Receive updates from a connected source
Use a webhook to receive an event from another system. Map the fields your process needs and use compatible custom connections where an API hand-off is required.
Business development and research
Turn business updates into a focused outreach review queue.
Bring connected news, financial filings and business updates into a review process. Evaluate each item against written criteria, keep the reasoning on a task board, and help your team decide where a relevant conversation may be worth starting.
How can a team decide which business updates deserve follow-up?
Connect a source of updates to a workflow, give an AI agent written evaluation criteria and let it create or update tasks on a dedicated review board. People can inspect the evidence and decide what to pursue. Scheduled summaries and instruction reviews help keep attention on the work and the criteria current.
Connect the signal, the decision and the next action
This anonymised example adapts a financial-filing classification pattern for business-development review. Other update types need a compatible feed and criteria suited to their content.
Make the criteria readable
Keep evaluation rules in a document the agent can read. Define what matters, the evidence needed and how to handle missing information, so reviewers can understand the basis of a decision.
Let your agent choose its own task actions
Give the AI step a Task Tool scoped to the review board. It can check for an existing item, create a task, update its classification and leave a rationale within the permissions you choose.
Review the work on a custom board
Use your own column labels for unreviewed, important, completed and lower-priority items. Keep the source facts, uncertainty, ownership and discussion on the task.
Bring outstanding work into chat
Schedule a separate summary agent to read the relevant tasks and prepare a digest. A Send Chat step delivers its output to the chosen person or conversation.
Use feedback to refine instructions
Configure a weekly review of decisions and comments against the criteria document. Make small, general improvements to the instructions while keeping a publication history.
Bring the update into the workflow with its context
A connected source can send an update to the workflow's webhook. The financial-update pattern uses five fields: company, report type, period, revenue and a signal describing the new information. A Combine step gathers those fields and the request content into one input for classification.
For example, a fictional payload could contain:
- Company: Company A.
- Report type: Trading update.
- Period: Q3.
- Revenue: GBP 76 million, with the unit stated explicitly.
- Signal: Management has withdrawn its outlook and announced a finance leadership change.
If the source supplies a reference or full announcement text, retain it with the task. A headline and one-line signal provide less evidence than a complete filing. The agent should identify what it received and what still needs verification.
A news or wider business-update feed can follow the same pattern once its events are mapped into suitable inputs. The feed connection, data rights and update content must be configured; this example does not assume a bundled subscription to every news provider.
For demonstrations, a manual input can pass sample values to a mock API that posts back to the webhook. That lets you exercise the intake path without representing mock data as a live vendor feed. Learn more about webhooks and APIs and reusable custom connections.
Separate the agent's job from the criteria it applies
The AI step has three distinct inputs to its behaviour. The incoming prompt contains the update to assess. Its standing instructions explain the job and the order of work. A selected document supplies the detailed evaluation criteria.
A classification instruction for an outreach-review adaptation could say:
Read the published evaluation criteria. Check the review board for identical information. For a new item, create an Unrated task and assess it against the criteria. For a repeat, log the repeated information on the existing task and reconsider its rating using the evidence and human comments. Explain the classification, identify missing facts and suggest what a person should verify before deciding whether to approach the company.
Keep the decision rules explicit. A financial-filings implementation can examine disclosure quality, financial pressure, governance changes and regulatory exposure. Its scoring anchors may use a 0–5 scale, with the meaning and direction defined separately for each criterion. A high disclosure-quality score and a high risk score describe different things.
For a business-development version, add the criteria that establish commercial relevance: whether the update is new, whether it suggests a need your team can credibly help with, and whether the timing makes a conversation appropriate. A risk signal alone does not establish that a company should be contacted.
Ask the agent to distinguish stated facts, interpretation and unknowns. Limited input should produce a limited assessment, with a request for verification where needed. Keep the criteria in a document, so the people responsible can read and revise them.
Give each agent the tools its role needs
The classification AI step is a workflow agent, separate from Butter in chat. Attaching tools lets it act on its assessment, rather than leaving a person to copy generated text into a board.
The pattern separates permissions by role:
- Classification agent: reads the selected criteria document and can read, create, update and comment on tasks in the selected review board. Its document access is read-only.
- Summary agent: reads the review board and comments on included tasks. It can prepare a notification without creating tasks, changing classifications or editing the criteria.
- Criteria-review agent: reads board decisions and feedback without changing the tasks. Its document tool can read, edit and publish the selected criteria document when that release policy is intended.
Set tool-round limits as well as board and document scopes. The agent chooses among the actions it has been given; the available tools and permissions define its reach. The task tool can be limited to this one board, rather than all the work someone can access.
In this configuration, Standard handles classification and summaries, while Pro is selected for criteria review. Model tier is a workflow setting, and availability depends on the plan. Explore workflow agents and tools and task-tool scoping.
Turn each assessment into a task people can review
Create a dedicated custom task board with column labels such as:
- Unrated: received and awaiting classification.
- Important: worth a person's attention under the current criteria.
- Not Important: below the current review threshold, with the reason retained.
- Done: the team has finished the review or follow-up it chose.
A new item begins as Unrated, then the agent sets the appropriate status and records its reasoning. The Important column represents a review state; task importance and urgency remain separate fields.
Keep the useful context together in the details and comments: the original update, the source reference where supplied, the criteria applied, the rationale, uncertainties and the proposed next check. Assign an owner and due date when the team decides who should act and when.
Before creating a task, the agent checks for identical information on the scoped board. A repeat can become a comment and a revised assessment on the existing task. This is an instructed matching process, so ambiguous matches still need checking; it should not be described as guaranteed duplicate prevention.
