, line by line
| Component we would have to build | Western agency | Nearshore agency |
|---|---|---|
| Document ingestion (web, SharePoint, email, SQL, code) with preprocessing and RAG search | $400k–$900k | $200k–$450k |
| AI agents, tool connections (MCP), REST API | $250k–$600k | $120k–$300k |
| Chatbot: streaming, citations, uploads, voice, human handoff | $150k–$350k | $80k–$180k |
| Admin and operations: connections, users, login, backup/restore | $100k–$250k | $50k–$120k |
| Mobile app, iOS and Android (hybrid; native $250k–$400k) | $60k–$150k | $40k–$90k |
| Total | ≈ $1.0M–$2.2M | ≈ $0.5M–$1.1M |
Rough estimates at Western $100–200/h and nearshore $50–90/h. They do not include Content Studio, Course Maker or training content. Maintenance after launch usually adds 15–25% of build cost every year.
What rebuilding the full platform would take
| Scenario | Cost | Team · duration |
|---|---|---|
| Rebuild from full specifications | $4M–$7M | 15–20 people · 18–24 months |
| Recreate from zero (strong product organization) | $6M–$10M | 20–25 people · 24–30 months |
| Fortune 500-style global systems-integrator program | $15M–$25M | 30–40 at peak · 24–36 months |
The estimate is about 450–650 person-months of specialized effort. It is an independent replacement-cost model, not a quote from any named firm and not a valuation of Trilivy.AI. AI coding tools make writing code cheaper, but they do not remove the architecture, integration, testing and product decisions; that is the journey buying lets us skip (see C6).
B4What Trilivy.AI costs
One price, everything included
The license covers the whole platform: ingestion and retrieval engine, chatbot, agents, REST API and MCP, administration, the bonus add-ons (Content Studio, Course Maker, mobile apps, Talk to a Person), the Workshop training and the full source code. There are no editions to compare.
What the 18% updates cover, and what they do not
Included
New platform versions, compatible features, security and maintenance patches, updated documentation and training, upgrade guidance. The first 12 months are included in the license.
Not included (budget separately)
Custom development, new ERP/CRM integrations, on-site consulting, 24×7 support or SLAs, and cloud / LLM / embedding usage charges.
If we do not renew, we keep perpetual rights to the version we have; we just stop receiving new releases. For comparison, SAP has publicly described its Enterprise Support at 22% of net license value.
The other costs, so nothing is hidden
- AI model usage (Claude, GPT, Cohere or a self-hosted model): [llm_text]. This varies with volume, and a self-hosted open-source model can cut it substantially at high volume.
- Hosting: [hosting_cost_text].
- Standing up and running the platform ([impl_mode]): [svc_text].
- AI Data Expert time (mandatory): [ade_pct]% of one FTE. This is the single biggest predictor of success.
B5Where the return shows up
The return does not appear as one line labelled "AI value". It appears as changes in numbers we already track:
| Value driver | How we measure it | Baseline today |
|---|---|---|
| Knowledge-worker hours avoided | Time to find an answer; repeated questions to experts | [base_time] |
| Tickets / contacts deflected | Questions resolved without a person; escalation rate | [base_tickets] |
| Faster onboarding | Weeks for a new hire to reach productivity | [base_onboard] |
| Future headcount avoided | Volume handled per person as the business grows | [base_volume] |
| Cycle time shortened | Days for reviews, proposals, approvals, research | [base_cycle] |
| Partner spend made productive | Consultant hours spent on outcomes, not rebuilding basics | [base_partner] |
On headcount: the case is to avoid future hiring, not to replace the people we have. The value is the hires we do not need to make when volume grows. Some roles will change as routine work is absorbed, and we will say so openly.
The cost curve bends the right way
The first use case pays for the foundation: ingestion, governance and retrieval quality. The second costs less, because it reuses that foundation, and by the tenth a new capability costs close to nothing. A tool-by-tool approach pays for the foundation again every time, so its cost per capability never falls.
B63-year cost, benefit and return
| Our inputs | Value | Our inputs | Value |
|---|---|---|---|
| Employees using it | [users] | License (one-time) | – |
| Hours saved per user per week | [hours] | Updates from Year 2 | [upd]% |
| Loaded cost per hour | $[rate] | Integrator / IT setup, one-time | $[svc] |
| Working weeks per year | [weeks] | Hosting per year | $[host] |
| Realization factor | [real]% | AI model usage per year | $[llm] |
| Hires avoided per year (from Y2) | [hires] | AI Data Expert time | [ade_pct]% FTE |
| Loaded cost per hire | $[hirecost] | AI Data Expert loaded salary | $[ade_sal] |
| Year-1 ramp-up | [ramp]% |
| Year 1 | Year 2 | Year 3 | 3-year total | |
|---|---|---|---|---|
| License (one-time) | – | – | – | – |
| Updates | included | – | – | – |
| Integrator / IT setup | – | – | – | – |
| Hosting + AI model usage | – | – | – | – |
| AI Data Expert time | – | – | – | – |
| Total cost | – | – | – | – |
| Hours recovered (value) | – | – | – | – |
| Hiring avoided | – | – | – | – |
| Total benefit | – | – | – | – |
| Net value | – | – | – | – |
Method: hours value = users × hours/week × weeks × hourly cost × realization factor, with Year 1 scaled by the ramp-up. Hiring avoided counts from Year 2. Payback is computed month by month, with one-time costs at month 0. Benefits not modelled (cycle time, onboarding, partner productivity, use cases 2 to N) are upside.
B7Financial risk, and how it is contained
| Risk | Containment |
|---|---|
| We pay and it does not work for us | Start with one narrow use case and measure at day 60 before expanding. The worst case is the license plus a short pilot. [pilot_terms] |
| Recurring costs creep up | Perpetual license; updates are optional at a fixed 18% policy rate; no per-user fees. |
| Vendor disappears or changes direction | We hold the full source code and our data stays in our own databases. The platform keeps running without the vendor. |
| AI model prices rise | The model is a replaceable component. We can switch provider or run a self-hosted open-source model by configuration. |
| Nobody owns the knowledge | A named AI Data Expert is a precondition of the project, not an option. |
| Adoption is low | A named business owner, a pilot audience, 10–25 suggested prompts, and the Workshop training tracks for every audience. |
B8Questions Finance will ask
"– is a lot for an AI tool we have not tested."
It is one-time, with no per-user fee. Compare it with one senior engineer for one year (about $150k–$250k fully loaded), or with a year of per-seat AI subscriptions for [subs_seats]. The pilot plan in Part D tests it on our own data before we expand.
"Why not wait until the AI market settles?"
The part that moves fast, the model, is the part Trilivy treats as replaceable. The part that takes time, organizing and governing our knowledge, does not depend on the model. Waiting postpones the slow part without lowering its cost.
"How do we price this against per-token vendor quotes?"
A per-token rate tells us how a vendor meters compute. It tells us nothing about the return. We evaluate on business outcome versus total cost of ownership, as in B6.
"What is the real ongoing cost?"
Updates (optional, 18% of the license from Year 2), hosting, AI model usage and our AI Data Expert's time are all in B6. Premium support or SLAs, if wanted, are priced separately: [support_terms].