A customer asks whether they can move a booking without a fee. Your administrator checks an old email, searches a shared folder, then messages the owner. Meanwhile, a quote waits because someone needs to confirm what the standard package includes.
If that sounds familiar, start your AI project with the answers people keep chasing.
Organising trusted business knowledge is useful before you buy anything. It gives staff a clearer place to look and gives any future AI assistant better material to work with. The essential work is deciding what is current, who can use it, and who is responsible when it changes.
What knowledge engineering means for a small business
In “Rise of the Knowledge Engineer,” Mintlify describes a role that keeps company information consistent across people, systems and AI agents. Its responsibilities include establishing authoritative sources, assigning owners, setting permissions and review requirements, and measuring unanswered questions.
Mintlify sells documentation software, and its article speaks largely to businesses with technical documentation. It is a useful account of the work, not independent proof that a particular tool will save your business money.
For a small business, the practical version is straightforward: make the answers needed to run a process findable, trustworthy and maintainable.
That could cover what belongs in a quote, how to onboard a customer, which cancellation policy applies, or when an administrator needs approval. Some answers may live in inboxes. Others may exist only in a staff member’s head.
Mintlify also argues for assigning accountability before creating a new job title. An office manager or operations lead could own this work, with time allocated to it. They do not need to become the expert on every policy. They need access to the people authorised to settle questions.
Start with one recurring interruption
Resist the temptation to tidy every folder before doing anything useful. Choose one process where people repeatedly stop to find an answer.
Quoting is a good candidate when standard inclusions, exclusions and approval limits are unclear. Customer onboarding may be better if staff keep asking which forms, checks or welcome messages to use.
For a few working days, record the questions that interrupt that process. Write down where the answer was found and whether someone had to confirm it. Include the awkward cases, such as a long-standing customer with different terms.
Then separate three things:
- Approved rules: The current package definition, policy or procedure.
- Individual agreements: A specific customer’s accepted quote or agreed exception.
- Unresolved questions: Decisions someone authorised still needs to make.
Do not turn an old email into company policy because it is the easiest answer to find. If two sources disagree, record the conflict and ask the responsible person to resolve it. AI should not decide which commercial promise the business intended to make.
A hypothetical quoting handover
Imagine a small property-maintenance firm. This is an illustrative scenario, not a Valdris customer case study.
The office coordinator prepares quotes, while the owner approves unusual jobs. A shared-folder cancellation policy contains an older notice period. The current standard booking terms use a different one. A repeat customer also has a separately agreed arrangement in an email thread.
An assistant searching all three documents could return a plausible answer without establishing which terms apply. Even a correct quotation from the old policy would be wrong for a booking governed by the current terms.
Before testing an assistant, the firm assigns the operations manager to maintain the quoting guidance. The owner confirms the current standard terms. The manager marks the old policy as superseded, removes it from the active reference collection and keeps any required historical record separately.
The guidance tells staff to check the customer’s accepted terms before answering a cancellation question. The customer-specific agreement stays in the customer record, with appropriate access restrictions. It does not become a general rule.
For quote preparation, the same guidance names the current package description, required job details and approval route for exceptions. A new administrator can follow the handover without being expected to remember which colleague knows each answer.
This does not guarantee accurate AI output. It establishes the sources and boundaries against which staff can check it.
Give each answer enough context to use safely
A folder of polished documents can still leave people guessing. For the pilot, use a short record for each recurring question:
| Field | What to record |
|---|---|
| Question | The words staff or customers use |
| Approved answer | A short explanation with clear steps where needed |
| Applies to | Relevant service, customer group or booking date |
| Exceptions | What changes the answer and when to escalate |
| Source | The approved policy, agreement or procedure |
| Owner | The person responsible for keeping this record accurate |
| Dates | Effective date, last review and next review |
| Access | Who may read it and whether it is suitable for customers |
For quoting, “manager approval required” is incomplete unless staff know which manager, what information to send and what to do while waiting.
Ask someone unfamiliar with the process to try the record. Watch where they hesitate. Their questions may expose missing instructions that the experienced team no longer notices.
Keep the records in an existing tool if it supports the access and version controls you need. A maintained shared document or knowledge base may be enough for this test. New infrastructure is not a prerequisite for resolving contradictory instructions.
Put boundaries around AI access and answers
Once the reference material works for staff, test whether AI makes finding and drafting answers easier. In this pilot, give it a retrieval and drafting role: locate the relevant guidance, show the source and prepare a response for review.
Keep approval of prices, policy changes and exceptions with named people.
Use a small, approved collection rather than connecting every inbox and drive. Exclude unrelated customer records, staff information and sensitive commercial material. Before adding business data, check the selected tool’s terms, retention settings and access controls; do not assume it inherits the permissions on your original files.
Test access with the accounts people will actually use. An administrator should not be able to retrieve owner-only information through a more permissive assistant.
For customer-facing replies, require a person to check the applicable source before sending any price, refund, cancellation or delivery commitment. Keep automatic sending and record-changing actions outside this first pilot.
Ask the assistant to flag missing or conflicting guidance rather than fill gaps. Test that behaviour deliberately. An instruction to say “I don’t know” is not proof that it will do so reliably.
Run a bounded pilot and count the checking
A practical starting proposal is a two-week test of one process with a small group of staff. This is a suggested trial design, not a research-backed optimum.
Set a preparation-time allowance and a software spending ceiling before starting. Use tools you already have where suitable. If connecting an assistant requires a costly integration, test the improved reference material on its own first.
During the first week, record the existing workflow and prepare the approved answers. During the second, try the revised process. If you include AI, compare it with using those same improved documents without AI, so you can distinguish the value of tidier knowledge from the value of the assistant.
Record:
- Time from receiving a question to finding an applicable, approved answer.
- Time spent checking or correcting a draft.
- Questions requiring escalation because guidance is missing or unclear.
- Incorrect answers, especially those involving customer commitments.
- Time spent maintaining the reference material and resolving conflicts.
Use comparable questions, including exceptions, rather than testing only easy examples. Keep a short error log stating what happened and what had to be fixed. Count any actual correction expense as well as staff time. Do not translate faster drafting into revenue you have not measured.
Continue only if the process helps after review and maintenance effort are included. An answer that arrives quickly but takes longer to verify has not solved the administrator’s problem.
Make updates part of the process
A review date is a backstop. Also require updates when something relevant changes: a new price list, revised booking terms or a different approval limit.
The owner should update the approved record, deal with superseded copies and confirm that the assistant, if used, retrieves the revision. Keep customer-specific terms separate and preserve historical records where required.
When staff correct an answer, capture the correction for review. Otherwise, the person who fixed it becomes the next person everyone has to interrupt.
To start, ask Valdris to review one recurring quoting or onboarding process: where its answers live, who approves them, and what a small, controlled knowledge pilot would need.
Sources
- David Isquick, “Rise of the Knowledge Engineer,” Mintlify, 16 September 2026. Primary source for Mintlify’s definition, stated responsibilities and recommendation to assign ownership before creating a role. The pilot and small-business examples above are editorial recommendations, not results reported by Mintlify.
- Hahnbee Lee, X post sharing the knowledge-engineering argument, 17 September 2026. Source supplied for this article; the canonical Mintlify article is cited for the substantive argument.