Tools and agents
Choose the tools around the work your team needs.
A CRM feature, an integration, a custom view and an AI agent solve different problems. Start with the person who needs the result, then compare the options using the same records, permissions and awkward cases. Keep what already works.
In this guide
Describe the complete handover
A staffing business wants to know when a former client starts hiring again. That job needs a current vacancy, the previous placement history, the right account owner and a reason to review the account. Finding a job advert alone does not complete it. Sending a message automatically adds another job with different approvals.
Write a short decision statement: “When an approved source shows relevant hiring at a former client, prepare an account brief for its owner to review.” Add what the owner will decide, where the brief should appear and what happens when the vacancy belongs to a different office. That statement lets you compare implementations on the same basis.
List what already works. Your CRM may hold a reliable ownership rule while the missing step is checking a source or resolving an uncertain match. Our seven-part system helps identify those connections without making the whole system compulsory. The workflow specification turns the trigger, decision and acceptance conditions into an implementation brief.
Decide where each fact belongs
A CRM, ERP, applicant tracking system or shared sheet may hold part of the answer. Name the account identifier, owner and accepted place for the next action. If the same business appears in two systems, establish how those records are matched before adding another connection.
Keep ownership at field level. Finance may own billing details, the account team may own the relationship and an external source may supply a dated observation. A newer website result should not silently replace an approved customer record. The shared customer data guide explains how to separate identity, source priority and changes across tools.
In a demonstration, change the account owner, run the same input twice and interrupt a connection. Ask to see the resulting records. A useful demonstration shows where the work lands and what remains unresolved, as well as the successful first run.
Check what the source actually supplies
An enrichment waterfall tries approved sources in sequence until it finds a usable result or reaches a stopping point. That sequence needs a definition of usable: the correct company or person, the needed field, a source and a date. An available email address does not establish a reason or permission to contact its owner.
Use a mixed sample of known accounts, difficult matches and records with missing details. Count correct results, wrong matches, unresolved cases and the cost of accepted information. Keep missing and wrong separate: an empty field can ask for review; a confident result for the wrong company can send the rest of the workflow in the wrong direction.
The CRM enrichment evaluation turns that sample into a repeatable comparison. Include the human time needed to verify a match, not only the price per lookup.
Choose how much freedom the task needs
Use a fixed workflow when you can describe the steps and branching rules in advance. An agent can choose intermediate steps within a task. Anthropic explains this distinction between predefined workflows and agent-directed work. Choose the freedom the particular task needs.
“AI sales agent” is too broad to approve. Separate research, brief preparation, CRM updates, scheduling and sending. A research assistant may need to choose which approved page to inspect next, while an ownership update can follow a fixed rule. Neither requires authority to change prices or make commitments to a customer.
Use an agent where that bounded choice is useful and testable. Write its operating contract before giving it tools. If the output cannot be checked, adding more freedom makes the operating problem harder to contain.
Four ways to prepare a former-client brief
The staffing team has reliable account ownership but checks vacancy pages manually. Compare four ways to prepare the same brief. Each option needs evidence that it can handle the team’s actual records and exceptions.
Existing feature
- What it would do
- Show a saved account view and assign a review task
- Evidence needed before choosing
- Confirm the current licence can expose the required history, owner and due date without manual copying.
Integration
- What it would do
- Bring an approved vacancy observation into the account record
- Evidence needed before choosing
- Test matching, repeated observations, source failure and the person who resolves an uncertain match.
Custom working tool
- What it would do
- Put conflicting office records and source evidence in one review screen
- Evidence needed before choosing
- Show why existing views cannot support the decision; define allowed edits and who maintains the screen.
Bounded agent
- What it would do
- Choose relevant pages within approved sources and draft a brief
- Evidence needed before choosing
- Require dated references, a limited search budget, an explicit stop and human review before any commercial action.
The team might choose an integration plus its existing review view if the sources are predictable. It might add bounded research when the evidence varies across pages. The choice follows the unresolved work; these options are not a maturity ladder.
Record the selected option, the alternative considered, the decisive test and the owner. Revisit that record when a source, volume or operating responsibility changes, rather than expanding access whenever an exception appears.
Separate reading, preparing and acting
Give each connection only the records and operations required for the agreed task. Reading an account, preparing a draft, editing an owner and sending a message are separate permissions. A review button is useful only if the underlying system also prevents an unapproved write or send.
OWASP’s excessive-agency guidance calls for constrained capabilities and permissions, with authorisation enforced outside the model. In this brief, the research component may read approved sources and save a draft; it cannot send outreach or alter commercial terms.
Check the reviewer’s access too. A brief should not expose a customer document to someone who could not otherwise read it. Keep credentials out of prompts, briefs and logs, and identify who can revoke the connection if its owner leaves.
Maya is the account owner. She reviews the brief and can pause the work.
- Research
- Read account history and approved source pages, retaining dates and references.
- Preparation
- Save a draft brief against the account.
- Sending
- Separate from research. The sending step waits for Maya’s approval of the draft and the agreed recipient checks.
- Prices
- Changes are not permitted.
- Refer to Maya
- Stop when an owner is missing, evidence conflicts or the account is excluded.
Treat sending as its own readiness decision
Before enabling outreach, name the audience, source of contact details, relevant markets, exclusions, suppression process and person responsible for replies and stop decisions. A positive account signal is an input to that review, not an instruction to send.
Authentication and delivery monitoring also need an owner. Google’s sender guidelines concern delivery to personal Gmail accounts; they do not establish permission to contact someone or guarantee an inbox placement. Use the sending-readiness record to keep technical checks and recipient decisions visible.
Ask for evidence of the difficult cases
Test a correct account, a duplicate observation, stale information, a missing owner and conflicting company identities. Agree the expected result for each case before the demonstration. An unresolved item that reaches the right review queue can be a correct result; an invented answer cannot.
Also try a forbidden action. A source page telling the agent to send a message should not expand its authority. Confirm that rejected operations leave the protected fields unchanged. If a connection fails after a write may have occurred, the result must stay uncertain until the operator checks the destination; a failed response is not proof that nothing changed. Give the operator the request reference and a clear recovery step.
Keep the inputs, observed outputs, reviewer and decision with the handover. A supplier’s general feature list cannot show whether your particular records and exceptions have passed.
Include the cost and responsibility of operating it
Compare the full cost of implementation and operation: licences, data, usage, human review and maintenance. Identify who pays each subscription, controls credentials, can export records and handles provider changes. Include the time needed to investigate held cases.
Name the person responsible for the complete handover, even when several tools are involved. They need to know how to pause the workflow, find exceptions and recognise a change that needs testing. Give them the operating record with the configuration and handover.
Choose the commercial scope separately from the tool. A Build campaign combines the agreed research, setup, outreach operation and improvement. A standalone connection or custom working tool receives its own scope and quote; AI Workflows covers bounded AI tasks. If an inherited setup needs investigation before its scope is reliable, agree a paid Plan assessment first. The starting prices for Build and AI Workflows apply to their defined offers; bespoke software is quoted for its own scope.
Choose delivery responsibilities separately. We can hand research and account evidence to your team, operate agreed outreach or share the work. A tool choice should support that arrangement. The working agreements explain scope, ownership and handover.
Which part of the job do your current tools leave unfinished?
We can review the handover, existing tools and hardest exception, then scope the missing research, connection or workflow. Start with the result the recipient needs.
Discuss the implementationA Fit Call is a free 30-minute video conversation. No preparation needed.