When AI output does not sound like your own team wrote it, the cause is usually not model capability but missing business context: which customers exist, what prices were quoted, which lanes are served, what rules the owner has set. Tencent's recent enterprise AI assistant markets "connecting enterprise context" as a core capability (reported by Silicon Star Pro on 2 November 2026; also covered by Jianshi from 31 October to 1 November 2026), which confirms the same reality: the model is strong, but it does not know your business. The practical order of work is to build enterprise context first and tune prompts second.
Why a strong model still produces off-brand output
Language models are general purpose and excel at producing sentences that every company in a category would write. Without enterprise context they fill gaps with generic probability, which produces three recurring failure modes: overstating qualifications, quoting against rules the company does not follow, and drifting tone over time. The model is not getting worse; it simply has no private information to narrow the distribution. Enterprise context exists to narrow general probability into what this specific company actually does.
Three frequently missed failure scenarios
First, over-promising. A dangerous goods carrier's first AI-drafted customer reply stated the company held qualifications for all dangerous goods categories, while in practice it covers only some. The model did not fabricate; it reproduced category-wide boilerplate.
Second, quoting logic mismatch. Whether existing and new customers share a rate, whether peak season surcharges apply, and which fees must be itemised often live only in the manager's and senior salesperson's experience, never documented, therefore not available to the AI.
Third, knowledge loss through staff turnover. When a senior salesperson leaves, their default rules leave with them; the reason earlier AI output was correct cannot be reproduced and the replacement starts over.
Build three tables before writing prompts
Sequence matters. Most companies start with dozens of pages of prompt templates and still get unstable output, because the enterprise facts the model needs are missing.
Table one, customer segmentation: tier, price sensitivity, transit time requirements, complaint history. It determines whether a given sentence may be said to a given customer.
Table two, quotation rules: price triggers, fee items that must be itemised, cases requiring escalation. It determines whether AI may quote on the company's behalf.
Table three, tone samples: the company's best emails, quotation notes and customer replies, quoted verbatim. It determines stylistic consistency.
The three tables need not be completed at once. Start with the most painful one, run it for two weeks, and backfill the sections where the AI errs. That is more realistic than exhaustively specifying rules up front, since most companies never fully documented their rules in the first place.
A simple test for what can be delegated to AI
Ask whether you can write down the decision criteria for the task. If you can, the task can be delegated; if you cannot, document the criteria first. Work whose criteria you cannot articulate will only have its confusion amplified faster by AI.
FAQ
If enterprise AI results are poor, should we switch to a stronger model? Usually not. Check enterprise context first; changing models does little for missing context.
Where should enterprise context be stored? Structured storage first (customer tables, rule tables), documents and prompts second; structured data is easier to maintain and reuse.
Who maintains the three tables? The business owner leads, frontline staff contribute, updated weekly. Maintenance delegated solely to IT tends to decay quickly.
Do small companies need this step? Yes. The more the rules depend on individuals, the more they should be documented, otherwise AI has no stable reference.