An owner of a fifteen-person services firm has been recording client calls for the last eight months. The transcripts sit in a folder somewhere inside her video conferencing platform. Last Tuesday a new account manager landed a question about how the firm has handled a particular contract clause for a specific client segment in the past. The honest answer is that the firm has handled it well, twice, in conversations that happened in March and September of last year. The account manager will not find those conversations. Nobody is going to scroll eight months of transcripts. The precedent is sitting on disk and is effectively invisible.
This is the hidden knowledge base. Many owner-operated firms hold thousands of unread conversations across Slack, Teams, email, and increasingly call recordings. The answers to the recurring “how do we handle this” questions have already been written down dozens of times. AI tools have finally made that archive retrievable, but only if the access design is proportionate, the accuracy risk is understood, and the starting corpus is small enough to govern.
What lives in the typical SME message archive, and why is none of it currently retrievable?
A services firm with ten to thirty staff typically holds somewhere between one thousand and twenty thousand unindexed conversations across Slack, Teams, email folders, and call recordings. These conversations were created to solve immediate problems, not to become a knowledge asset. Channels were named for projects that have since ended, email is filed by sender and date rather than by question type, and call recordings sit in a folder almost nobody opens.
The reason none of it surfaces when a new question lands is structural, not technological. Traditional email and chat search work on exact keywords, and the new person handling the question does not know which keywords appear in the relevant thread. Nobody searches “how did we handle the clause about liability caps for a tech-startup client” because that exact phrase is not how anyone wrote about it at the time. The precedent exists, the conversation is on disk, but it is invisible to the person who needs it.
Why does this matter for your business?
It matters because every novel-seeming question your team solves from first principles is work the firm has already paid for once. McKinsey’s 2025 State of AI survey identifies knowledge management as one of the most-experimented AI functions, with the majority of firms still stuck in pilots rather than scaled retrieval. The opportunity is concrete: faster onboarding, fewer repeated commercial mistakes, and senior judgement being captured rather than re-extracted from individual inboxes every quarter.
The capability that has changed in the last eighteen months is threefold. Semantic search treats “how much legal risk are we taking on” and “client contract limitations” as the same conversation, even when the words differ. Multi-thread summarisation extracts the decision points from a forty-message email exchange and presents them as a paragraph. Pattern detection across hundreds of calls surfaces the recurring objections, common misunderstandings, and team members who consistently handle particular client types well. None of this required your team to tag, label, or curate the archive in advance.
Where will you actually meet the proportionate access design?
You will meet it the first time your operations lead asks whether the new sales hire can search the entire Slack history. The answer is that they should not be able to. The archive holds genuinely useful material on proposals, objections, and delivery problems. It also holds commercial pricing, staff performance discussions, and personally identifiable information spoken aloud on calls. Access policy is the first question, not the last.
A proportionate setup for a small firm has three layers. Role-based access controls determine who can retrieve what, with junior staff seeing procedural and project material while pricing strategy and HR-adjacent conversations sit behind a tighter gate. A retention schedule defines what gets deleted when, defensibly grounded in the ICO’s storage-limitation principle: six months for raw call recordings, twelve to eighteen months for transcripts, two years for internal decision threads, and seven years for client communication aligned to tax and contract dispute windows. An exclusion list keeps automated notifications, marketing emails, vendor newsletters, and personal back-channel chat out of the indexable set entirely. Most of the work is policy, not configuration.
When should you trust a retrieved precedent and when should you ignore it?
Trust a retrieved precedent when it is recent, when the surrounding context has not materially changed, and when it is being used as a starting point for human judgement rather than as a final answer. Ignore it, or at least verify it, when the conversation is more than twelve months old, when it touches anything regulated, or when the firm’s commercial position, risk appetite, or standard terms have shifted since the conversation took place.
This is the subtle failure mode of AI retrieval, and it gets a lot less attention than the privacy risk. A confidently-surfaced fourteen-month-old email thread on how to structure a contract clause reads as current guidance even when the regulatory landscape has moved on or your firm’s policy has changed. The AI does not flag staleness on its own. The mitigation is partly technical, retrieval systems should rank recent conversations above old ones for the same query, and partly organisational, one named person needs to spot-check what the archive is surfacing and retire guidance that no longer holds. Without that ownership, knowledge management systems degrade into unreliable repositories inside a year, which is the failure pattern Agility Portal’s practitioner research describes in detail.
What does a starter pattern look like, and what comes next?
The starter pattern that works for owner-operated firms is a deliberately small corpus, a named owner, and a defined process. Index the last six months of recorded client calls, the last twelve months of decision channels in Slack or Teams, and the last eighteen months of email discussion threads where commercial or operational reasoning was worked through. That produces somewhere between two and ten thousand documents, which is governable. Indexing the eight-year archive is not.
The timeline for getting this working is roughly three to six months of elapsed effort, not a year-long programme. The first month is policy and access design, deciding what gets indexed, who can search what, and what the retention schedule is. The second month is onboarding the initial corpus and verifying that retrieved results are relevant. The third month is embedding retrieval into the workflow people already use, so that relevant precedents surface inside the CRM, the email client, or the messaging tool rather than requiring a separate search interface. Months four to six are about adoption, accuracy review, and refining what is in or out of the indexable set. The work is in the policy and governance, not the technology. The proportionate starting point is small enough to govern, large enough to be useful, and disciplined enough to compound.
If you would like help scoping the first corpus for your firm and writing the access and retention policy that goes with it, book a conversation.
