Define the strategic signal proactive agent

A proactive agent is an AI assistant that does assigned work on its own, on a schedule or when something changes, within boundaries you set. You do not prompt it each time.

It needs a named role, such as Operating Review, and a one-paragraph purpose statement. The purpose keeps it from drifting into collecting news. Its job is to surface changes that matter to your current priorities.

Example for Northline, a 300-person maker of commercial refrigeration units:

Operating Review signal agent. Before each monthly operating review, check our approved sources and report only changes that affect our three current priorities: entering the Spanish market, protecting margin against Coldex, and securing compressor supply. Output feeds the signals section of the monthly operating review pack. Do not summarise general industry news.

The last two sentences matter most. Write down which decision or review each output feeds. If you cannot name the meeting or decision, the agent does not yet have a job.

You choose how the agent runs.

• A single run is one pass. Use it to prepare one specific review, or to test the setup before you rely on it.
• A continuing brief repeats on a schedule, for example every Monday before the monthly review, until you pause or retire it.

Start with a single run. Read the output with your reviewer. Ask: did it use only approved sources, did it report the right changes, did it ignore what it should? Fix the purpose, sources or criteria, then run again. Move to a continuing brief only once results are reliable across a few runs.

Northline's CEO ran the agent once before the March review. It reported a Spanish trade show date change that was irrelevant, so she tightened the criteria. The second single run was clean, and only then did she set it to run weekly.

A continuing brief that starts with a weak setup repeats the same mistakes every week.

This course uses two platforms for proactive agents: OpenAI Dots and Claude Duet. Pick one using three questions, in this order:

1. Approval. Has your organisation approved this platform for business use? If only one is approved, that settles it.
2. Connections. Can it connect to the sources you plan to use, such as your shared drive, CRM or a regulator's public site, and is that connection permitted?
3. Data location. Where is your data processed and stored, and does that meet your policies and any client commitments?

Do not assume a feature exists because it did in another tool or in a demo. Confirm current scheduling options, connection types and data handling in the platform's own documentation and with your IT or security lead. Features change, and this course cannot promise what either platform offers on the day you build.

Write your choice down with reasons, for example: Claude Duet: approved by IT in January, connects to our SharePoint, data stored in the EU region per our contract. Confirmed with Head of Security.

An approved source is one you have permission to use and connect. Keep the list short; five to eight is plenty for one agent. Every extra source adds changes someone must review.

For each source, record four things:

Source	Owner	Covers	Changes	Permission
Spanish energy-efficiency regulator site	Public	Rules for our product category	Monthly	Public, approved
Coldex price list page	Public	Competitor list prices	Irregular	Public, approved
Internal sales dashboard	Head of Sales	Bookings by region	Daily	Approved by Sales

Then set a schedule that fits the review. Northline's review is monthly, so the agent runs weekly and the final run lands two days before the meeting.

The sliders show why size matters. Six sources that each change about five times a week give 30 changes to check. If one in ten meets your criteria, the digest holds about three items. Double the sources and the noise doubles too.

Materiality criteria tell the agent what counts as a relevant change, and what to ignore. Link each one to a current priority or commitment, so every report item has a reason to be there.

Northline's criteria:

• Report any new or changed energy-efficiency rule in Spain affecting commercial refrigeration (priority: Spanish entry).
• Report any Coldex list-price change on models that compete with ours (priority: margin).
• Report any announced delay or price change from our two compressor suppliers (priority: supply).

Then state what is out of scope:

• General industry news, opinion pieces and conference announcements.
• Competitors other than Coldex, unless they enter Spain.
• Rule changes in markets we are not entering.

The out-of-scope list matters as much as the report list. Without it, the agent fills gaps with anything loosely related. When a criterion no longer links to a live priority, remove it.

Board discussions, acquisition plans and employee details stay out of unapproved cloud tools. That rule holds even if a platform is convenient, and even if the information would make a signal clearer.

In practice:

• Connect only sources you are permitted to use. If the board-papers folder is not approved for the platform, it is not on the source list.
• Name and stop. If a signal touches a sensitive matter, the agent records which source it came from and that it relates to a sensitive area, then stops. It does not copy, quote or summarise the content.
• Write the rule into the setup. Put it in the purpose or criteria so the agent follows it every run.

Example: Northline is quietly exploring buying a small Spanish distributor. A public news item mentions that distributor. The digest line reads: Source: El Economista, 12 May. Relates to a sensitive matter. Content not copied. Refer to CEO. The CEO reads the article herself, outside the platform.

Agree the report format before the first run, so every output looks the same.

Signal digest: one line per change with:

• Source and date
• What changed
• Why it may matter, naming the priority
• How certain it is: confirmed by the source, reported second-hand, or unclear

Competitor map: a one-page view limited to what the signals need. For Northline, that is Coldex's competing models and current list prices, not a full market study.

Human reviewer: name one person who checks each digest before it goes into the review pack, for example the Chief of Staff. Nothing is used until they have checked it.

Pause and retire: write down who can pause the agent, how (using the platform's pause or disconnect option, confirmed in its documentation), and when, such as a source behaving oddly or the reviewer being away. Retire it when the priorities it serves end, and record that you did.