Coding tools: review the change, not the confidence
AI coding assistants now extend beyond completing a line of code. Tools such as GitHub Copilot and Cursor can help explain repositories, propose edits and support multi-file work, depending on the product and plan. They are particularly useful for repetitive changes, test scaffolding and navigating unfamiliar code. They can also introduce plausible but incorrect assumptions at considerable speed.
Evaluate them on accepted changes that pass your normal review and test process. Measure review time, defects, security findings and maintainability alongside implementation speed. Do not use generated lines of code as a productivity metric. A smaller, clearer patch may be much more valuable than a large automated rewrite.
Restrict access to secrets and production credentials. Review repository indexing, retention, model-provider terms and organisation-level controls before enabling the tool. Run proposed commands in a sandbox where possible. Require explicit approval for installing dependencies, changing infrastructure or accessing external services; the development environment is a powerful tool surface.
Enterprise search: permissions are part of relevance
Enterprise search assistants, including Glean and knowledge features within existing workplace suites, aim to answer questions across scattered company information. Their value depends on connector coverage, indexing freshness, source quality and permission enforcement. A fluent summary of an obsolete policy is still a bad answer.
Test with real access boundaries: an employee, a manager and a contractor should not receive the same results when their permissions differ. Verify that access changes and document deletions propagate within an acceptable period. Check whether citations support each material claim and whether the assistant admits when the available evidence is incomplete.
Begin with a curated collection rather than connecting every shared drive. Remove duplicate policies, assign content owners and record effective dates. Search AI exposes information-management problems; it does not automatically solve them. Include content maintenance and connector administration in the business case, not just the per-user subscription.
Browser agents and voice systems need clear limits
Browser agents can interact with interfaces that lack convenient APIs. They may help transfer records, gather public information or prepare routine submissions. Their flexibility comes with fragility: page layouts change, sessions expire and visual elements can be misidentified. Use isolated browser profiles, domain allowlists and explicit confirmation before purchases, external messages or account changes.
For repeatable work, prefer a stable API when available. It offers clearer inputs, errors and permission boundaries. A browser agent can be a useful bridge for legacy systems, but should not silently become the only route through a critical business process. Keep a manual fallback and evidence of what was changed.
Voice AI combines speech recognition, conversational reasoning and speech generation. Potential uses include appointment handling and basic service triage. Test end-to-end latency, interruptions, accents, noisy environments and transfer to a person. Tell callers when they are interacting with AI, and assess applicable recording and privacy requirements before retaining audio. Natural-sounding speech does not establish identity; use a separate verification process for sensitive requests.
Interoperability helps. It does not remove trust decisions.
The Model Context Protocol, or MCP, standardises aspects of how AI applications connect to tools and resources. It can reduce bespoke integration work. It does not certify that a connected server is trustworthy, make every tool safe or replace authorisation. Review server provenance, permissions, deployment settings and update practices before granting access.
| Tool category | Primary success measure | Release gate |
|---|---|---|
| Coding | Accepted, tested changes | Review and security checks |
| Search | Correct, supported answers | Permission and freshness tests |
| Browser | Verified task completion | Action approval and recovery |
| Voice | Correct resolution or handoff | Disclosure and identity controls |
Use a four-week evaluation with a named owner, a small user group and a fixed spending cap. Capture a baseline before introducing the tool. Record both time saved and new work created, including reviews, corrections and administration. Ask users which tasks became easier and which they stopped trusting.
Commercial models may combine seats, premium requests, minutes, storage and usage credits. Confirm overage behaviour and cancellation terms before rollout. Adopt only when the tool passes your quality and security gates and delivers a repeatable benefit. Keep promising but immature capabilities on a watchlist; deciding “not yet” is a valid technology strategy.
Back to the system
Explore all five layers of business AI ↗