Skip to content
EncryptTechBD
Home Industries Our Work Blog Pricing About Contact Discuss Your Project
বাংলা (BN)
AI & Agentic Workflows 6 min read

AI governance for SMEs: problems and fixes

Troubleshoot the common failures of AI governance in small and mid-sized businesses, from shadow tools and slow approvals to model changes and agent actions.

E
EncryptTechBD · 0 comments
Small humanoid robot standing on a table
Photo: Jiuguang Wang / Wikimedia Commons

AI governance in a small business rarely fails because the policy is missing. It fails because the policy describes a company that does not exist: one with a central AI committee, a complete tool register and staff who wait for approval. In practice, someone signs up for a tool on a credit card, a team pastes a customer file into a chat window, and the first governance meeting happens after a complaint.

This troubleshooting guide covers the failures that recur in SMEs and the fixes that fit a small team. Each entry follows the same pattern: the symptom you notice, the likely cause, the fix and the prevention that stops it returning.

Symptom 1: The policy exists but nobody follows it

Likely cause: the policy was written as a legal document, not as a set of decisions staff make during the day. It bans vague categories such as "unapproved AI" without saying which tools are approved or what to do when a customer asks for something the approved tool cannot do.

Fix: replace the general prohibition with a short list of approved tools, the data each may receive, and a named person to ask when a new use appears. Publish the list where people work, not only in a shared drive.

Prevention: review the list monthly and add a one-page intake form. A policy that cannot be applied in five minutes will be ignored in five seconds.

Symptom 2: Nobody can produce a complete list of AI uses

Likely cause: AI arrived through individual subscriptions, vendor features switched on by default and experimental scripts. The inventory was built from invoices, which miss free tiers and embedded features.

Fix: ask each team to walk through its daily work and name every tool that generates, summarises, classifies or recommends. Include features inside existing platforms, such as support desk summarisation or accounting anomaly detection.

Prevention: add an AI-use question to procurement and to the change request for any new system. Make the inventory part of the annual system review rather than a separate exercise.

Symptom 3: Approvals are slow, so teams bypass them

Likely cause: every request goes to the managing director, or the approval group has no quorum and no service target. The process is designed for the highest-risk case and applied to low-risk tools.

Fix: create two paths. A low-risk path for tools that process public or internal non-sensitive content, approved by a team lead within a stated time. A high-risk path for customer data, employment decisions, credit, health or anything that acts without a human check.

Prevention: publish the response time and the escalation route. Measure how long approvals take; a process that is slow by design teaches people to avoid it.

Symptom 4: Outputs change after a vendor update

Likely cause: the model is a hosted service that changes without notice, and nobody recorded the version, prompt or evaluation set used at approval.

Fix: record the model name, version where available, prompt text, settings and a small set of test cases with expected behaviour. Re-run the tests after a material change and compare results.

Prevention: ask vendors for change notice and a deprecation window in the contract. Keep a fallback path for the most important use cases, even if it is a manual process.

Symptom 5: Customer or staff data appears in a prompt

Likely cause: the tool is convenient and the data rule is implicit. Staff do not know whether a customer's phone number, order history or medical note counts as sensitive, so they paste the whole record.

Fix: define data classes in plain language and give examples for each team. Configure the tool to block or mask sensitive fields where the platform allows it, and remove retention where it is not needed.

Prevention: include data handling in onboarding and in the quarterly refresher. Test the rule with a realistic scenario rather than a policy quiz.

Symptom 6: There is no record of what was approved or why

Likely cause: approvals happen in chat, decisions are verbal, and the rationale disappears when the person leaves. The organisation can describe the tool but not the judgement behind it.

Fix: create a one-page decision record for each use: purpose, data involved, affected people, human review, known limits, approver and review date. Store it with the system record.

Prevention: make the record a condition of access, not a follow-up task. Review the register quarterly and close uses that are no longer active.

Symptom 7: An agent takes an action nobody authorised

Likely cause: the agent was given broad credentials to make the workflow work, and the approval step was left to the prompt rather than enforced by the system.

Fix: restrict the agent to the minimum permissions it needs, require human approval for irreversible actions, and log the input, decision and result. Add a kill switch that a manager can use without engineering help.

Prevention: classify agent actions by reversibility and financial exposure. Test the approval path with a deliberate failure before the agent handles live work.

Symptom 8: The board asks for assurance and receives anecdotes

Likely cause: governance reports describe activity, such as the number of policies or training sessions, rather than evidence that risks are controlled.

Fix: report a small set of measures: the number of active AI uses, the share with a named owner and review date, the number of incidents, the time to disable a failing use, and the exceptions past their deadline.

Prevention: agree the measures with the board before the next reporting cycle and keep the definitions stable so trends mean something.

A 30-day triage sequence

  1. Days 1–7: inventory uses and identify the three with the highest potential harm.
  2. Days 8–14: freeze new high-risk uses, fix data exposure and restrict permissions on those three.
  3. Days 15–21: publish the two approval paths, the decision record and the tool list.
  4. Days 22–30: test one agent or automated decision with a failure scenario, then set the quarterly review calendar.

Our Business Automation division designs AI workflows with approval gates, logging and a documented fallback. The pricing page explains how assessment, pilot and rollout are scoped. For the procurement side of the same decision, read AI governance for SMEs: a buying guide.

Sources

Services that deliver this

Want this implemented properly?

Pricing →

Every engagement starts with a written scope. If a requirement is outside it, we quote it separately before starting.

Responses (0)

No comments yet. Add the first one below.

Leave a response

Your email address will not be published. Required fields are marked *.

Keep reading

More field notes