Who to hire for Zoomin knowledge work on Salesforce

Ask ten people in a service organisation where the answer to a customer's question lives, and you will hear nine places. The manual is a PDF on a shared drive. The workaround is a Slack message from March. The policy is in a SharePoint library that three teams edit. The official article exists, but it is two releases out of date. An AI agent that is given all of that will answer confidently from whichever fragment it finds first. Hiring for Zoomin work is hiring someone to stop that happening.

#What was acquired and where it went

Salesforce completed the acquisition of Zoomin on 1 November 2024. The announcement said that "Zoomin enables the integration of unstructured data from various sources", and that Zoomin had been an AppExchange partner since 2018. Salesforce's announcement is the primary source for those statements.

The visible product result is Enterprise Knowledge powered by Data Cloud. Salesforce's own blog said of it: "These capabilities are the first milestone in integrating Zoomin’s technology to augment Data Cloud’s unstructured data pipeline post-acquisition." The same post described out-of-the-box connectors, with Zendesk, Google Drive and SharePoint named as places to begin, and a harmonisation engine that converts formats such as HTML and Word into a consistent form stored in Data Cloud. It also said edits and deletions in the source files sync through. Data Cloud has since been renamed Data 360, and Trailhead now lists a Salesforce Certified Data 360 Consultant credential. Read the Enterprise Knowledge post before any scoping call, so that you can tell the difference between what Salesforce ships and what a partner has built around it.

One caution. We did not find a Salesforce statement that maps every Zoomin feature to a named Data 360 feature, so do not accept a partner's claim that "Zoomin does X" without a current Salesforce document showing the equivalent today. The what happened guide lists what the announcements actually said.

#Skills and credentials that matter

For the Salesforce side, ask for the Salesforce Certified Data 360 Consultant, the Salesforce Certified Agentforce Specialist, the Salesforce Certified Platform Data Architect and the Salesforce Certified Platform Integration Architect. Each is listed on Trailhead's consultant, developer and architect pages. If the knowledge will feed a service agent, add the Agentforce Service Consultant for the people who design the service side.

The harder skills do not have exams. They are information architecture, which decides how documents are split, tagged and versioned, and retrieval evaluation, which measures whether the right passage comes back for a real question. A good candidate will talk about chunk size and overlap, metadata filters, citation, multilingual content and what happens to tables and diagrams in PDFs. A weak one will talk about the number of connectors.

#The corpus audit as a first engagement

Before a connector exists, ask the partner to audit what you hold. A sound audit produces a short, usable document, not a slide deck of aspirations. It should cover:

• An inventory of sources, with an owner, a last-edited date and an access model for each.
• A sample of documents classified by type: policy, product manual, how-to article, release note, contract, informal chat.
• A list of duplicates and contradictions, with a view on which version wins.
• The permission rules that must carry across, such as partner-only content, region-specific terms and anything covered by confidentiality.
• A set of fifty to a hundred real questions, drawn from cases and chats, with the document and passage that should answer each.
• A ranked list of sources to connect first, and those to leave out.

The question set is the most valuable output because it lets you measure retrieval before and after every change. It also gives you a fair way to compare partners, since each can be tested against the same questions. The implementation guide shows how the audit feeds a first connector.

#Where knowledge projects go wrong

Permissions are the hazard people discover last. If a SharePoint library restricts a file to the finance team, the retrieval layer must respect that restriction when an agent answers a question for someone else. Ask how the partner mirrors source permissions, how they test it with a user who should not see a document, and what happens when a person's access changes. This is a matter of data protection, not polish.

Freshness is next. Salesforce's post says deletions and changes sync, which is only helpful if the source is actually maintained. A partner should ask who retires superseded content, and should propose a visible review date on every document the agent may cite.

Then comes grounding. An agent answer that does not cite its source teaches users to distrust it. Ask whether answers show the passage they relied on and how a user can report a wrong answer, and who triages those reports.

Finally there is scope creep. Everyone wants the whole intranet connected. Connecting everything at once multiplies conflicts and makes the failures hard to diagnose. Start with the three or four sources that answer the most questions and expand only when the question set shows improvement.

