3 Things AI: The “Don’t Make the Assistant Guess” Edition

1. Put client context where the handoff happens

A 20-person advisory firm can still lose an afternoon to one client update. The account lead checks email, a consultant searches meeting notes and someone else opens the project plan. An assistant can draft the update, but it is guessing if the current decision, budget and client feedback live in different places.

Asana made Asana Client Management generally available on September 28. The company says it combines a client directory and portal with meeting summaries, assigned actions, capacity and budget views, plus four customizable AI Teammates that work from the account’s shared record. That is a product announcement, not independent proof that every agency will save time.

The benefit is that a draft, risk flag and client-visible status can draw from the same current information. The tradeoff is concentration: permissions, retention and client visibility become more important, and a shared record can spread an error if nobody owns corrections.

Choose one low-risk account and list the five facts people repeatedly hunt for. Put only those fields and one week of work into a trial workspace. Ask a teammate who did not set it up to prepare the normal update, then compare the time, missing facts and corrections with the current process. Availability does not establish fit with your contracts, Canadian privacy obligations or existing tools, so test information quality before considering a migration.

2. Give a browser agent a path it can actually follow

An online wholesaler wants an assistant to check supplier stock before staff promise a delivery date. A person can work around a shifted button or a slow page. An automated browser may click the wrong element, miss a warning or fail silently when the layout changes.

Cloudflare announced a September 28 update to its Kitesurf browser for AI agents. Kitesurf now supports WebMCP, a way for websites to expose named functions to agents instead of making them imitate clicks. Cloudflare also says Kitesurf has full coverage across its Browser Run API and passes more than 730,000 Web Platform Test subtests.

The benefit is a more explicit route: an agent calls an offered function and receives structured information rather than inferring meaning from pixels. The tradeoff is that WebMCP is emerging, websites must expose useful functions, and a callable tool is risky when permissions are too broad. Test coverage is encouraging evidence, not proof that your supplier portal works.

Start with a read-only task. On a site you control, expose one function that returns a test product’s availability without customer or payment data. Run 25 checks including an unavailable item, a delayed response and an unexpected field. Record wrong answers, retries and the evidence shown to the reviewer. Cloudflare’s figures are vendor-reported, and sites without compatible tools still require page interpretation; authentication and final-action approval remain your responsibility.

3. Keep the record before deciding what to remember

A service manager asks why a warranty exception was approved six months ago. The weekly summary says it happened but omits the supplier email and its condition. A shorter memory made everyday questions faster, yet the detail needed for this decision disappeared during compression.

A new research paper, Just-In-Time Agent Memory, posted September 28, proposes keeping complete raw histories in a structured store with short navigational summaries. For each request, a research component retrieves and assembles relevant evidence instead of relying only on a summary prepared before anyone knew the question. The authors report stronger benchmark results than ahead-of-time memory approaches while using fewer resources than earlier trained agent-memory systems.

The benefit is that rare but important details remain available when the question changes. The tradeoff is more retained data, more retrieval work and a larger privacy and deletion obligation. Keeping everything without access limits can turn useful history into unnecessary exposure, and runtime research can add latency.

Pick 20 questions where staff already know the right answer. Compare the summary-only assistant with a small prototype that searches approved records and cites the passage used. Score accuracy, evidence quality, response time and whether retrieval surfaces material that should have been deleted or restricted. This is early research, not a production guarantee; benchmark gains may not transfer to your documents or controls.

The bigger pattern

Client work improves when context travels with the handoff, browser work when the interface is explicit, and long-running assistance when compression does not destroy the original record. More context is not automatically better; it creates cost, access and maintenance duties. A reasonable option is to make one answer less dependent on guessing, then measure whether accuracy improved enough to justify the added system.

Where does your team lose the most time today because an assistant—or a person—has to reconstruct missing context?