A person can correct the status and explain why in a comment. That creates feedback the later criteria review can inspect, rather than losing the decision in a separate conversation.
Illustrative example
From a trading update to a considered next step
Company A's fictional trading update arrives with the GBP 76 million revenue figure, withdrawn outlook and finance leadership change. The classifier checks the selected board, finds no identical item and creates Company A: trading update for review.
Its task details could contain:
- Reported facts: outlook withdrawn; finance leadership change announced.
- Why review it: the combination may matter to a team offering relevant finance or operational support, under that team's agreed criteria.
- What is missing: the full announcement, the reason for the leadership change and the scope of the revised outlook.
- Classification: Important for human review, with the uncertainty stated.
- Suggested next action: inspect the primary announcement and decide whether there is a relevant, appropriate reason to approach the company.
The reviewer might instead find that the leadership change is a planned retirement already announced earlier. They can change the assessment and comment: This is a routine succession already covered. A personnel mention alone does not create a new reason to approach.
A separate routine update from Company B, containing no new qualifying signal under the chosen criteria, could remain Not Important with an explanation. Both examples are fictional illustrations of the process, not investment recommendations or claims about a real company's circumstances.
The workflow creates a review queue. The person responsible decides whether to contact anyone, what to say and which facts have been verified. This example does not automatically send prospecting messages.
Send a useful summary of the work still needing attention
A separate scheduled branch reads tasks in the Important column and prepares a chat summary. Configure the weekdays, time and timezone to fit the team, and state the selection window clearly in the summary instructions. Decide whether the digest concerns newly received items or an outstanding backlog.
The summary agent has read and comment access to the review board. It adds a notification comment to included tasks, then returns the final message for a Send Chat step. The message goes to the selected person or conversation.
For the fictional example, a digest could say:
Updates awaiting review
Company A: outlook withdrawn and finance leadership change reported. Verify the full announcement before deciding whether there is a relevant reason to approach. Open the task for the original facts, assessment and discussion.
Keep commentary about preparing the summary out of the final notification. A notification comment records what the workflow reported doing; it is not proof that the recipient read the message. Configure logging at the appropriate point if delivery confirmation is required.
People can continue the review through their tasks and conversations, including on the mobile app.
Use a weekly review to improve the written criteria
Configure a weekly Scheduled Trigger for a separate instruction-review agent. When enabled on the published workflow, it can read the board's classification decisions, human comments, reclassifications and general feedback, then compare them with the criteria document.
The agent's brief is to identify clear mistakes and make small, general improvements. A useful change addresses the cause of an error, rather than adding a rule for one named company.
For the fictional succession example, a refinement might say:
Distinguish a routine, previously announced succession from a new and unexpected leadership departure. Explain the evidence for material change and mark missing context for verification.
This weekly calibration changes the written instructions used by future runs. The underlying AI model remains unchanged. The process needs representative, well-explained feedback, and its output still needs evaluation. It should not be presented as a guarantee that every future classification improves.
In an automatically publishing configuration, the reviewer has write and publish permission for the selected document. It updates the draft and publishes the revised criteria, so subsequent classifications can read them without changing the workflow canvas. Published document revisions provide a history of the changes.
If the team needs a person to release criteria changes, use a different permission and review configuration: keep the proposed revision in draft and add an explicit human review and publication step. That is an additional control to configure, not an approval gate already present in the automatic-publishing pattern.
Keep changes to workflow logic separate. Editing the workflow's steps or permissions affects its draft and needs workflow publication before those changes govern triggered runs. Documents and workflows explain the two publication lifecycles.
Explore the platform features behind the process
- Tasks and custom boards: organise assessments, owners, evidence and reviewer feedback.
- Documents: maintain readable criteria, prepare draft changes and keep published revisions.
- Workflows and AI tools: combine classification, scoped actions, schedules and chat delivery.
- Custom connections: reuse compatible API calls and webhook triggers.
- Webhooks and APIs: bring events from an external source into the workflow.
- Butter: work with your assistant in chat, separately from the agents inside the workflow.
- Mobile app: follow tasks and conversations away from your desk.
Questions about business-update triage
Can the same approach handle news as well as financial filings?
Yes, where a compatible source supplies those updates and the inputs are mapped into the workflow. Adapt the criteria to the content and commercial purpose. A news headline should not be evaluated as though it contained a complete set of financial accounts.
What happens when the feed provides only a short summary?
The assessment should state the limits of the supplied evidence and identify what a person needs to verify. Receiving a filing label and revenue figure does not establish that the agent has read the complete filing.
Does an Important task mean we should contact the company?
It means the item meets the configured threshold for review. A person still needs to verify the information, assess relevance and decide whether an approach is appropriate. The example does not automatically send outreach or make an investment recommendation.
What does the weekly learning loop change?
It calibrates the published criteria document using decisions and feedback from the review board. Future classifications can use the revised instructions. It does not change the underlying AI model's parameters.
Can criteria changes require human approval?
Yes, with an explicitly configured review and release process. The automatic-publishing variant gives the reviewer document-write and publish rights. To require approval, keep proposed changes in draft and configure a human review and publication step.
Can the agents be limited to one board and one criteria document?
Yes. Scope their Task Tool and Doc Tool to the selected resources and enable only the actions each role needs. The classifier, summariser and criteria reviewer can have different permissions.
When do scheduled summaries and instruction reviews run?
Only when their schedules are enabled and the workflow is published. Configure the weekdays, time and timezone separately for each branch. Draft changes and an inactive schedule do not start the recurring process.
Start with the updates your team already reviews.
Write down what makes an item worth attention, connect one source and give the team a board to review the results. Use its feedback to refine the next set of decisions.
No card. Your data stays in your organisation.