#How the pipeline works, in plain terms

It helps to hold a simple picture of the machinery so that you can follow a candidate's explanation. A connector reads documents from a source such as a drive or a content system. A harmonisation step converts them into a consistent format. The text is split into passages, and each passage is turned into a numerical representation so that similar meanings sit close together. When an agent needs an answer, a retriever fetches the passages closest to the question, and the agent writes a reply grounded in them. Every stage can fail quietly. A connector may skip a folder, a conversion may strip a table, a split may cut a sentence in half, and a retriever may favour a long page over the short one that actually holds the answer.

That is why evaluation matters more than any single feature. A candidate who can describe which stage they would inspect first when an answer is wrong, and how they would tell, has done this before. One who answers every failure with "We will tune the prompt" has not. Ask them to walk through a real fault from a past project and the order in which they checked each stage.

#Service knowledge versus enterprise knowledge

Salesforce customers already have knowledge in more than one form. Salesforce Knowledge holds curated articles with authors, versions and publication states. Enterprise knowledge, in the sense of the Salesforce post, is wider: manuals, policies, product documentation and files that were never written as articles. The two behave differently and a good partner treats them differently.

Articles carry structure and an approval trail, so they are usually the first and safest thing to ground an agent on. Raw documents carry no such guarantees. A policy may exist in four drafts, a manual may describe a retired product, and a slide deck may contain a figure that was never approved for customers. Ask the partner how they decide what an agent may cite, and whether they would rather convert the best documents into reviewed articles than expose every file directly.

The answer also affects who does the work afterwards. If your service team already maintains articles, adding unstructured sources means a second editorial process. Make sure someone owns it, with time allocated, because an agent that quotes a stale policy does more harm than one that says it does not know.

#Ten questions to put to every candidate

1. Which unstructured-knowledge projects on Salesforce have you taken live, and what was in the corpus?
2. How do you audit a document set, and what does the output look like? May I see a sample with the client removed?
3. How do you measure retrieval quality, and which questions will you test with?
4. How are source permissions carried into retrieval, and how do you test that a restricted file stays hidden?
5. How do you handle versions of the same document, such as a manual for each product release?
6. What do you do with tables, images and scanned PDFs?
7. How do you deal with multilingual content?
8. Who retires or corrects content after go-live, and how do you make that person's job easy?
9. Which Data 360 and Agentforce capabilities have you used, and which things did you build yourselves?
10. What would make you advise me to leave a source out?

#Evidence and references

Ask to see a retrieval report: the question set, the passages returned, and the score before and after tuning. If the partner has none, they have not measured. Ask for a reference from the person who owns the knowledge, not the sponsor, and put this to them: how often does the agent cite something wrong, and how quickly was it fixed? Then run your own test. Give each shortlisted firm the same twenty questions and a small set of your documents, and compare what comes back. The differences are usually obvious within an hour.

Include some questions that have no answer in your documents. A trustworthy system says so, and an untrustworthy one invents a reply from the nearest passage. Include a question whose answer changed recently, to see whether the old version still surfaces. And include one that only a restricted document can answer, asked by a user without access. These three cases separate a partner who has built a demonstration from one who has built something fit to run in front of customers, and they cost nothing to set up.

#Commercial models and contract terms

Fix the audit first, as a short and separately priced piece of work, because its output decides the size of everything after it. Build can then be priced against the sources and the question set, with a stated pass mark. Tuning after launch suits time and materials, with a cap and a monthly review.

The statement of work should name each source and its owner, the permission model and how it is tested, the question set and the pass mark, who maintains content, and what the partner hands over. Salesforce licensing, including any consumption that Data 360 involves, is a separate conversation with your account team. Keep it out of the services quote, as the licence guide and the cost guide explain.

#When a small team can skip a partner

If your knowledge sits in a single, well-kept source, such as a Salesforce Knowledge base with a named editor, and you have an administrator who has completed the Data 360 and Agentforce paths, you may be able to run a first grounded agent in-house. Bring in a partner when content spans several systems, when permissions differ between groups, or when wrong answers carry regulatory or safety risk. Even then, you can keep the build in-house and buy only the audit and the evaluation. The migration guide is the place to start if you currently run Zoomin itself and need to decide what to do with it.