Enterprise AI Agents and Knowledge Bases: How to Split the First-Year Budget
When enterprises build AI agents and knowledge bases, what runs out of control is rarely the technology choice. It is the budget structure: most of the money goes into hardware up front, the data is never governed, no scenario goes live, and by year two nobody uses it. The workable approach is to split the first-year budget into three parts and fund them in phases.
Part one: compute and models, three routes
| Route | Form | Public price reference | Best for |
|---|---|---|---|
| Public cloud API | Per-token billing, elastic | Prices fell sharply after the April 2026 adjustment; new pricing from August 2026 introduces peak and off-peak rates | Volatile volume, low sensitivity, scenario validation first |
| All-in-one appliance | Integrated hardware and software with pre-installed models | Public product lines run from entry to enhanced tiers, priced from several hundred thousand to several million RMB | Data that cannot leave the intranet, fast go-live |
| Self-built cluster | Servers and accelerators purchased directly | No single public benchmark; estimated per card type and scale | Long-term heavy load, existing operations capability |
The appliance market has matured into tiers. Telecom operators, for example, offer training-and-inference and inference-only variants in entry, general and enhanced configurations, and some vendors now support single-machine inference of very large models on domestic GPUs. When choosing, separate training-and-inference from inference-only: for knowledge base Q and A and workflow agents, an inference machine is usually enough.
Part two: data and knowledge governance
This is the most underestimated item. Retrieval-augmented generation depends on corpus quality, and the real work sits in four tasks: document cleaning and de-duplication, chunking strategy, permission tiers for who may ask what, and content refresh. Miss any one and the system typically answers the wrong question or cites stale facts. Policy places governance up front as well: the Cyberspace Administration of China's guidelines on deploying large models in government affairs, issued in October 2025, stress systematic planning, intensive development and secure, prudent rollout, with governance as pre-deployment work.
Part three: scenario development and integration
The third part is scenario work: tool calling and workflow orchestration for agents, integration with existing systems such as OA, ERP, ticketing and customer service, plus permissions and auditing. The public tender market confirms the size of this demand. In the first half of 2026 there were 2,770 application-type large model contract awards in China, 60.7% of all large model awards, up from 58% for the whole of 2025, showing that procurement is shifting from buying models to buying working scenarios. For comparison, the first half of 2025 saw 1,810 awards worth RMB 6.4 billion, and the government sector grew fastest: in 2025 government large model awards rose 593% in count and 275% in disclosed value, ranking second by count after education and research and first by value.
Three-phase budget allocation for year one
| Phase | Duration | Main spend | Judgement |
|---|---|---|---|
| Validation | 1–2 months | Scenario selection, small sample data, proof of concept | Can one real question produce a usable answer |
| Pilot | 3–6 months | Data governance, knowledge base build, first scenario live | Real user activity and resolution rate |
| Rollout | 6–12 months | Multi-scenario replication, operating routine, cost optimisation | Cost per resolved question falling, departments covered |
Four common mistakes
- Buying hardware before choosing a scenario: the appliance sits for three months while the first use case is still undecided
- Skipping governance and dumping documents in: retrieval quality collapses and users abandon it after three tries
- Treating the Agent as a project rather than an operation: go-live marks the end, with no content refresh or effect measurement
- Leaving data and model ownership out of the contract, so the boundary around data, prompts and fine-tuning output stays unclear
One recommendation: spend most of year one on data governance and getting scenarios live rather than on one-off procurement. Models and compute are the adjustable variables; the knowledge base and the operating routine are the constants that decide success.
FAQ
- What does a first-year AI knowledge base cost?
- Three parts: compute and models (metered public cloud API, or an all-in-one appliance publicly priced from a few hundred thousand to several million RMB), data and knowledge governance, and scenario development plus integration.
- Cloud API or appliance?
- Depends on data sensitivity and volume. Metered APIs got very cheap after April 2026, but peak-and-off-peak pricing introduced in August 2026 makes cost modelling necessary. Appliances keep data on-premises at publicly listed prices from several hundred thousand to several million RMB.
- Should knowledge governance come first?
- Yes, and it is the most underestimated item. Retrieval-augmented generation depends on corpus quality: document cleaning, chunking strategy, permission tiers and refresh mechanisms are all required.
- How should the first year be split?
- Three phases: one to two months of scenario selection and proof of concept, three to six months of data governance and first scenario go-live, and six to twelve months of replication and operating routine. Weight the budget toward pilot and operations rather than one-off hardware.