Enrichment is not a one-off job
Firmographic data ages. An enrichment run is a schedule, not a project. Re-verify, decide what may overwrite, and never silently replace a value a human typed.
Notes from putting an agent on a live queue.
Permissions, exceptions, write-back, and the one-workflow package. Written for Heads of Ops, CX, and RevOps.
Topics
Articles
No articles in this topic.
What to tell the team before the agent goes live in their queue
For the champion. Name which steps move and which stay, before rumour does it. The agent takes the repeatable work. The team keeps anything that can change money or a customer.
The document piles that are not ready for an agent
Handwriting, unstable formats, a queue about to be replaced, a process with no rule. If that is your pile, do not buy an agent yet. Fix the pile, or wait.
What one live workflow costs to build and to keep running
The cost lines behind $9,900 to go live and $2,900 a month: model calls, the integration surface, the exception reviewer's time, and change work when your systems move. No invented savings.
Your routing SLA is probably measuring assignment, not response
The clock usually starts when the lead is assigned. The buyer feels first human action. An agent can enrich, score, and prepare the next step. It cannot make a rep pick up the phone.
The exception queue is the part that decides whether this lasts
Separate exception types by cause, name an owner for each, keep ageing visible, and attach the evidence. One undifferentiated queue is how the automation becomes a second manual process.
Checking an agent on real items before it is allowed to write
Read-only first, then shadow mode against decisions a person already made, then a graduated write scope. A sandbox hides the parts that break you.
When the feature you already pay for is enough
What native helpdesk, CRM, and ERP AI actually covers, where it stops, and the cases where you should turn that feature on before you call a partner.
A one-workflow business case that survives a finance review
Structure, not invented savings. What to measure before you start so the after-number means something, and why a named queue can be approved when 'AI for customer service' cannot.
Build it, buy it, or have someone put it live
Three ways to get one workflow automated, the failure mode of each, and the cases where buying the tool you already pay for is the right answer and we are not.
MuleSoft vCore pricing: what to do with a bill you can no longer read (vCore)
Salesforce does not publish a public vCore list price. If you are still on a vCore contract, the unit that matters is production runtime on the order form. Bring that, plus utilisation, and we will tell you which workers to shrink, merge, or leave.
What you will need to show about a live agent
Dated 4 September 2026, and hedged. Named overseer, authority to stop the writes, what the log has to contain. Most support, sales, and document workflows are probably not high-risk under the EU AI Act. The evidence is still worth having.
What your IT team should ask before an agent gets write access
A list you can forward to the person who can veto the project: the permission set, where the model call goes, what is in the log, and what we cannot answer without seeing the tenant.
Choosing the first workflow to put an agent on
The five conditions a first workflow has to meet, the disqualifiers that matter more than any score, and why the most valuable process is usually the wrong one to start with.
Deciding which agent actions need a person
Sort the actions by what happens if the agent is wrong and how hard the mistake is to undo. Confidence scores are the wrong gate.
Extraction is the easy half of invoice processing
Reading the PDF is largely solved. Validating what you read against the purchase order and the goods receipt, and routing what does not reconcile, is the work that is still yours.
Letting an agent write into Salesforce without dirtying the pipeline
The write pattern that keeps a CRM trustworthy when an agent is updating it: proposal fields, a deterministic gate, and conflict detection.
Automating helpdesk triage without letting it answer the customer
How to put an agent on ticket triage in the helpdesk you already run: which fields it writes, which it must not, and what breaks first.
Why the pilot worked and the go-live did not
Four things a convincing AI demo leaves out, and what has to be settled before an agent is allowed to write into a live system.
Practical AI automation without losing control
Connecting agents to real workflows, systems, and human approval paths. Older note, kept because it still describes the control question.
How to evaluate a technology outsourcing partner
What enterprise technology leaders should look for in team formation, governance, and delivery continuity.